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.1000701 | Non-Linear Neuronal Responses as an Emergent Property of Afferent Networks: A Case Study of the Locust Lobula Giant Movement Detector | In principle it appears advantageous for single neurons to perform non-linear operations. Indeed it has been reported that some neurons show signatures of such operations in their electrophysiological response. A particular case in point is the Lobula Giant Movement Detector (LGMD) neuron of the locust, which is reported to locally perform a functional multiplication. Given the wide ramifications of this suggestion with respect to our understanding of neuronal computations, it is essential that this interpretation of the LGMD as a local multiplication unit is thoroughly tested. Here we evaluate an alternative model that tests the hypothesis that the non-linear responses of the LGMD neuron emerge from the interactions of many neurons in the opto-motor processing structure of the locust. We show, by exposing our model to standard LGMD stimulation protocols, that the properties of the LGMD that were seen as a hallmark of local non-linear operations can be explained as emerging from the dynamics of the pre-synaptic network. Moreover, we demonstrate that these properties strongly depend on the details of the synaptic projections from the medulla to the LGMD. From these observations we deduce a number of testable predictions. To assess the real-time properties of our model we applied it to a high-speed robot. These robot results show that our model of the locust opto-motor system is able to reliably stabilize the movement trajectory of the robot and can robustly support collision avoidance. In addition, these behavioural experiments suggest that the emergent non-linear responses of the LGMD neuron enhance the system's collision detection acuity. We show how all reported properties of this neuron are consistently reproduced by this alternative model, and how they emerge from the overall opto-motor processing structure of the locust. Hence, our results propose an alternative view on neuronal computation that emphasizes the network properties as opposed to the local transformations that can be performed by single neurons.
| The tiny brains of insects of about 1mm3 smoothly control a flying platform while avoiding obstacles, regulating its distance to objects and search for objects of interest. This is largely achieved through a complex hierarchical processing of signals from the multitude of ommatidia in their eye to a set of highly specialized neurons that are optimized to respond to specific properties of the visual world. One of these neurons, the Lobula Giant Movement Detector (LGMD) of the locust, has been recently shown to perform a functional multiplication of its synaptic inputs. If true, that would make the LGMD neuron a unique and highly sophisticated neuron that raises questions about the non-linear operations other neurons in other neuronal systems would be able to perform. Hence it is crucial to understand its properties, its role in behaviour and to evaluate whether its responses can be explained in simpler terms. Our results emphasize the role of network architecture and distributed computation as opposed to local complex non-linear computation. We show that our model reliably reproduces the known properties of the LGMD and can be used to control a high-speed robot.
| Since the introduction of the neuron doctrine about 100 years ago, a central question has become what local operations the primitive elements of nervous systems can perform. So far, the only operation that has clear experimental support is the threshold operation that converts the depolarization of the membrane into action potentials. However, also other local non-linear operations such as multiplications and divisions have been proposed. For instance, the Elementary Motion Detector (EMD), a well-established model of motion detection in the fly visual system that relies on multiplication in order to explain the neural responses of the Horizontal and Vertical System (HS, VS) visual interneurons [1]. In addition, it has been proposed that attentional modulation can result in a multiplicative gain of neuronal response to sensory stimuli [2]. Another example is the divisive inhibition that is assumed to underlie some of the non-linear adaptation properties of cortical neurons [3],[4], while several other studies have investigated how neuronal noise or dendritic saturation could contribute to divisive gain control [5],[6]. Moreover, theoretical studies on neocortical pyramidal cells have suggested that multiplicative dendritic integration could account for non-linear sensory processing enhancing stimulus classification [7],[8]. Despite the above examples, its computational attractiveness and the fact that some data can be satisfactorily described in non-linear terms, it remains unclear how the biophysics of single neurons could implement these operations.
One particular case in point is the Lobula Giant Movement Detector (LGMD) visual interneuron of the locust. Recently it has been shown that the responses of this visual interneuron can be explained in terms of a local product of two high-level features of visual stimuli, their angular size and angular speed by means of a non-linear transfer function of the neuron [9],[10]. If correct, this is the most explicit case reported in the experimental literature that supports the notion of local non-linear neuronal operations and it will have important consequences for our understanding of the computations that the nervous system can perform, as it significantly increases the computational power we can ascribe to single neurons. Hence, given the implications of this finding, it is important to investigate whether the non-linear relationship between the responses of the LGMD neuron and the visual stimuli it is exposed to can be understood in alternative terms, yet consistent with our current knowledge of the system. Here we investigated the alternative hypothesis that the non-linear responses of the LGMD can be explained in terms of an emergent non-linear operation that results from the integration of distributed computations performed by the neurons of the processing architecture as a whole as opposed to being a multiplication operation that is local to a single unit, i.e. the LGMD.
The LGMD neuron is a wide-field neuron that is known to respond preferentially to looming stimuli [11],[12]. Initially, it was first thought to be an on-off neuron due to its integration of neuronal responses generated in the afferent medulla layer that correlate with the onset and offset of local visual features [13]–[15]. More recently the relationship between properties of looming stimuli and the firing rate of the LGMD have been extensively documented, including the non-linear relationship between firing rate and time to collision (TTC), the constant relation between peak firing rate and angular size, the independence of the peak firing rate of the stimulus speed, shape and texture, and the linear relationship between the TTC of the LGMD peak firing rate and the apparent looming stimulus' speed [9], [16]–[18]. The LGMD has been the target of a number of theoretical studies that either investigated its collision detection capabilities [19]–[22], or its putative non-linear integration properties [9], [16]–[18]. The first model was published in the late 90's [22]. Rind et al. have shown that the integration of on- and off-channels by a LGMD model can account for aspects of its looming sensitivity and subsequently this model has been applied to collision avoidance by roving robots [22]–[26]. Recently, it has been shown that all of the known response properties of the LGMD can be accounted for in terms of the multiplication of the angular velocity (θ′) with the angular size (θ) of a looming stimulus [9], where θ and log (θ′) are directly conveyed to the LGMD via separate inhibitory and excitatory pathways (Figure 1). The membrane potential (Vm) deflection is subsequently assumed to be proportional to this multiplication that is subsequently expressed in a firing rate, f−l, via an exponential mapping:(1)where,(2)and θthreshold is an animal and species dependent parameter that specifies the approaching object's angular size at which the LGMD firing rate is maximal [27]. Hence, by performing an exponential on the summed inputs an effective multiplication occurs. This model indeed provides for an excellent fit of the LGMD responses to looming stimuli, and as such constitutes a useful benchmark for any model of the LGMD [10]. Nevertheless, this local multiplicative model makes a number of strong assumptions and overlooks the role of the neurons pre-synaptic to the LGMD. More concretely: how does the fan-in to the LGMD delineate an “object” of which θ′ and θ can be assessed, given that an “object” has been defined, how are log (θ′) and θ computed, how is this high-level information represented by the massive fan-in to the LGMD, and how are the parameters related to the approaching stimulus (θ′ and θ) extracted and conveyed to the LGMD in the early visual system of the locust? Moreover, this proposal assumes that the excitatory and inhibitory inputs to the LGMD respond to high level information about the visual stimulus (θ′ and θ) and that the role of the LGMD is to compute a functional multiplication on those. By definition, the functional multiplication attributed to the LGMD heavily depends on having the two above mentioned features reliably computed and delivered to distinct pathways. However, in mathematical terms, there is not a unique combination of input signals to the LGMD that could give rise to the above described firing rate pattern (eq. 1), and thus, no reason to exclude this possibility. Indeed, our model suggests that this is the case (Figure 1, layers C–D right panel). Would the LGMD in that case still perform a functional multiplication or just a non-linear mapping of its inputs? In fact, the putative multiplicative properties of the LGMD have already been a matter of debate [28]–[30]. In this study we approach the above mentioned points from a system and architectural point of view. We evaluate the alternative hypothesis that the non-linear relationship between the responses of the LGMD neuron and the stimuli the organism is exposed to result from the interaction of many neurons in the sensory processing architecture, i.e. it is an emergent non-linearity that is read-out by the LGMD. In particular, we will assess the contribution of each processing layer in the visual processing hierarchy of the locust, how and what information is conveyed to the LGMD, and the resulting integration at the LGMD level. The empirical validation of this alternative hypothesis, however, is currently unpractical since it requires simultaneous in-vivo measurements from large numbers of neurons under well-controlled behavioural conditions. Hence, to assess the validity of our alternative “emergent non-linearity” hypothesis we resort to a computational approach and use a computational model that is consistent with the anatomy and physiology of the locust visual processing hierarchy, including the ommatidia, medulla, lobula, LGMD and the Descending Contra-lateral Movement Detector (DCMD). Using this model we show that all properties of the LGMD neuron that can be described in terms of a local non-linear operation can be explained as emerging from the structure of the network as a whole. Above all, we show that the inputs to the LGMD are directly driven by the stimulus dynamics rather than resulting from a process of segmentation or computation of the speed of the approaching objects. Despite the differences with Gabbiani's et al. model, the model proposed here displays identical responses to its biological counterpart on all standard stimulation protocols reported in the literature. We demonstrate that the emergent non-linear operations are strongly dependent on the details of the synaptic organization of the locust's visual system. In addition, we apply our model to a high-speed mobile vehicle and show that it reliably stabilizes the movement trajectory and robustly avoids collisions. Hence, our model not only suggests that the functional non-linear response properties of the LGMD emerge out of the network as a whole but also shows robust and realistic real-world properties.
The structure of our model consists of four layers that capture the most relevant processes involved in the pathway to the LGMD, and both the output of the LGMD and the population responses for each layer are considered (Figure 1). We model the photoreceptor layer with Linear Threshold (LT) units that are driven by a CCD camera with automatic gain control (see Experimental Procedures) (Figure 1A). The lamina is modelled with a centre/surround connectivity that produces an edge enhancement [31] (Figure 1B). Subsequently, neurons in the medulla layer produce onset and offset sensitive responses [13]–[15] (Figure 1C). The connectivity between the medulla and the lobula layer transduces the excitatory input to the LGMD (Figure 1D). Post-synaptic inhibition onto the LGMD is modelled through the integration of the activity of the onset and offset sensitive neurons in the medulla where the summed activity inhibits the excitatory projections onto the LGMD from the second chiasma. The LGMD is modelled as an Integrate and Fire (I&F) neuron that integrates the above mentioned excitatory and inhibitory inputs from the medulla and produces spikes (Figure 1E). All neurons in our model are standard leaky I&F or leaky LT neurons [32],[33] (see Experimental Procedures for model details and dynamic equations).
In the context of this study, we present an exhaustive analysis of the responses of our model to a set of standard LGMD stimulation protocols that allow us to validate our model with respect to the biological system. Additionally, the contribution of each neural layer of the model to the LGMD responses' properties is assessed experimentally (Figure 1A–E) as well as analytically, and its real-world properties are evaluated using a fast moving robot.
In our first experiment we evaluate the model by using a looming stimulus consisting of a solid square with 10 to 21 repetitions performed per each l/|v| pair (where l stands for the half length of the object and v for its linear velocity) (see the Experimental Procedures for further details). This ratio determines the time course of the angular size (θ) of the looming stimulus in an independent fashion from the actual stimulus properties (eq. 3). This experiment replicates the protocol used in [17].
Our model of the LGMD displays the typical response of this neuron to an approaching stimulus (Figure 2A); as the angular size of the retinal projection of the stimulus increases, the firing rate increases, peaks and decays before the collision occurs. This response closely resembles that of the biological data with the multiplicative model (r = 0.98) (Figure 2A, middle panel). We observe that the fit of the peak firing rate and the TTC versus the l/|v| ratio is consistent with that observed in the biological system, and is well captured by the multiplicative model that was derived from LGMD recordings (eq. 1) (Figure 2B).
The response of the LGMD neuron has been shown to peak when the angular size of the projection of the looming stimuli onto the retina of the insect reaches a specific size, known as the angular threshold [9],[10],[16],[17],[34],[35]. Moreover, the time at which the response of the LGMD peaks, that is, when the stimulus reaches the angular threshold, depends linearly on the l/|v| ratio. This reflects a robust detection of the angular threshold over a wide range of l/|v| ratios since the time at which the response of the LGMD peaks is proportional to l/|v|. The linear relationship between TTC of the peak firing rate and the l/|v| ratio is a known property of the LGMD [17],[18] (eq. 2), that is reliably replicated by our model (r>0.99) (Figure 2C).
We propose a specific connectivity for the LGMD pre-synaptic fan-in such that the projections from the medulla to the lobula integrate oriented contrast boundaries (see Experimental Procedures). These projections are retinotopic and integrate the activity of a set of on-off neurons of the medulla that surround its location at distances δx and δy (surround excitation). Consequently, δx and δy define the width and height of the region within which the boundaries of a looming stimulus have to fall in order to achieve maximal excitation, what defines the angular threshold. To further test this aspect of the model, we performed a control experiment in which we varied δx and δy to define a surround receptive field of 25, 29 and 36 degrees of the camera's field of view. The predicted behaviour of our model is that the changes in δx and δy would affect the angle of the peak firing rate, and therefore the TTC. Indeed, we obtained a change of the slope of the linear regression between the frequency peak and the l/|v| ratio which correlates with the changes in δx and δy; the bigger δx and δy, and hence the angular threshold. The later the LGMD response firing rate reaches its maximum and the flatter the slope is (Figure 2C). In conclusion, our model is consistent with the known properties of the LGMD [9],[16],[17] and shows that the response peak is defined by the topology of the projections from the medulla to the LGMD.
It was shown that the responses of the LGMD are largely independent of the shape of the stimulus and its texture [17]. In a series of experiments, we assessed whether our model shows similar invariant properties (Figure 3). To do so, consistent with previous experiments [17], we used four different shapes. For all stimuli tested, and over the whole range of l/|v| ratios (from 5ms to 50ms), the model's responses show the same linear relationship with the TTC as reported for the biological system, with a correlation coefficient between the model's responses and the regression lines of r>0.99 (Figure 3C).
The response invariance to the approach angle of looming stimuli is biologically highly relevant in a system that can serve to detect potential predators, as is the LGMD. This reported property of the LGMD was investigated in the last set of experiments. The invariance was assessed by using the same experimental protocol as previously employed, but now aligning the camera at different angular orientations with respect to the projection screen as was reported in [17]. In the following we refer to 0% of the visual field when there is a complete alignment of the camera orientation and screen, and to 100% when the centre of the screen is at the edge of the camera's visual field (Figure 4B, insertion).
We found that our LGMD model shows a robust response invariant to the approach angle up to an angle that represents approximately 75% of the visual field (Figure 4). A one-way ANOVA analysis of the distribution of the model responses revealed that a significant difference in the mean number of spikes only occurs at an angle exceeding 75% of the total visual field of the camera (approximately 30°) (p<0.05), i.e. when the stimulus was partially lying outside of the visual field of the camera. Although the fields of view of the locust eye and our camera are not equivalent, yet we have designed it to have a similar angular resolution of 2.33° per pixel [36]. Additionally, the fraction of field of view where the response is invariant is comparable to the one of the biological system [16] (Figure 4A).
Subsequently, we investigated the linear relationship of the TTC of our LGMD model over a wide range of l/|v| ratios and approach angles. Our results show that the invariance of the response properties can be seen as well in the TTC domain (Figure 4B). Here, the correlation coefficients of the data and its linear regression are above 0.9 for both a perfect alignment between the camera and the screen and in case of a misalignment of 75% of the visual field. Thus, even though the activity of the neural model is significantly reduced due to the loss of stimulation by the looming stimulus at a very shallow approach angle (Figure 4A), the intrinsic linear dependence of the TTC with respect to the l/|v| of the LGMD is preserved (Figure 4B).
In order to understand better and to be more specific about the nature of the inputs to the LGMD, we propose the use of additional stimulation protocols that can be applied to the locust using currently available experimental technologies. For instance, in the multiplicative model, the firing rate of the LGMD is defined by the product of the angular speed (θ′) and a value related to the object's angular size (θ) (eq. 1). If those two variables were indeed the input to the LGMD, it would imply that for an object that is approaching at a constant angular speed the LGMD should display a completely different time response. In fact, since the angular approaching speed of the stimulus (θ′) would be constant, the predicted output by the multiplicative model would be an exponentially decreasing firing rate. Hence, we explicitly evaluate the different responses between our model for each neural layer and the multiplicative one by using objects that show a uniform increase in size (Figure 5, layers A–E left panel). We observe that, whereas the multiplicative model displays the expected exponential decreasing response, our emergent non-linearity model still displays a peak at the preferred angular size. This stimulation protocol was previously used by Hatsopoulos et al. showing a response profile consisting of a fast increase of the firing rate, a peak and subsequently followed by a slower decrease of the activity [18]. Although some of the data could eventually be approximated by an exponential function, a more quantitative analysis of the LGMD responses is required in order to find the relationship between stimuli and rising, peak and decay properties of the responses of the LGMD under this protocol. Additionally, we see that the predicted excitatory input to the LGMD with our model differs from the constant factor predicted by the multiplicative fit (Figure 5, layer D left panel). Thus, a further examination of the LGMD responses under this protocol is essential to unveil what the real input to the LGMD is, and therefore to understand whether it computes a product of the object's angular size (θ) and angular speed (θ′) or responds to a different processing as suggested by our results.
Next, we analyze the activity of each layer of our model to identify the relationship between receding stimuli and the intensity of the LGMD response (Figure 5, layers A–E right panel). The responses are consistent with a number of experiments of stimulus selectivity of the LGMD that showed a preference for looming stimuli and its diminished response to receding ones [11],[12],[35]. Two hypothetical peaks in the TTC curve to receding stimulus are predicted depending on the weighting of the post-synaptic inhibition (Figure 5, right panel). We show that the specific amplitude-time course of this response depends on the gain of the inhibitory projections onto the LGMD.
So far we have shown that we can account for all known aspects of the responses of the LGMD neuron to looming stimuli with a model that relies on the transformations performed in the complete pathway from the photoreceptors to the LGMD as opposed to a local multiplication. We now want to assess the behavioural validity of our model by applying it to a high-speed impeller driven based robot called “Strider”. Given its structure, the Strider is highly sensitive to inertia and friction forces, yet it delivers high-speeds. For the robot to be sensitive to shallow approach angles, its camera was equipped with a wide angle lens (190 deg. field of view). Although the aim of the robot is to have dynamics comparable to that of a flying insect, our robot has longer reaction times due to its increased mass, i.e. it operates at a higher Reynolds number than a flying insect. We therefore use a course stabilization system to guarantee that the robot is able to follow straight trajectories. This course stabilization system is based on the fly's Elementary Motion Detectors (EMD) and uses directional motion information from the visual input to correct for drifts, and has been previously deployed on flying vehicles [37].
The real-world behavioural task of the robot is to drive forward on a straight course until an imminent collision is detected. The modelled LGMD neuron will detect this upcoming collision and induce a collision avoidance reaction that consists of two phases: first deceleration of the robot, and then change of heading direction. To deal with the inertia of the robot, the braking manoeuvre is realized by driving the impellers backwards at full speed for one second. The change in heading direction is achieved by a turn-in-place manoeuvre of 1.25s duration. The LGMD model reports the detection of an imminent collision when its firing rate exceeds a specific threshold, and will trigger avoidance actions until its firing rate decreases below the above mentioned threshold value.
The following analysis is based on 16 experiments in a confined environment of 3.5 by 4.5 meters that lasted approximately 3 minutes each, where both course stabilization and collision avoidance systems were active. Additionally, we performed 5 control experiments where the LGMD neuron model was active but the course stabilization system was disabled.
The experimental results confirm the necessity of a course stabilization system: when the robot is solely controlled by the LGMD model it displayed an erratic behaviour dominated by multiple loops in either one direction or the other (Figure 6A, right panel). When the LGMD model is combined with the EMD-based course stabilization system, the robot exhibited longer periods of translation exploring a larger area, and had a less variable heading direction (Figure 6A, polar plots). The nearly uniform distribution of the variation of the heading direction during the control experiments (Figure 6A, right panel polar histogram) is the result of the continuous changes that result from the complex dynamics of the Strider robot. When both the LGMD model and the course stabilization system were combined, this distribution was significantly different and reduced to a few preferred heading directions (Figure 6A, left panel polar histogram) (p<0.01, Kolmogorov-Smirnov). To further demonstrate the effect of the course stabilization system in the control of the behaviour of the robot, a linear segmental fitting of the behavioural traces, consisting of finding a sequence of linear segments that keep the Mean Square Error (MSE) of the fit below a threshold value, was performed (Figure 6A). This measure allows quantifying the straightness of the trajectory. That is, the longer the segments are on average, the straighter the overall trajectories are (Figure 6D). In order to assess the dependency of the fit upon the threshold value, different threshold values were tested. All tested values yielded comparable results. Although it is not the objective of this study to evaluate our course stabilization model, these data serve to illustrate the complex dynamics of the Strider robot. A statistical analysis of the segment length distribution (two-sample Kolmogorov-Smirnov) showed that in case of the combined system, the distributions of the linear segments were significantly different (p<0.01). Longer segments and a higher variance were obtained for the combined system (Figure 6D); concluding that the stabilization system contributes significantly to the straightness of the trajectory. Therefore, the course stabilization system we included is an essential component in order to deal with the dynamics of the Strider, and allows us to perform and evaluate the collision avoidance task with a high-speed robot.
To evaluate the performance of the LGMD component of the robot system, all collision detections were classified into three groups: correctly detected, false negatives (missed), and false positives. These data have to be read in the context of this fast moving robot, that on average detects a collision 0.5m away from a wall while moving at a mean speed of 1.2m/s. Hence, if the robot does not dramatically change its speed at the moment of detection, it collides in less than half a second. Collisions detected 20–100cm away from the walls were considered as correct, while all collisions detected closer than 20cm from the wall were considered to be detected too late, and thus missed (false negative). Conversely, collisions detected farther than 100cm from the walls, were considered false positives (Figure 6A, grey dashed region). In total, 87.8% of the detections were correct, 4.9% were false positives and 7.3% were missed (Figure 6B). The distribution of the number of detections vs. the distance to the wall at the time of detection peaked at 0.5m, and decreased exponentially further away from the wall (Figure 6C). Thus, the behaviour of the robot directly results from the non-linear nature of the response of the LGMD model (Figure 2). Since the responses of the DCMD neuron feed directly into the thoracic motor ganglia of the locust that control the wing muscles, this seems to suggest that the amplitude-time course of the LGMD defines a particular collision avoidance strategy that minimizes the number of false positives as the distance to objects increases. In conclusion, these behavioural experiments suggest that the exponential transfer function of the LGMD neuron [9] is more related to its role in the regulation of behaviour rather than to the computation of object approach per se.
The question whether neurons can perform non-linear operations is of great relevance to answer what computations neuronal systems can be expected to perform. It has been argued, on the basis of the physiology of the LGMD neuron, that these neurons can perform a multiplication of high-level features of visual stimuli in order to detect pending collisions [9], [10], [16]–[18]. Gabbiani et al. proposed a model that provides for an excellent fit of the LGMD responses to looming stimuli, and as such constitutes a useful benchmark for any model of the LGMD. Using a biologically constrained model of the locust visual system, we have demonstrated that an alternative interpretation can not be excluded. In this alternative view, the local non-linear transfer function of the LGMD neuron can be accounted for in terms of the physiological and anatomical properties of its afferent visual processing hierarchy. We tested our model using simulated analogues of the locust experiments reported in the literature and assessed the real-world validity of our model using a high-speed robot. We showed that our model is able to account for all reported properties of the LGMD neuron without assuming any non-linearities other than thresholding that is intrinsic to standard leaky I&F and leaky LT neural models [32],[33] (see Materials and Methods). Consistent with our model, recent findings support the existence of a retinotopic mapping of the LGMD pre-synaptic network and suggest that a topographic map would be used to magnify the dendritic sampling of the acute zone [38]. Our model proposes an alternative view that suggests that a non-linear transfer function between stimulus and response can emerge out of the interaction of many distributed neuronal operations and their specific mapping through synaptic topologies. Moreover, our simulations show that the computation of angular speed and angular size pre-synaptic to the LGMD is not necessary to explain its properties. It has been reported that the LGMD shows an exponential relationship between the membrane potential and the firing frequency [9],[17]. Such properties are standard to integrate and fire neurons and can be explained in terms of their sigmoid transfer function [39]. As such, we believe that the LGMD has a similar transfer function and we have included it in our model. Additionally, our experiments reveal that this non-linear transfer function does not play a significant computational role in the detection of a collision, but rather that it shapes the LGMD response with respect to the behaviour requirements of collision avoidance, as demonstrated with our robot experiments.
In our analysis we have presented a plausible model of how motion selective responses can arise from the interaction between on-set and off-set sensitive neurons. The idea of having selective motion detection via delayed on-off interactions has been previously used to model visual motion-selective neurons in the mammalian neocortex [40]. The analysis of our “emergent non-linearity” hypothesis shows that the non-linear responses of the LGMD are caused mainly by the particular connectivity through the second chiasma and the parameters of the neurons in the network. It is the contribution of the restricted and local non-linearities in the medulla and structures pre-synaptic to the LGMD that give rise to the non-linear responses of our model. This mechanism is akin to the way a multilayer perceptron can approximate any continuous function with an arbitrary accuracy based on a distributed set of non-linearities [41]–[43].
Nonetheless, ours is not the first connectionist model proposed to explain the responses of the LGMD neuron to visual stimuli. In fact, Rind and Bramwell proposed a model that accounts for the looming sensitivity and selectivity when stimulated with approaching, translating or receding objects over a decade ago [44]. Consistent with ours, Rind and Bramwell's model is a feed-forward model with transient detectors (on and off-set sensitive neurons) and a feed-forward inhibition that brings the LGMD activity back to baseline. Moreover, Rind and Bramwell's model has been successfully applied to mobile robots [23]–[25],[45]. Although the model has been shown to provide a similar functionality to that of its biological counterpart, there are a number of aspects of LGMD computation that it does not account for since this model was proposed before many of the properties of the LGMD were unveiled. Thus, it does not address aspects such as the emergence of the angular threshold or the non-linear responses of the biological LGMD with respect to the specific properties of the visual stimulus (angular size and angular velocity).
Our model goes a step beyond Rind's model, making clear anatomical predictions on how the specific properties of the LGMD arise and showing that a non-linear interaction in the form of a multiplication between stimulus' angular size and velocity is not required to account for the known properties of the LGMD neuron. In our predictions, we test new stimulation protocols that would help us to better understand the functional aspects of the LGMD encoding of visual stimuli.
We have considered other possible, and probably simpler, explanations of the responses of the LGMD such as the idea that all the non-linear behaviours of this neuron could be driven directly by the input dynamics (see Text S1 for further details). Interestingly, as proposed by Rind and Simmons [11], the second derivative of the size of the looming stimulus displays a very similar time course to the actual LGMD responses. However, the second derivative model is unable to explain the invariance of the LGMD response since can not guarantee that the peak firing rate does always occur at the same angular size of the object (Figure S1). Although this stimulus dynamics based explanation cannot account for all the known LGMD properties, it does provide an alternative approach to explaining the LGMD response dynamics. To understand to what extent a direct linear mapping between input and output would suffice to explain the LGMD responses, a multivariate Least Squares linear regression method was used to fit our model's responses to a sequence of raw input images of an approaching object (Figure S1). This linear input-output mapping is indeed able to reproduce the responses of our LGMD model, as well as of its biological counterpart. Yet, as a linear mapping is not able to capture directional motion information, it fails to predict our model's responses when it was tested against receding stimuli. These two observations strongly suggest that the standard stimulation protocol used to study the LGMD neuron is under-constraint, and yields results that are insufficient to fully understand the input-output transformations it performs. In fact, what is needed are new stimulation protocols that independently manipulate both angular size and speed under different conditions – as in the linearly increasing object case – to demonstrate that the LGMD does compute the product of angular speed and size. Though some steps have been undertaken to investigate new stimulation protocols such as multiple simultaneously approaching objects, they do not capture all functional aspects of LGMD encoding of visual stimuli [31],[32],[41],[42].
Some of the stimulation protocols that we propose have recently been used in the context of the behavioral responses of the Locust to approaching predator like objects. In particular, the behavioral responses to looming and uniformly increasing angular size stimuli were studied when triggering an escape response [46]. In this study it was found that a hindleg flexion reaction (cocking) always occurred with a fixed delay after the stimulus reached a fixed angular size, independent of speed and type of approach of the stimulus. Moreover, the timing of this behavioral reaction changes in a linear fashion with the l/|v| ratio, as does the peak of the firing rate in the LGMD (Figure 2C, Figure 3). Nonetheless, there seems to be a discrepancy between these findings and the ones reported by [47], where this relationship was not found. If correct, the findings of Santer et al. would be consistent with the fact that the LGMD fires maximally when the stimulus reaches the angular threshold and thus with our predictions (see Results section). However, according to Gabbiani's model [18], the LGMD would not show a peak in its firing rate for uniformly expanding objects (Figure 5). Interestingly, sectioning the contralateral nerve cord (the stimulated DCMD) did not prevent cocking from occurring, but it just increased its variability [46]. Thus, there seem to be other parallel mechanisms that also contribute to this visually mediated behavioral response. These results seem to suggest that the role of the LGMD in this context is more related to timing of the escape action rather than the selection or execution of it.
Although there are valuable data on different stimulation protocols, there remains the need for a more detailed quantification if we want to pinpoint the underlying principle that gives rise to the non-linear responses of the LGMD. Specifically, to assess how the different parameters of different stimulation protocols (angular size and angular velocity) do affect the shape of the responses of the LGMD (the timing of the peak firing rate, the slope of the rising and declining phases, etc).
We have used our model to make functional, structural and testable predictions of the response of the LGMD. These predictions can help to explain the sub-linear behaviour found by Krapp and Gabbiani [48] when mapping the LGMD sensitivity to local motion stimuli, as well as aid in explaining the functional role of the post-synaptic inhibition. Recently, picrotoxin, a chloride channel blocker, was used to investigate the functional contribution of the feed-forward inhibition to the LGMD [35]. The main conclusion of that study was that the feed-forward inhibition contributes actively to the termination of the LGMD response to looming objects. This post-synaptic inhibition increases in an approximately exponential manner as the stimulus expands, and it is followed by a fast decay. These results are consistent and match the behaviour observed in our model. Yet, recent research has shown that other mechanisms can not be disregarded, such as spike frequency adaptation or synaptic plasticity, which can further contribute to the sharpening of the looming selectivity of the LGMD neuron [34],[39].
Finally, we implemented the LGMD model in the context of a behavioural robot experiment that demonstrates the reliability of the system to detect imminent collisions on a high-speed and inertial robot system. It has been shown that high frequency spikes of the LGMD are involved in triggering escape manoeuvres to lateral looming predators [50]. The responses of the LGMD have been shown to be correlated with cocking behavior [46], and to be sufficient to trigger gliding behavior [50]. Contrary to gliding, cocking is not necessarily triggered by the LGMD responses in isolation [46]. In fact, gliding has been shown to be triggered when the spikes of the DCMD summate significantly in the MN84 neuron, the second tergosternal flight motor neuron [50]. In this case the timing of the gliding responses is not directly related to the angular size of the visual stimulus, as in the case of cocking, but to high frequency activity (>150Hz) produced by the LGMD neuron. The difference between relying on the angular size of the approaching object or on high frequency activity from the LGMD supports the notion that gliding is triggered as a “last resort” when the other existing mechanisms to evade a thread fail [51],[52]. In the case of our high-speed robot experiments we have used a very similar approach to what occurs in gliding. That is, the robot only triggers an avoidance reaction when the responses of the LGMD summate over a threshold in a motor neuron responsible for the avoidance reactions (see Robot Experiments section).
Furthermore, our experiments show that the exponential transfer function of the LGMD could play an important role in minimizing the probability of false positive detection at long distances from obstacles without compromising the performance of the system. We thus propose that the exponential Vm to firing rate mapping of the LGMD may more be related to its role in the regulation of behaviour than to its putative computational role in input processing.
In the evaluation of our model we employed a twofold strategy: On the one hand, we characterized our model using protocols identical to those reported in the literature, i.e. approaching stimuli with different speeds, shapes and textures, which were displayed on a LCD screen and captured with a CCD camera. Additionally, new stimulation protocols were used to make predictions of the responses of the biological system. On the other hand, we studied the behavioural implications of our model by applying it to a high-speed robot.
The rate of expansion of an approaching object of half length l with velocity v was reproduced by a simulated looming stimulus. Any object approaching at a constant speed shows a typical slow angular speed that rapidly increases as it gets closer to the camera. The angular size of this approaching object can be described as a function of l and v, where l is the half-size of the object length and v its linear velocity.(3)Consistent with previous studies (Gabbiani et al., 2002; Gabbiani et al., 1999; Gabbiani et al., 2001), looming stimuli with l/|v| ratios that range from 5 to 50ms, with a 5ms step size, were used, with 10 to 21 repetitions for each stimulation condition. Using these stimuli we have assessed the relationship between the responses of the model LGMD and stimulus properties, including the relationship between the TTC and the l/|v| ratio and the invariance of the angular threshold of the LGMD response over the whole range of l/|v| ratios used. In these experiments the stimuli are presented as a solid shape (square) and the centre of the screen is aligned with the centre of the camera in both azimuth and elevation. Subsequently, we performed a set of measurements in order to establish the dependence of the LGMD response on the shape and texture of the stimulus using stimuli reported in the literature: a solid square, a solid circle, a square with a checkerboard pattern and a square with a pattern consisting of concentric squares [17]. Finally, we investigated the invariance of the responses of the LGMD model to the approach angle considering presentation angles of stimuli corresponding to 0%, 33%, 55%, and 75% of the visual field of the camera, where 0% represents the alignment of the camera with the screen, and at 100% the looming stimulus lies outside of the visual field of the CCD camera.
A high-end CCD camera (EVI-D31, Sony Corp., Japan) placed 10cm in front of the screen was used as input to our model. The camera was positioned such that its image covered the complete display, resulting in a visual field size of 74.65°H×56.25°V. To present the looming stimuli, a LCD screen with a resolution of 800×600 pixels was used. The spatial resolution of the screen (0.019cm per pixel) corresponds to an angular resolution of 0.0933° per pixel. The highest luminance (LHigh) value reported by our video acquisition system (mean of value for the RGB color channels) was defined as 255 and the lowest as 0 (LLow) on a 0 to 255 scale. The stimuli were generated with an ideal luminance contrast of infinity, where CR = 1 indicates no contrast. The acquisition rate of the camera was 25Hz (PAL) and the refresh-rate of the LCD monitor was set to 60Hz. For the purpose of the simulations presented here, the system was not required to run in real-time. Thus, we simulated a processing power of 100 images per second for our model. For the acquisition, we approximated a uniformly distributed compound eye of 32×24 ommatidia/photoreceptors. This is obtained by sub-sampling the image that is acquired from the camera, making the step size increase of the looming stimulus negligible. The resulting angular resolution corresponds to 2.33° per pixel, a good match to the real photoreceptor acceptance angle of the locust which is close to 1.5° in light conditions and 2.5° when dark adapted [36].
We evaluated the behavioural implications of our model using a ball caster based robot platform called “Strider”, specifically designed to have low frictional forces with the surface and that uses a propulsion system that allows it to deliver high-speeds, with the advantage of a low deployment and maintenance effort (Figure 7, left panel). The Strider is about 16cm long and it is equipped with three passive wheels (ball casters) (Euro Unit 15mm, AlwayseEngineering Ltd, United Kingdom), and propelled by two ducted fans (GW/EDF-50, Grand Wing Servo-tech Co., Ltd., Taiwan). The base platform on which the wheels are mounted connects to the upper part via a servo (Microservo FS 500 MG, Robbe Modellsport GmbH & Co, Germany), allowing the robot to turn in place, a task difficult to achieve with ducted fans alone. The lift-strength of one ducted fans is 30g, allowing the robot to move at a maximum speed of about 3m/s which corresponds to 19 body lengths per second. Similarly, the locust displays a free flight speed of about 4m/s [53].
Two separate lithium-polymer batteries (t-technik, Germany) are used as independent power-supplies for controller-board and sensors, and motors respectively. The total weight of the robot is 280g. A Bluetooth® link is used to send control signals to the motors of the robot and to read sensor states from the robot. The robot carries a wireless camera (1.2GHz Mini Wireless Camera Kit, ZTV Technology Co., Ltd, China) with a 190° wide-angle lens.
The robot experiments were performed in a 3×4m arena (Figure 7, right panel). The walls of the arena (0.5m high) were covered with random textures consisting of vertical and horizontal stripes to provide the robot with visual cues. The behavioural data was acquired in real-time with a custom-built general purpose video tracking system called “AnTS” developed by the authors. The AnTS tracking system receives its input from a B/W CCIR camera (CSB-465C, Pacific Corporation, Japan) with a wide-angle lens fixed on a 2.2m high tripod. To obtain an undistorted planar view of the arena, correction algorithms for perspective and wide-angle lens distortions were built into the AnTS tracking software. As a compromise between sampling frequency and spatial accuracy, a QVGA image resolution (320×240 pixels) was used; this resulted in a spatial resolution of 1.56cm for the 3×4m arena and an update frequency of 35Hz. The behavioural data recorded with AnTS was acquired synchronously with the states of the model of the locust visual system (see below).
Two standard neuron types are used in these simulation experiments: Leaky Integrate & Fire (I&F) and leaky Linear Threshold (LT) neurons [32],[33]. Both neuron models are equivalent to a circuit built from a capacitor C and a resistor R connected in parallel to ground on one end and driven by current on the other end [39]:(4)For a constant input current the voltage is defined by:(5)The voltage at the membrane of both neural models will increase asymptotically to . While the voltage is below the firing threshold () the neuron remains silent, and once is reached the neuron's output is equal to the membrane potential in the case of LT, or it produces an action potential (spike) and resets the membrane voltage to zero in the case of the I&F. The charging time constant of the membrane potential is defined as .
Our model captures the basic processes found in the locust visual system and can be divided into three sequential processing steps (Figure 1). First, the centre-excitation/surround-inhibition connectivity among the signals received from the photoreceptors in the lamina layer that provides an edge enhancement [31]. Second, the interaction of neurons in the medulla layer yields onset and offset sensitive responses [13]–[15]. Third, the lobula layer provides a specific connectivity that contributes to the transformation of the onset/offset signals into the response of the LGMD. Our model is structured exclusively with leaky Integrate and Fire (I&F) and leaky Linear Threshold (LT) neurons (see Experimental Procedures for the dynamic equations) and implements the three layers described above.
An edge enhancement on the input image is achieved via a centre-excitation/surround-inhibition connectivity from the photoreceptors to the lamina layer, modelled as LT neurons. Our model implements onset and offset responses of the medulla by combining the activity of one excitatory and one inhibitory neuron with the same visual sensitivities from the lamina onto a common third neuron, where the inhibition is time delayed relative to the excitation in case of onset detection, and time advanced relative to the excitation in case of offset detection (Figure 8).
When we assume that a transition of activity in the receptive fields of the on and off neurons is a moving edge, there exists a unique arrangement of on and off cells with a combined response that is maximal whenever the moving edge is being displaced in a specific direction, i.e. neighbouring cells placed along the movement axis, where a first offset sensitive cell and a second onset sensitive cell synapse onto a common neuron. The post-synaptic neuron is maximally excited only when both pre-synaptic cells are active at the same time, i.e. when an offset and an onset stimulus coincide. Hence, the only situation that can provoke this kind of response is a moving edge passing out of the off cell's receptive field, generating an offset event, to the receptive field of the on cell, generating an onset response. Therefore, as outlined above, a pair-wise combination of on and off transient detectors can encode for directionally selective motion pre-synaptic to the LGMD. This neuronal processing structure is consistent with our knowledge of the pre-synaptic structure of the LGMD since the 1970s [13]–[15].
One of the most important and most studied properties of the LGMD is related to the angular threshold (θthreshold), which is defined as the angular size of a looming object for which the LGMD produces the maximal firing rate. It has been shown that there is a constant relationship between the peak firing rate of the response of this neuron and the angular size of the looming stimulus, independent of the approach speed and angle, object shape, texture and contrast [9],[16],[17]. To account for the angular threshold properties (θthreshold) we propose a specific connectivity between the on-off cell ensembles onto the LGMD, referred to as the LGMD pre-synaptic fan (Figure 9). It is central to our hypothesis that the projections from the medulla to the lobula are such that the excitation on the target cells is maximal when the collection of detected oriented contrast boundaries reach a specific size. The LT neurons connecting the medulla with the LGMD through the second chiasma collect the activity of a set of surrounding on-off neurons in the medulla with a particular directional selectivity at distances δx and δy (Figure 9). These LT neurons have lateral interactions with the neighbouring cells via a lateral excitation that spreads and smoothes their activity over the pre-synaptic excitatory fan of the LGMD (Figure 9A). The δx and δy define the width and height of the connectivity where the expanding boundaries of a looming stimulus lie to maximally excite that post-synaptic neuron. This connectivity pattern is applied to each of the neurons that mediate the excitatory pathway to the LGMD across the second chiasma and receive input from the onset/offset sensitive cells. These neurons will concentrate a spot of high activity for looming stimuli approaching the angular threshold size whereas a sparse distribution of activity will occur for other stimuli (receding, translating, etc) (Figure 9B).
It is now possible to define the exact values of δx and δy that make the LGMD maximally excited for a given object angular size and excited below maximum otherwise. The accuracy with which we can define the angular threshold is given by the resolution of our model, being ±2 • acceptance angle of one pixel (approximately ±5°).
In our implementation of the model, only four type of ensembles of on and off neurons with different directional sensitivities were used (Figure 9A). By means of a thresholding mechanism, the LT neurons that cross the second chiasma respond only when a number of the surrounding (δx and δy) pre-synaptic motion sensitive ensembles detect expanding moving edges. Hence, by looking at the neural activity of this layer of neurons it is possible to extract the position of the looming stimulus in the visual field (Figure 9B). Subsequently, the spatial integration by the LGMD pre-synaptic fan of those responses discards the position information, and in this way introduces the important property of response invariance to object position and approach angle. The structure of the feed-forward network up to this point supports the consistency and invariance of the angular threshold (θthreshold), i.e. the independency of the approach angle, position inside the visual field, object shape and looming speed. In the last processing stage, the LGMD receives a post-synaptic inhibition from the activity of the on-off neurons in the medulla (Figure 8). The role of this inhibition is to bring the LGMD neuron's activity back to baseline after the looming object reaches the angular threshold size.
For the data analysis, a Gaussian smoothing filter with a window size of 20ms was applied to our raw data, consistent with previous LGMD studies [9]. The membrane potential of the LGMD was computed in the simulation while the used Vm/F transfer function of the LGMD neuron is consistent with the one reported in the literature [16],[17]. A one-way ANOVA analysis was used to evaluate significant differences between the data sets obtained during the experiments.
The simulations were performed on a 2GHz Pentium4 personal computer (Intel, Santa Clara, USA) under the Linux operating system. The neural simulation software iqr, an open source simulation software (iqr.souceforge.net), was chosen for the implementation and evaluation of the neural model, including the robot experiments [54]. All creation of visual stimuli was performed using openCV (the Open Source Computer Vision library, Intel, Palo Alto, USA) while the analysis was performed using Matlab (Mathworks, Natick, Massachusetts, USA).
|
10.1371/journal.pntd.0004714 | Trypanosoma brucei DHFR-TS Revisited: Characterisation of a Bifunctional and Highly Unstable Recombinant Dihydrofolate Reductase-Thymidylate Synthase | Bifunctional dihydrofolate reductase–thymidylate synthase (DHFR-TS) is a chemically and genetically validated target in African trypanosomes, causative agents of sleeping sickness in humans and nagana in cattle. Here we report the kinetic properties and sensitivity of recombinant enzyme to a range of lipophilic and classical antifolate drugs. The purified recombinant enzyme, expressed as a fusion protein with elongation factor Ts (Tsf) in ThyA- Escherichia coli, retains DHFR activity, but lacks any TS activity. TS activity was found to be extremely unstable (half-life of 28 s) following desalting of clarified bacterial lysates to remove small molecules. Stability could be improved 700-fold by inclusion of dUMP, but not by other pyrimidine or purine (deoxy)-nucleosides or nucleotides. Inclusion of dUMP during purification proved insufficient to prevent inactivation during the purification procedure. Methotrexate and trimetrexate were the most potent inhibitors of DHFR (Ki 0.1 and 0.6 nM, respectively) and FdUMP and nolatrexed of TS (Ki 14 and 39 nM, respectively). All inhibitors showed a marked drop-off in potency of 100- to 1,000-fold against trypanosomes grown in low folate medium lacking thymidine. The most potent inhibitors possessed a terminal glutamate moiety suggesting that transport or subsequent retention by polyglutamylation was important for biological activity. Supplementation of culture medium with folate markedly antagonised the potency of these folate-like inhibitors, as did thymidine in the case of the TS inhibitors raltitrexed and pemetrexed.
| There are few validated and fully characterised targets suitable for drug discovery against African trypanosomes, causative agents of sleeping sickness in humans and nagana in cattle. Here we report the biochemical properties of the bifunctional enzyme, dihydrofolate reductase–thymidylate synthase (DHFR-TS), and its susceptibility to a range of classical inhibitors normally used in the treatment of cancer, bacterial or protozoal infections. Some of these drugs are extremely potent against the isolated enzyme, but much less so against the intact trypanosome. We have found that modulating certain medium components can affect drug sensitivity, presumably by either competition for uptake and competition for the active site of DHFR-TS. In the case of one human TS inhibitor (raltitrexed) the inhibitor is more potent against the intact parasite. We propose that addition of extra glutamic acid residues not only improves retention in the cell, but also increases potency against TS, as it does in human cells.
| Human African trypanosomiasis (HAT) is an infectious disease caused by two distinct subspecies of the protozoan parasite Trypanosoma brucei (T. b. gambiense and T. b. rhodesiense). Existing therapies for this otherwise fatal disease are limited due to toxicity, difficulty in administration, emerging drug resistance and cost. As such, new safe and affordable drugs are required for the continued treatment and control of HAT. Enzymes of essential metabolic pathways in T. brucei, such as N-myristoyltransferase [1] and trypanothione synthetase [2,3], are of continuing interest as novel targets for the development of new treatments, while a number of other putative drug targets remain to be fully exploited. One example is the bifunctional folate and pyrimidine-metabolising enzyme dihydrofolate reductase-thymidylate synthase (DHFR-TS). In T. brucei, this enzyme is expressed from a single gene as a homodimer comprising of an N-terminal DHFR domain fused via a linker peptide to a TS domain at the C-terminus. In contrast, DHFR and TS are expressed separately from independent genes in many other organisms, including humans. In trypanosomatids, DHFR catalyses reduction of dihydrofolate (DHF) by NADPH to form tetrahydrofolate (THF) which is then converted to N5, N10-methylenetetrahydrofolate (CH2THF), either via the glycine cleavage system or by serine hydroxymethyltransferase (the latter is absent in T. brucei). CH2THF serves as carbon donor for the reductive methylation of deoxyuridine monophosphate (dUMP) to form thymidylate (dTMP) catalysed by TS [4]. dTMP is ultimately phosphorylated to thymidine triphosphate (dTTP) and used for DNA synthesis and DNA repair (Fig 1). T. brucei can also salvage extracellular thymidine by-passing de novo synthesis. Unlike apicomplexan parasites, trypanosomes lack the ability to synthesise folate and take up serum folate or 5-methyltetrahydrofolate via putative folate transporters. Concentration and retention of folate may involve polyglutamylation as in other organisms, although this has not been established for T. brucei.
DHFR-TS is essential for cell survival and has been previously validated (both genetically and chemically) as a potential drug target in T. brucei [6]. Despite significant evolutionary separation between protozoa and mammals, T. brucei TS is highly similar to its human homologue with 60% identity and an active site that is identical at the amino acid level. T. brucei DHFR is less well-conserved with only 28% identity with the human enzyme. Indeed, the DHFR domain from several protozoan species has been successfully exploited as a drug target, most notably in the treatment of malaria by the DHF-competitive inhibitors pyrimethamine and cycloguanil [7] which, based on their structural similarity to natural folates, belong to the class of antimetabolites known collectively as the antifolates. These compounds deplete the cellular THF pool, which in turn inhibits dTMP and DNA synthesis resulting in what is known as ‘thymineless-death’ [8,9]. To date, antifolates have not been evaluated as chemotherapeutics in animal models of HAT. Newer antifolates such as nolatrexed [10], pemetrexed [11] and raltitrexed [12] have been designed to directly inhibit TS and have proven useful as cancer chemotherapies; however, these compounds only possess low potency against trypanosomes in thymidine-rich medium [6]. In contrast to Leishmania DHFR-TS, the TS domain of TbDHFR-TS has long proven to be an elusive drug target due to an inability to express the enzyme in an active recombinant form [13], which precluded a thorough characterisation of its activity and its sensitivity to inhibitors. Here we describe the first successful recombinant production of bifunctional T. brucei DHFR-TS (TbDHFR-TS) in the form of a fusion protein incorporating Escherichia coli elongation factor Ts (Tsf) [14]. We also biochemically characterise the two activities of TbDHFR-TS and describe their sensitivities to a variety of known inhibitors, along with corresponding in vivo potencies in wild type T. brucei. We show that the challenges faced in recombinant TbDHFR-TS production are the result of instability in the TS domain, rather than proteolysis, as was previously hypothesised [13], and how a combination of TS-stabilising small molecules and macromolecules can overcome this limitation, suggesting that an as-of-yet unidentified TS-stabilising factor could be present in T. brucei and possibly other species as well. Through comparisons of in vitro and in vivo potencies of known DHFR and TS inhibitors, we also show that additional targets for these compounds remain to be identified in T. brucei.
T. brucei strain 427 was the original source for DNA used in recombinant enzyme production. All reagents were of the highest quality available from Sigma, unless otherwise specified. Recombinant protein expression employed a previously described TS-deficient (thyA-) E. coli strain [6], derived from Invitrogen BL21 Star (DE3). Restriction enzymes and Pfu DNA polymerase were from Promega. Site-directed mutagenesis was performed using the QuikChange Site-Directed Mutagenesis Kit, Stratagene. DHFR and TS inhibitors were sourced as follows: methotrexate, 5-fluorouracil, 5-fluorodeoxyuridine monophosphate (FdUMP), trimethoprim and pyrimethamine from Sigma Aldrich; nolatrexed, pemetrexed and raltitrexed from Sequoia Research Products; and trimetrexate from Tocris Bioscience.
The solubility enhancing factor Tsf [14] was engineered into a modified pET15b expression vector containing a Tobacco Etch Virus (TEV) protease recognition sequence in place of a thrombin recognition sequence (pET15b_TEV) [15]. The Tsf open reading frame was amplified by PCR from the genomic DNA (gDNA) of E. coli strain K12 using specific oligonucleotides (EcTsf_s and EcTsf_as, S1 Table) and pfu polymerase. The stop codon in the Tsf gene was replaced with a threonine-encoding ACC codon and the PCR product (866 bp) was cloned into the NcoI restriction site on the pET15b_TEV vector resulting in an expression cassette containing Tsf-(His)6-TEV. The open reading frame of DHFR-TS was amplified by PCR from T. brucei gDNA using specific oligonucleotides (TbDHFR-TS_s and TbDHFR-TS_as, S1 Table). To express TbDHFR-TS on its own or in frame with EcTsf, the PCR product (1597 bp) was cloned into the BamHI restriction site in either pET15b_TEV or pET15b_Tsf-TEV to generate the pET15b_TEV-DHFR-TS and pET15b_Tsf-TEV-DHFR-TS expression constructs, respectively. To create a pET15b_Tsf-TEV-TS fusion construct without the DHFR domain, TS (884 bp) was PCR-amplified using oligonucleotides TbTS_s and TbTS_as (S1 Table) from pET15b_Tsf-TEV-DHFR-TS and cloned into the BamHI restriction site on pET15b_Ts-TEV. To express DHFR without the TS domain, a stop codon (TAA), immediately after the last amino acid (Arg 239) of DHFR, was introduced into the above DHFR-TS expression constructs (using oligonucleotides TbDHFR_mut_s and TbDHFR_mut_as, S1 Table) by site-directed mutagenesis (Stratagene), as per manufacturer’s instruction. The accuracy of all constructs was verified by DNA sequencing (http://www.dnaseq.co.uk).
Expression constructs carrying a TS domain from T. brucei, Leishmania major and human TS (pET15b_Tsf-TEV-TbTS, pET15b_Tsf-TEV-TbDHFR-TS, pET15b_LmDHFR-TS and pET17b_hTS, respectively) were expressed in a TS-deficient E. coli strain (thyA-), while TbDHFR without TS (pET15b_TEV-DHFR) was expressed in the parental thyA+ strain. Transformants were selected on LB agar plates containing carbenicillin (50 μg ml-1). Plates were initially incubated at 37°C for 18 h and those not displaying colonies were incubated at room temperature for a further 3–6 days until colonies appeared. Single colonies were used to set up starter cultures to inoculate 1 litre of auto-induction media [16] containing 50 μg ml-1 carbenicillin. Cultures were incubated at room temperature with shaking at 200 r.p.m for 72 h and aliquots of 50 ml harvested by centrifugation (2,000 g, 10 min, 4°C). Cell pellets were stored at -80°C before use. Pellets were resuspended in lysis buffer (100 mM HEPES, 100 mM NaCl, 1 mM EDTA, 1 mM DTT, pH 7.0), lysed by sonication (3 × 30 s, 10 micron amplitude), clarified by centrifugation (20,000 g, 5 min, 4°C) and supernatants were analysed for DHFR or TS activity.
To purify the recombinant proteins, cultures (1 litre) were harvested by centrifugation (2,800 g, 30 min, 4°C), resuspended in lysis buffer containing cOmplete Protease Inhibitor Cocktail (Roche) and lysed using a cell disruptor (Constant Systems) at 30,000 psi. Lysates were clarified by centrifugation (50,000 g, 30 min, 4°C) and recombinant Tsf-TbDHFR-TS purified using methotrexate affinity chromatography, as previously described [13]. To cleave the Tsf tag, TEV protease was used in a 5:1 (mass to mass) ratio, estimated from DHFR specific activity, at 4°C for up to three days, either in assay buffer following purification or prior to purification in E. coli lysate treated with up to 40% glycerol. A methotrexate agarose column (5 ml) was loaded by recirculation, monitoring DHFR activity until the column was saturated, and then washed exhaustively with buffers consisting of 50 mM HEPES, 1 M KCl, pH 7, 10% glycerol, followed by 0.5 M KCl, until no further change in absorbance at 280 nM could be detected. Protein was eluted with one column volume of 50 mM HEPES, 0.5 M KCl, pH 8, 10% glycerol with 5 mM DHF. Up to 1 mM dUMP was added to buffers and the column operating temperature reduced to 4°C in an effort to preserve recombinant TS activity. The relative molecular mass of the cleaved recombinant enzyme was determined by size exclusion chromatography on a Superdex 200 column using Bio-Rad gel filtration standards.
All animal experiments were approved by the Ethical Review Committee at the University of Dundee and performed under the Animals (Scientific Procedures) Act 1986 (UK Home Office Project Licence PPL 60/4039) in accordance with the European Communities Council Directive (86/609/EEC).
T. brucei trypomastigotes were purified from blood of infected Wistar rats by anion exchange chromatography [17]. Parasites were resuspended (2.5 x 109 cells ml-1) in lysis buffer plus cOmplete Protease Inhibitor Cocktail (see above) and biologically inactivated by three rapid freeze-thaw cycles before lysis using a one-shot cell disruptor (Constant Systems) at 30,000 psi. Aliquots (500 μl) were stored at -80°C and clarified by centrifugation (20,000 g, 20 min, 4°C) before use.
DHFR activity was determined spectrophotometrically at 340 nm [18]. DHFR (5 nM) was pre-incubated with 100 μM NADPH (Medford) in assay buffer (50 mM HEPES, pH 7.4, containing 100 mM KCl) for 1 min at 25°C (1 ml final assay volume), before the addition of 100 μM DHF (Sigma Aldrich). Initial rates were calculated from the combined molar extinction coefficient for NADPH oxidation and DHF reduction (ε = 12,300 M-1 cm-1). The Kmapp values for DHFR substrates were determined by varying the concentration of one substrate in the presence of a fixed saturating concentration of the other. IC50 values were determined using 8-point doubling dilutions of inhibitor under the above standard assay conditions.
The initial characterisation of TS was also carried out using a spectrophotometric assay [19,20]. Owing to the pronounced instability of the recombinant protein, clarified ThyA- E. coli lysates were used for characterisation, where the concentration of TS was calculated based on DHFR activity. To determine the Kmapp for dUMP, TS (200 nM) was pre-incubated in DHFR-assay buffer containing varying amounts of dUMP (1.56–100 μM) in 1 ml assay volumes. Enzymatic reactions were initiated by the addition of CH2THF (200 μM, Shircks Laboratories) and initial rates of CH2THF oxidation to DHF monitored by the increase in absorbance at 340 nm (ε = 6,200 M-1 cm-1). This method is not suitable for determination of the Kmapp for CH2THF and a radiometric method was used instead [21]. This method measures the release of tritiated water from 5-[3H]-dUMP (American Radiochemicals, 14.3 Ci mmol-1). Assays (40 μl final volume) contained 200 nM TS and varying amounts of CH2THF (37.5 μM– 4.8 mM) in DHFR-assay buffer. Reactions were initiated by adding 200 μM [3H]-dUMP (5.55 × 105 dpm nmol-1) and stopped after 10 min by the addition of 20 μl trichloroacetic acid. Residual 5-[3H]-dUMP was removed by the addition of 200 μl of 10% (w/v) activated charcoal (Sigma). Aliquots (100 μl) of the supernatants were added to 2 ml of scintillation fluid (Pico-Fluor 40™, Packard Bioscience) and radioactivity determined using a Beckmann LS 6500 Scintillation Counter. To determine TS activity in clarified T. brucei lysates the incubation time was increased to 30 min and the assay volume was increased to ~200 μl. To assay overexpressed recombinant activity under comparable linear conditions, working stocks of bacterial lysates were prepared by diluting 20- to 40-fold with 200 μM dUMP, unless otherwise noted. For stability experiments all lysates were desalted using 0.5 ml Zeba Spin Desalting Columns (7K MWCO). Protein concentrations were determined using the BioRad protein assay based on the method of Bradford [22]. Inhibitors (8-point doubling dilutions) were assayed using the radiometric method in the presence of 100 μM CH2THF, 200 μM dUMP and 200 nM Tsf-TbDHFR-TS.
Results were analysed by non-linear regression using GraFit v 5.0.13 (Erithacus Software). Kmapp values were determined using the Michaelis-Menten equation. IC50 values were determined using eq 1 and Ki values calculated using the Cheng-Prusoff eq 2 [23]. For tight binding inhibitors, where the Hill slope of the IC50 equation was > 1, the modified Morrison eq 3 [24,25] was used to calculate Kiapp to compensate for the effective reduction in total free enzyme concentration.
Equation 1. IC50
y=100%1+(xIC50)s
(1)
(where y is the % activity remaining, x the inhibitor concentration and s the slope factor)
Equation 2. Cheng-Prusoff equation
Kiapp=Ki(1+[S]Km)
(2)
(where Km is the Michaelis-Menten constant, S is the substrate concentration and Kiapp is the apparent Ki, expressed here as IC50, see eq 1)
Equation 3. Modified Morrison equation for tight-binding inhibition
viv0=1−([E]T+[I]T+Kiapp)−([E]T+[I]T+Kiapp)2−4[E]T[I]T2[E]T
(3)
(where [E]T is the total enzyme concentration and [I]T the total inhibitor concentration).
Wild type (WT) T. brucei bloodstream-form ‘single marker’ S427 were cultured in HMI9T medium [26] supplemented with 2.5 μg ml-1 G418 to maintain expression of T7 RNA polymerase and the tetracycline repressor protein [27]. HMI9T medium (standard media for T. brucei cell culture) contains high concentrations of folate (~9 μM) and thymidine (~160 μM) principally from IMDM and 10% Serum Plus components [6]. A medium based on HMI9T, only lacking Serum Plus, folate and thymidine was prepared in-house, named trypanosome base media (TBM; the residual folate is provided by the serum component). A comparison of these media is described in S2 Table. WT T. brucei cells grow normally in TBM and the rate of growth is similar to HMI9T in TBM with no supplementation, supplementation with 9 μM folate, 160 μM thymidine or supplementation with both folate and thymidine (7–8 h doubling time).
EC50 of antifolates against T. brucei were determined in 96-well microtitre plates. Serial doubling dilutions of antifolate drugs (10–50 mM stocks prepared in DMSO) were prepared in 100 μl of the appropriate medium and trypanosomes (resuspended in the same medium) added in 100 μl to give a final concentration of 2.5 × 103 cells ml-1. All wells, including controls, contained a final volume of 0.5% DMSO. Cultures were incubated for 72 h at 37°C / 5% CO2 before cell density was determined using a resazurin-based assay [28]. EC50 values were calculated using GraFit v 5.0.13 (Erithacus Software) with a 3-parameter non-linear regression from triplicate readings. Antifolates were tested against parasites cultured in TBM; this allowed for the addition of thymidine (160 μM) and folate (9 μM) respectively.
Initial attempts to express recombinant His6-tagged TbDHFR-TS resulted in a low yield of soluble, enzymatically active protein, as previously reported (Table 1) [13]. Although DHFR activity in clarified E. coli lysates was ~100-fold above background (determined by spectrophotometric assay), the equivalent assay was insufficiently sensitive to detect any TS activity. The functionality of TS could only be confirmed by complementation studies using TS-deficient (thyA-) E. coli strain (S1 Fig). ThyA- cells were unable to grow in the absence of thymidine supplementation and growth, albeit very slow, was restored in cells transformed with the pET15b_TbDHFR-TS plasmid with bacterial colonies appearing after 3–6 days incubation. In stark contrast, numerous colonies were formed within 18 h for the positive controls of thyA+ E. coli and thyA- cells complemented with L. major DHFR-TS (S1 Fig). Re-plating of TbDHFR-TS-complemented cultures resulted in comparable numbers of colonies as seen with positive controls; however, slow growth persisted. These results, together with the undetectable activity of TS in lysates, suggest that the TS domain of TbDHFR-TS could be highly unstable. SDS-PAGE of E. coli lysates also revealed that the majority of TbDHFR-TS was present in the insoluble pellet of clarified lysates, indicating formation of inclusion bodies.
To determine whether the DHFR or TS domain was responsible for the poor solubility, constructs separating the two domains were generated, based upon previously reported boundaries [13]. Expression of the domain-specific constructs resulted in a ~10-fold increased soluble DHFR expression (Table 1), whereas the independent TS domain failed to complement thyA- E. coli (S1 Fig).
Attempts to improve the solubility of TS-active TbDHFR-TS using common fusion partners, such as NusA and thioredoxin, were unsuccessful. As an alternative approach, E. coli elongation factor Ts (Tsf) was examined as a solubility-enhancer [14]. Tsf was engineered upstream in frame with the His6-TEV site within the pET15b expression vector, to generate a Tsf expression construct (pET15b_Tsf-His-TEV). This was used as an expression cassette for the bifunctional DHFR-TS and the individual domains (Tsf-TbDHFR-TS, Tsf-TbTS, Tsf-TbDHFR).
Functional TS activity of Tsf-TbDHFR-TS was confirmed by its ability to complement thyA- E. coli cells, whereas Tsf-TbTS did not (S1 Fig). These results indicate that the DHFR domain provides a structural contribution for TS to be active, which is consistent with previous findings for Trypanosoma cruzi DHFR-TS [29] and Plasmodium falciparum DHFR-TS [30], suggesting TS is only functional when in complex with DHFR.
DHFR activity in lysates of E. coli expressing Tsf-TbDHFR-TS was ~6-fold more active than those expressing His6-TbDHFR-TS (Table 1). In addition, TS activity which had proved elusive in the His6-protein could now be detected. Crucially, the addition of dUMP to assay buffer prior to addition of the enzyme appeared to enhance TS activity. The possible role of dUMP in the stabilisation of TS was therefore further investigated. TS activity in clarified bacterial lysates was stable for up to 72 h at 4°C, whereas less than 5% TS activity was retained if small molecules and metabolites (<1,000 Da) were removed using a desalting spin column. The addition of 200 μM dUMP immediately following desalting preserved the activity of TS, consistent with several other studies reporting substrate-mediated TS stabilisation by dUMP in other organisms [31–34]. In some cases dUMP is reported to have a synergistic action with CH2THF; however, in the case of TbDHFR-TS, there was no stabilisation observed with CH2THF, possibly due to the low affinity for this substrate (see below). Other pyrimidine nucleotides, including the uracil-containing ribonucleotides and deoxyribonucleotides, and the thymidine-containing deoxyribonucleotides, were also unable to stabilise TS. Furthermore, common stabilisers such as 10% glycerol, 1% BSA and 1 mM EDTA were ineffective in preserving TS activity. To determine if the oxidation of the TS catalytic cysteine could be a reason for its inactivation, 2-mercaptoethanol (10 mM) was tested. However, not only was the reducing agent ineffective, it was found to inhibit TS activity at higher concentrations (EC50 ~100 mM), suggesting cysteine oxidation is probably not the cause of TS-inactivation. Thus, preservation of TS activity seemed to be specific for dUMP.
To characterise the stabilisation of recombinant T. brucei TS by dUMP, clarified thyA- E. coli lysate containing Tsf-TbDHFR-TS was diluted 100-fold into assay buffer containing 100 μM CH2THF and pre-incubated for different times before initiation of the reaction by the addition of dUMP (Fig 2A). In the absence of dUMP following desalting, TS activity decayed rapidly with only 1% residual activity remaining after 3 min. The data was fitted to a single exponential decay yielding a rate constant of 1.46 ± 0.11 min-1, from which a half-life of inactivation (t½) of 28 ± 2 s can be derived (Fig 2A, inset). The half maximal concentration of dUMP required to stabilise TS activity in the absence of CH2THF was 12.6 ± 2.5 μM (Fig 2B). Over longer incubation times the addition of 200 μM dUMP markedly increased the stability of the diluted enzyme by 700-fold (t½ = 330 min), but failed to completely stabilise TS activity (Fig 2A).
The stabilising effect of dUMP on Tsf-TbDHFR-TS was compared with LmDHFR-TS and human TS expressed in thyA- E. coli as controls (Fig 2A). Both trypanosomatid enzymes were found to be relatively unstable compared to the mono-functional human TS. Human TS showed little or no activity loss over a five hour period in either the presence or absence of dUMP. In contrast, LmDHFR-TS behaved more like TbDHFR-TS, but was ~20-fold more stable than TbDHFR-TS (t½ = 9.7 min). TbDHFR-TS was ultimately only partially stabilised by dUMP, as evidenced by the loss of 50% total activity over a five hour period, whereas LmDHFR-TS was effectively stable in presence of dUMP.
To establish if instability of TbDHFR-TS was also the case with the native enzyme, DHFR and TS activity in lysates of T. brucei and thyA- E. coli expressing Tsf-TbDHFR-TS were compared (Table 2). Tsf-TbDHFR-TS in desalted lysates is considerably less stable at higher temperature, hence, lysates were incubated at 37°C for 1 h prior to activity determination. Before incubation, the ratio of DHFR- to TS-activity for Tsf-TbDHFR-TS was calculated to be 53:1. Both DHFR and TS activities in the recombinant enzyme decreased drastically following incubation. Once again, the addition of dUMP improved the stability of TS while DHFR activity was unaffected. The ratio of DHFR:TS activity (5:1) of the native enzyme before incubation was in good agreement with values previously reported for T. brucei gambiense and T. lewisi lysates [35]. DHFR activity of the native DHFR-TS remained unchanged after incubation, while the decrease in TS activity was comparably less drastic compared to the recombinant enzyme. The addition of dUMP to T. brucei lysates, however, did not protect against loss of TS activity. These results suggest that the T. brucei endogenous TbDHFR-TS does not suffer the same inactivation as the recombinant enzyme. In the event that a hitherto unknown TS activating factor might be present in T. brucei lysate, native and recombinant clarified lysates were combined in a 1:1 ratio; this however resulted in no appreciable improvement in stability.
Having established the importance of dUMP in stabilising recombinant TbDHFR-TS, it was subsequently included in all buffers used for the purification of Tsf-TbDHFR-TS and its TEV-cleaved counterpart. Despite the presence of dUMP, TS activity was still completely lost during purification regardless of whether TEV cleavage occurred at the start or end of the procedure. Cleavage of Tsf-TbDHFR-TS prior to purification was only possible when glycerol was added as a stabiliser to prevent rapid protein precipitation. Cleaved TbDHFR-TS was then purified by methotrexate agarose affinity chromatography to near homogeneity in ~10% yield, along with some residual un-cleaved Tsf-TbDHFR-TS (Fig 3A). The specific activity for the purified enzyme was 24.3 U mg-1 for DHFR, with no detectable TS activity. The cleaved protein behaved as a homodimer on size exclusion chromatography (Fig 3B) and sequence identity confirmed by mass spectrometry fingerprinting with >70% sequence coverage, including the C-terminal amino acid shown to be crucial for TS activity [36]. MALDI-TOF determination of the exact total mass was not possible due to difficulties associated with desorption, although low-resolution data also confirmed the presence of dimeric TbDHFR-TS.
In comparison, affinity chromatography of the un-cleaved Tsf-tagged TbDHFR-TS resulted in a ~7-fold greater yield, suggesting Tsf significantly improved the stability of this enzyme. Following elution from the methotrexate column, some minor contaminating proteins were visible by SDS-PAGE; these were identified by mass spectrometry fingerprinting as Ef-Tu, the binding partner of the Tsf tag, and E. coli hsp90. The latter could be removed by washing the column with ATP prior to elution with methotrexate. In contrast, the control protein (LmDHFR-TS) could be purified to homogeneity with retention of TS activity (DHFR 21.2 U mg-1; TS 0.89 U mg-1).
Purified Tsf-TbDHFR-TS was subsequently used for the kinetic characterisation of DHFR domain and clarified crude lysates for the TS domain (S2 Fig). Using the spectrophotometric assay, DHFR displayed a classical bell-shaped pH-optimum profile, with an optimal pH of ~5.5 (S2A Fig). In contrast, TS had a pH optimum of 7.0. The optimal ionic strength for both enzymes required 100 mM KCl, with DHFR displaying 2.5-fold activation and TS 4.5-fold activation (S2B Fig). This is consistent with previously reported KCl-dependent activation of this enzyme, although our pH optimum profiles disagree [13]. TS could also be activated 4.5-fold with 10 mM MgCl2 (30 mM ionic strength) and this effect was not additive with activation by KCl. Higher concentrations of MgCl2 were inhibitory to TS, consistent with a previous report [37].
Since the intracellular pH of T. brucei has been reported to be 7.4 [38], a standardised assay buffer consisting of 50 mM HEPES, pH 7.4 and 100 mM KCl was used for all subsequent studies. Under these conditions, Tsf-TbDHFR-TS obeys simple Michaelis-Menten kinetics with all four substrates (Fig 4). The Kmapp of Tsf-TbDHFR-TS and Tsf-cleaved TbDHFR-TS for DHF were identical (4.1 ± 0.6 μM for Tsf-TbDHFR-TS and 4.2 ± 0.5 μM for the cleaved enzyme) demonstrating that the tag did not interfere with the enzyme activity. The catalytic efficiency for reduction of DHF was 6.8 x 106 M-1 s-1 consistent with DHFR from other organisms (Table 3). Active site titration with methotrexate confirmed DHFR activity corresponding to one site per monomer. Folic acid and the structurally related pterins (biopterin, dihydrobiopterin, sepiapterin and neopterin) were inactive as substrates for T. brucei DHFR (<17,000-fold and <2,300-fold compared to DHF as substrate for folate and pterins, respectively). The inability of TbDHFR-TS to reduce folate is in agreement with the L. major enzyme [39]; folate is presumed to be reduced to DHF by PTR1 (Fig 1).
The Kmapp for dUMP (8.2 ± 0.6 μM) is in good agreement with TS from other organisms and is consistent with the half maximal concentration of dUMP (12.6 ± 2.5 μM) required for stability of TS. In contrast, the Kmapp values for CH2THF were considerably more variable between TS from various organisms. The affinity of the T. brucei enzyme was more similar to that from the more distantly related C. fasciculata [37,41] (Table 3). The catalytic efficiency (1.5 x 103 M-1 s-1) was considerably lower compared to those reported for recombinant T. cruzi [29] and L. major [44] enzymes. These discrepancies could be due to the inherent instability of the T. brucei enzyme or due to competing metabolism of CH2THF by other enzymes in the crude thyA- E. coli lysate
Known inhibitors of DHFR and TS from other organisms were tested for their potencies against recombinant Tsf-TbDHFR-TS (Table 4). The greatest degree of TS inhibition was seen with 5-fluorodeoxyuridine monophosphate (FdUMP), a dUMP-competitive TS-specific inhibitor which displayed tight-binding inhibition. The most potent DHFR inhibitors were the classic antifolate methotrexate and the trimethoxy-substituted trimetrexate. These compounds were found to be tight-binding inhibitors with picomolar Ki values, while other antifolates exhibited linear competitive inhibition with respect to the substrate, DHF. The diaminopyrimidine antifolates trimethoprim and pyrimethamine were found to specifically inhibit DHFR, as did the diaminoquinazoline trimetrexate. Trimethoprim, pyrimethamine and raltitrexed were found to behave as potent competitive inhibitors of DHFR with Ki values of 11.4 ± 1.2, 17.6 ± 2.3 and 70.4 ± 7.2 nM, respectively, (Fig 5). These values are in good agreement with those determined using the IC50 method in Table 4. Other antifolates possessed varying degrees of both DHFR and TS inhibition. Apart from the TS substrate analogue FdUMP, the only inhibitor that possessed greater inhibition of TS than DHFR as tested was nolatrexed, showing 10-fold TS selectivity. Chemical structures of the inhibitors are shown in Fig 6.
Antifolate drugs had highest potency against bloodstream forms of T. brucei when tested in a medium deficient in folate and thymidine, with the exception of the lipophilic drug nolatrexed where potency did not change between media types (Table 5). Indeed methotrexate, pemetrexed and raltitrexed possess nanomolar potency in a thymidine and folate deficient media. The addition of folate and thymidine reduced the potencies of the antifolates, except nolatrexed. For methotrexate, pyrimethamine and trimethoprim the addition of folate had a greater effect in reducing potency than the addition of thymidine. For raltitrexed and pemetrexed the addition of thymidine had a greater effect in reducing drug potency than the addition of folate. For the lipophilic inhibitor trimetrexate the addition of folate or the addition of thymidine had a comparable effect on reducing drug potency.
The TS activity of recombinant TbDHFR-TS is highly unstable (t½ 28 s) compared to other organisms, with the T. brucei enzyme proving to be the least stable TS yet reported. Addition of dUMP increases enzyme stability, as in other organisms, but proved insufficient to achieve purification of active enzyme. Other stabilising agents, including mercaptoethanol, did not prevent inactivation, unlike human TS that can be stored at 4°C for 3 months without loss of activity [50]. This remarkable instability could account for the inefficient complementation and slow growth of TS-deficient E. coli expressing TbDHFR-TS. However, the basis for instability is not known. A previous report suggested that sequential degradations at the C-terminus together with internal cleavage in the TS domain may be responsible [13]. Our purified recombinant protein showed no evidence of proteolytic cleavage by either SDS-PAGE or MALDI-TOF MS and the C-terminus of TS was identified by MS fingerprinting, including the final residue required for catalysis. Thus, proteolysis can be discounted, as can oxidation of the catalytic cysteine since thiols did not stabilise the protein. Other possible stabilising agents include parasite-specific interacting macromolecules (e.g. mRNA or protein chaperones), or parasite-specific post-translational modifications. Further research is required to test these possibilities. A variety of DHFR and TS inhibitors were examined using bifunctional recombinant TbDHFR-TS, all of which, apart from FdUMP, can be categorised as antifolates. Overall, T. brucei DHFR appears to be more exploitable in terms of selective inhibition over the human homologue than TS.
In the current study, the only molecules to possess picomolar Ki values against DHFR were trimetrexate and methotrexate, consistent with a previous report [51]. Maximal methotrexate potency in vivo was found to be dependent on the absence of thymidine from T. brucei growth media, thus confirming that thymineless death is part of its mode of action. However, the inability of thymidine to completely reverse methotrexate toxicity suggests that additional targets also exist beyond DHFR and TS. One likely candidate is pteridine reductase 1 (Kiapp 11.1 nM) [51], another validated target in these parasites [52,53]. Compared to methotrexate, the pronounced potency of trimetrexate against DHFR is not reflected against whole cells likely due to the fact that trimetrexate is lipophilic and lacks a terminal glutamyl moiety for polyglutamylation and increased retention in the trypanosome (Fig 6).
Trimethoprim and pyrimethamine are both competitive inhibitors with intermediate potency against T. brucei DHFR. Our Ki values are 30- to 60-fold higher than first reported [13], but consistent with a subsequent report [48]. For these diaminopyrimidines, this translates to modest selectivity at best between parasite and host DHFR, compared to the > 100,000-fold selectivity of the antibacterial trimethoprim between human and E. coli DHFR [54]. A variety of novel diaminopyrimidines reported by Chowdhury et al. have been shown to be potent against T. brucei DHFR with nanomolar Ki values and selectivity up to 610-fold over the human enzyme [48]. However, toxicity and poor in vivo potency were limitations associated with these compounds. Both trimethoprim and pyrimethamine displayed a marked drop-off in potency from target to cell, when cultured in a medium deficient in thymidine and folate. Like trimetrexate these are lipophilic antifolates and do not contain a terminal glutamyl moiety. The addition of thymidine had little impact on cell potency of pyrimethamine or trimethoprim, implying that thymineless death is not their sole mode of action. This is consistent with a previous report which showed that the potency of pyrimethamine was not significantly affected by knocking out dhfr-ts [6].
With regards to antifolates possessing selectivity for TS over DHFR, compounds with a primary amine at position 4 correlated with stronger inhibition of DHFR, as was expected from previous reports, whereas a carbonyl substituent in this position is known to favour TS inhibition by means of additional hydrogen bonding provided by the oxygen atom [55]. TS-targeted antifolates also frequently include a terminal glutamyl moiety which is polyglutamylated in vivo by the enzyme folylpolyglutamyl synthetase (FPGS) [56]. This results in tighter binding to TS, with little effect on DHFR, and improved cellular retention. Although a candidate gene for FPGS is present in the T. brucei genome [57], it has not yet been studied in this species; however, in the related trypanosome Leishmania, intracellular folates possess on average 3–5 glutamates [58], thus this is likely also the case in T. brucei. Of the TS-targeted antifolates tested, the only compound which cannot exploit polyglutamylation was nolatrexed. Without the need for polyglutamylation, nolatrexed was found to be 10-fold TS selective; however, in vivo data showed moderate whole cell potencies. By comparison, the monoglutamate forms of raltitrexed and pemetrexed, generally thought of as TS-targeted antifolates in their polyglutamate forms, were found to be more potent against DHFR than TS, as has previously been observed with the monofunctional human enzymes [11,12]. Raltitrexed and pemetrexed are more potent against whole parasites than they are against DHFR or TS suggesting that polyglutamylation is likely to occur inside T. brucei. Off-target effects can be discounted since the trypanocidal action of these drugs is completely abrogated by thymidine and suggests that these inhibitors are likely to be TS-specific in vivo. Further work is required to substantiate this hypothesis.
Based on our observations regarding the inhibition of DHFR and TS by known antifolates, and their potency in vivo with regards to thymidine bypass, it is clear that additional targets must exist in T. brucei. Given their structural similarity to folate metabolites, possible alternative targets would be the FPGS, the glycine cleavage system, methionine synthase (Fig 1) or the bifunctional N5,N10-methylenetetrahydrofolate dehydrogenase-N5,N10-methenyltetrahydrofolate cyclohydrolase (DHCH). If the inhibition of one or more of these proposed targets can be identified then they could potentially be explored as alternative targets for the disruption of folate metabolism, which could be of interest for future drug discovery should TbDHFR-TS prove difficult to exploit.
|
10.1371/journal.pntd.0003980 | Taenia solium: Development of an Experimental Model of Porcine Neurocysticercosis | Human neurocysticercosis (NC) is caused by the establishment of Taenia solium larvae in the central nervous system. NC is a severe disease still affecting the population in developing countries of Latin America, Asia, and Africa. While great improvements have been made on NC diagnosis, treatment, and prevention, the management of patients affected by extraparenchymal parasites remains a challenge. The development of a T. solium NC experimental model in pigs that will allow the evaluation of new therapeutic alternatives is herein presented. Activated oncospheres (either 500 or 1000) were surgically implanted in the cerebral subarachnoid space of piglets. The clinical status and the level of serum antibodies in the animals were evaluated for a 4-month period after implantation. The animals were sacrificed, cysticerci were counted during necropsy, and both the macroscopic and microscopic characteristics of cysts were described. Based on the number of established cysticerci, infection efficiency ranged from 3.6% (1000 oncospheres) to 5.4% (500 oncospheres). Most parasites were caseous or calcified (38/63, 60.3%) and were surrounded by an exacerbated inflammatory response with lymphocyte infiltration and increased inflammatory markers. The infection elicited specific antibodies but no neurological signs. This novel experimental model of NC provides a useful tool to evaluate new cysticidal and anti-inflammatory approaches and it should improve the management of severe NC patients, refractory to the current treatments.
| Neurocysticercosis (NC) is caused by the implantation of the larval stage of Taenia solium in the human central nervous system. Although NC diagnosis, treatment, and prevention have clearly improved in the last 40 years, the disease still causes significant morbidity and mortality in endemic regions of Latin America, Asia, and Africa. In industrialized countries, the number of diagnosed cases has increased in recent years due to immigration. In this paper, we introduce a new experimental model of T. solium neurocysticercosis in pigs. Activated oncospheres were surgically implanted in the subarachnoid space of the cerebral convexity in piglets. Then, the animals were observed during 4 months. An increase in anti-cysticercal antibodies was detected, along with an inflammatory reaction surrounding the established parasites. This experimental model of T. solium NC will improve our knowledge on the pathogenesis of the disease; additionally, it will let us evaluate new promising treatments for inflammation and improve the effectiveness of cysticidal drugs.
| The larval stage of Taenia solium can establish itself in different tissues of swine and human hosts after they ingest T. solium viable eggs [1]. The adult intestinal tapeworm develops when humans consume cysticercus-infected, improperly cooked pork meat. The adult worm produces millions of eggs, which are released to the environment by the host in feces and may contaminate the water, soil, and food [2]. Endemicity is clearly related to poor hygienic standards and sanitary conditions; i.e., absence or inadequate use of latrines, open-air defecation, traditional pig farming, lack of meat inspection, inadequate water supply, and lack of drainage [2,3]. These conditions prevail in developing countries of Latin America, Asia, and Africa, where cysticercosis is endemic and poses a major health and economic challenge [4,5]. Recently, the World Health Organization (WHO), the Food and Agriculture Organization (FAO), and the UK Department for International Development (DFID) listed T. solium infection as one of the 17 neglected zoonotic diseases that can be effectively controlled [6].
In humans, the metacestode frequently establishes in the central nervous system, causing neurocysticercosis (NC), the most severe form of the disease [1]. In pigs, cysticerci are usually found both in muscle tissue and in the brain [2].
One of the main challenges in human NC is the low efficacy of anti-cysticidal and anti-inflammatory treatment when cysts are located in the subarachnoid or ventricular spaces. Frequently, anti-cysticidal drugs (albendazole and praziquantel) are only partly effective in these extraparenchymal NC forms [7–9]. Moreover, the neuroinflammation that accompanies these NC forms frequently results in arachnoiditis and vasculitis, which increase the disease severity. Currently, corticosteroids are given to NC patients to control neuroinflammation [10]. However, the administration of high corticosteroid doses administered for long periods to control neuroinflammation frequently promotes severe peripheral side effects, like steroid-induced diabetes [11]. This situation points to the need of investigating the effectiveness of other cysticidal drugs and more specific anti-inflammatory drugs to treat these patients.
In this regard, a suitable experimental model for cysticercosis will be a useful tool to search and evaluate new therapeutic options. Several experimental models have been used to study cysticercosis. An artificial infection caused by the inoculation of T. crassiceps cysticerci (ORF strain) into the abdominal cavity of mice has been the most extensively used one. This model has contributed to our understanding of the impact of immune, sexual, genetic, endocrine, and behavioral factors on the infection [12–15]. The model was also employed to test promising antigens for vaccination against T. solium, based on the cross-reactivity between T. crassiceps and T. solium antigens [16,17]. However, the intraperitoneal environment in this experimental model hardly resembles the conditions prevailing in the central nervous system. Two recent reports of intracerebral infection with Taenia crassiceps offer hope on its potential to evaluate NC treatments [18,19]. A murine intracerebral infection with Mesocestoides corti was also developed [20,21]. Nevertheless, any extrapolation of the results obtained in those intracerebral models should be made with caution, due to the differences between these cestodes and T. solium.
With respect to porcine cysticercosis, a T. solium intramuscular model has been developed, but it does not allow studying NC [22]. On the other side, orally infected pigs have been used in some studies [23–25]; unfortunately, infection rates are low and variable, particularly in brain tissue, preventing its use to study the response of brain cysticerci to treatment. Naturally infected pigs were also used in some studies, evaluating the cerebral infection by Magnetic Resonance Imaging (MRI) [26]. While this approach is interesting, access to MRI in endemic countries is restricted even for humans, and MRI studies in pigs are not feasible.
Considering the limitations of the available experimental models, the results obtained in the development of a porcine NC model are herein presented.
This study was approved by the Institutional Animal Care and Use Committees of the Facultad de Medicina Veterinaria y Zootecnia (FMVZ), UNAM and of the Instituto Nacional de Neurología y Neurocirugía, Mexico. All guidelines in the Official Mexican Norm (NOM-062-ZOO-1999) on the technical specifications for the production, care, and use of laboratory animals were followed.
The adults that were treated to find the pork tapeworm (Taenia solium) provided written informed consent. In cases of minors, written informed consent from the person in charge of the minor was also required before any intervention.
Twenty-four 2-month-old crossbred York-Landrace piglets of different sexes were purchased from a technically operated, cysticercosis-free farm, and then transferred to the FMVZ, UNAM, to be employed in the two experiments reported in this study. All animals were bled before infection and every 20 days after infection until sacrifice; sera were separated and frozen until used for immunological tests.
Adult Taenia solium worms were retrieved from human patients living in Mexican rural endemic areas, who reported proglottid expulsion.
Oral treatment with niclosamide (Bayer, S.A., Mexico) in a single 2-g dose, followed by intestinal purge one hour after, was administered. Once obtained, adult tapeworm specimens were macro- and microscopically inspected to distinguish between T. solium and T. saginata, based on morphological characteristics. Species was confirmed by PCR using a previously described procedure [27].
Afterwards, T. solium proglottids were washed three times with physiological saline solution and maintained in PSS (sodium polystyrene sulfonate) and PBS (phosphate-buffered saline) with antibiotics (penicillin-streptomycin) at 4°C until use. Parasites were conserved under refrigeration for 4 weeks in the first experiment and for 7 days in the second one.
A few hours before surgery, in vitro egg hatching was performed under sterile conditions using 0.75% sodium hypochlorite in water as previously reported [28]. After extensive washing with PBS, subsequent oncosphere activation was slightly different in the two experiments. In the first one, RPMI-1640 added with 10% trypsin and 5% pig bile was used, while artificial intestinal fluid with 1% pancreatin, 0.2% anhydrous sodium carbonate, 10% trypsin and 0.5% pig bile in RPMI-1640 (Gibco) was used in the second one. In both trials, oncospheres were incubated in a water bath for 1 h at 37°C, shaking the oncospheres every 15 minutes. At the end of incubation, activated oncospheres (those with detectable movement) were counted and their viability was assessed using trypan blue (Sigma). Oncosphere activation and viability was about 20% and 85% for the first and the second experiment, respectively. For surgical implantation, activated oncospheres were thoroughly washed with sterile saline solution to eliminate all enzyme content.
Animals were anesthetized by intramuscular administration of xylazine (2.2 mg/kg), ketamine (2.2 mg/kg), and Tiletamine-Zolacepam (4.4 mg/kg) followed by isoflurane. Craniotomy was performed for subarachnoid and ventricular oncosphere implantation. Craniotomy was centered at the intersection planes located 2 cm above the external auditory canal in the coronal plane and 2 cm lateral to the midline. For subarachnoid implantation, sulcus dissection was done and oncospheres were implanted at the deep sulcus surface. For ventricular implantation, oncospheres were delivered via direct puncture at the frontal horn of lateral ventricle with a latex catheter (internal diameter 6 Fr). The volume of sterilized saline solution inoculated was 100 μl for 500 oncospheres and 200 μl for 1000 oncospheres.
In the first experiment, seven piglets were inoculated in the subarachnoid space; five received a high dose of oncospheres (1000–1500) and two received a low dose (100–150). Five animals were inoculated via ventricle, two with a high dose and three with a low dose. In the second experiment, eight piglets were inoculated in the subarachnoid space, four with 500 oncospheres and four with 1000 oncospheres. As controls, two pigs were inoculated in the muscle of the right rear leg, one with 500 and one with 1000 oncospheres, and two animals were operated but not inoculated (sham controls).
After surgery, all animals were kept in the facilities of FMVZ, UNAM for 4 months. Clinical status was checked daily, and blood samples were collected every 20 days by puncture of the anterior vena cava. Samples of cerebrospinal fluid (CSF) were also retrieved during surgery and at necropsy.
Serum samples from subject pigs were collected before and after infection to measure specific IgG antibodies and the HP10 secretion antigen by ELISA. All samples were run in duplicate.
Anti-cysticercal antibodies and the HP10 antigen were detected by ELISA, as previously described [29]. Briefly, for antibody detection, polycarbonate Immulon I plates (Nunc, Roskilde, Denmark) were sensitized with 1 μg/well of T. solium cysticercal antigens (TsAg) in carbonate buffered saline, pH = 9.6 overnight at 4°C. The plate was washed and blocked with 200 μl PBS containing 1% w/v bovine serum albumin and 0.3% v/v Tween 20 and left for 60 min at 37°C. Serum samples were diluted 1:100 and the reaction was detected with 100 μl/well of HRP-goat anti-pig IgG (Fc) (Serotec) diluted 1:60,000. The reaction was developed with 100 μl/well of tetramethylbenzidine (TMB) (Zymed, San Francisco, California, USA) for 11 min at 4°C in the dark and stopped by adding 100 μl 0.2 M H2SO4 (Baker, Estado de Mexico, Mexico). OD values were measured at 450 nm in an ELISA reader (Opsys MR Dynex Technology, Chantilly, Virginia, USA).
For HP10 antigen detection, Immulon I plates (Nunc, Roskilde, Denmark) were sensitized with 1 μg/well of McAb HP10 diluted in 0.07 M saline buffered with 0.1 M borate, pH = 8.2, overnight at 4°C. Plates were washed four times with 200 μl/well of 0.15 M saline containing 0.05% v/v Tween 20; then they were blocked with 200 μl/well of PBS containing bovine serum albumin 1% w/v and 0.05% v/v Tween 20, and left for 60 min at room temperature. Later, 100 μl/well of undiluted serum samples were added and incubated for 30 min at 37°C, followed by incubation with biotinylated McAb HP10 diluted 1:500 for 30 min at 37°C and 100 μl/well of streptavidin-peroxidase conjugated (Amersham Ltd), diluted 1:4000 and incubated in the same conditions. The reaction was developed by adding 100 μl/well of tetramethylbenzidine (TMB) (Zymed, San Francisco, California, USA) for 30 min at 4°C in the dark and stopped with 100 μl/well of 0.2 M H2SO4 (Baker, Estado de Mexico, Mexico). OD values were measured at 450 nm in an ELISA reader (Opsys MR Dynex Technology, Chantilly, Virginia, USA).
A sample was considered as positive for HP10 and anti-cysticercal antibodies if the mean OD value at 450 nm was higher than the cut-off value, which was set based on the mean OD plus 2 SD in serum before infection. Cut-off for antibodies was 0.3, while cut-off for HP10 was 0.22.
Four months after infection, the pigs were humanely killed and all tissues were inspected. The brains were extracted and macroscopic examinations were performed to detect external parasites. Afterwards, the brains were sliced for histological studies [30]. Parasites were regarded as vesicular when cyst membranes were thin and the liquid content was clear. In the colloidal stage, cyst membranes thicken and the liquid within the cyst turns opaque.
The muscles where cysticerci were inoculated were also inspected, and the number and degenerative stage of cysticerci were registered.
Tissue samples taken at necropsy were fixed in Zamboni solution (1.6% [w/v] paraformaldehyde, 19 mM KH2PO4 plus 100 mM Na2HPO4·7H2O in 240 ml of saturated picric acid and 1600 ml H2O), embedded in paraffin and stained with hematoxylin-eosin. Fifteen days after, 0.5-μm slides were prepared and microscopically observed to identify and evaluate the size, location, and degenerative stage of cysticerci. The degree of inflammatory reaction surrounding each cyst was evaluated according to the scale by Vargas and de Aluja, 1988 [30]. Briefly, this classification describes seven inflammation grades according to the cellular characteristics of the tissue surrounding parasites and the traits of parasites themselves. Grade 0: No inflammatory reaction. Grade 1: Discrete focal infiltration, mainly of lymphocytes, plasma cells, and eosinophils. Grade 2: Increase of infiltration, predominating lymphocytes and plasma cells; eosinophils are numerous and macrophages appear. Grade 3: The same cell populations are present; eosinophils adhere to the vesicular wall of the parasite, and the tegument becomes swollen and vacuoles appear within; macrophages begin to line up in a palisade pattern. Grade 4: The inflammatory reaction surrounds completely the parasite and the aggregates of lymphoid cells are larger; the capsular tegument shows marked degeneration and the bladder cavity is filled with acidophilic material and necrotic cells. Grade 5: The parasite is completely degenerated; lymphocytes, plasma cells and eosinophils are less frequent. Grade 6: Inflammatory cells are scarce.
Brains were fixed in Zamboni solution for at least one week. Afterwards, specimens were dehydrated and embedded in paraffin, and 5-μm sections were cut. Endogenous peroxidase was inhibited by incubation with 0.3% (v/v) H2O2 in PBS for 10 min. After washing twice with PBS, heat-mediated antigen retrieval method was performed by microwave treatment with 0.1 M sodium citrate solution (pH 6.0) for 5 min. Then, slides were rinsed three times in PBS buffer and sections were preincubated in a blocking solution consisting of 2% BSA (bovine serum albumin; Sigma-Aldrich) for 30 min. After two washes with Tris-EDTA buffer, sections were incubated with the primary antibody (described below) diluted in PBS buffer overnight at 4°C. After washing three times in PBS/A-T (1% BSA in PBS, plus 0.1% Triton X-100), 5 min each, slides were covered with secondary antibody conjugated with horseradish peroxidase (Dako-Kit) for 30 min at 37°C and rinsed with PBS/A-T. Peroxidase activity was visualized by incubating the samples for 2 min with 3-diaminobenzidine tetrahydrochloride (DAB, DAKO). Reaction was stopped with water, and sections were counterstained with hematoxylin, dehydrated, cleared, and mounted with permount (Fisher Scientific). The single labeled sections were examined by light microscopy Leica Galen III, and digital color video camera (SSC-DC14), on a Pentium IV, Windows 2000 computer.
The primary antibodies used in this study recognized: glial fibrillary acidic protein (GFAP; polyclonal rabbit DAKO, Glostrup, Denmark 1:100 dilution), vimentin (mouse clone V9, DAKO; 1:100 dilution), neuronal nuclear protein (NeuN; mouse clone MAB377, IgG; Chemicon, Temecula CA, USA; 1:1000), nestin (mouse anti-nestin monoclonal antibody, 1:100; Chemicon, Millipore Billerica, MA, USA), IL-4 (anti-human monoclonal antibody, 1:200 dilution, Biolegend), IL-6 (anti-human monoclonal antibody, 1:250 dilution Biolegend), IL-10 (anti-human monoclonal antibody, 1:100, Biolegend), IL-17A (anti-human monoclonal antibody, 1:150 Boise’s), TNF-α (anti-human monoclonal antibody, 1:500, Biolegend), CD54 (anti-human monoclonal antibody, 1:200 Biolegend), CD69 (anti-human monoclonal antibody, 1:200 Biolegend), CD80 (anti-human monoclonal antibody, 1:300 Biolegend), CD106 (anti-human monoclonal antibody, 1:250, Biolegend).
Anti-inflammatory (IL-4), immunoregulatory (IL-10), and proinflammatory (IL-6, IL-17A, TNF-α) cytokines were evaluated. CD54 was used to assess the expression of intercellular adhesion molecule-1 (ICAM1); CD69 was used to evaluate the activation of T lymphocytes; CD80 was used to evaluate the activation of B lymphocytes; and CD106 was used to evaluate the expression of the vascular cell adhesion protein-1 (VCAM1). The expression of glial fibrillary acidic protein (GFAP) and vimentin evaluates the activation status of astrocytes; nestin expression indicates the presence of immature neurons; and NeuN expression shows the presence of viable neurons.
Brain sections from distal areas to the parasite location or brain sections from sham-operated pigs were used as control.
Statistical analysis was performed with the SPSS software. Student’s t-test was used to evaluate statistical differences between means. Linear regression was used to test changes in OD values between sampling times.
One pig from the group inoculated with a higher dose of oncospheres in the subarachnoid space showed fever and ataxia 20 days after the challenge. In spite of the treatment with antibiotics and anti-inflammatory drugs, its neurological status worsened and it was killed 30 days after infection. An abscess and a vesicle of 0.2 cm × 0.05 cm in size were found in the brain by macroscopic examination. Under microscopic examination, the vesicle was confirmed as a vesicular cysticercus surrounded by inflammatory reaction consisting of lymphocytes, plasma cells, and macrophages.
No other animal in this first experiment showed any neurological sign until sacrifice, 4 months after infection. A single parasite was observed in two pigs, both of which were inoculated with a higher dose of oncospheres in the ventricular system. In the first case, the cyst was located in the subarachnoid sulcus of the parietal lobe, near to the cortex. It was a vesicular cyst, 0.5 cm in diameter; it presented a scolex, as confirmed by microscopic examination. A grade-4 inflammatory reaction was observed around the cyst. In the second case, a colloidal cyst was located in the subarachnoid space of the convexity, in the left hemisphere near the interhemispheric fissure. It was 0.4 cm in diameter and it was surrounded by a grade-4 inflammatory reaction. No parasite or inflammatory reaction was detected in any other animal.
During this experiment and until sacrifice, pigs were in good health condition, with no neurological signs. Macro- and microscopic examination revealed the presence of cysticerci in all animals. Cysts were located either in the subarachnoid space when the infection was induced in this compartment or in muscles when oncospheres were intramuscularly inoculated (Figs 1 and 2). All parasites were located in the proximity of the site of infection.
As shown in Table 1, 63 cysticerci from the brain and 11 cysticerci from the muscle were macroscopically examined. The number of cysts in the brain was higher when 1000 oncospheres were inoculated than those recovered when 500 oncospheres were used (36 versus 27). The efficiency of infection (number of developed cysticerci / number of oncospheres inoculated) was higher when lower doses of oncospheres were inoculated (3.6% with 1000 oncospheres versus 5.4% with 500 oncospheres, P = 0.1). The proportion of vesicular parasites was higher when higher doses of oncospheres were inoculated (33.3% (9/27) with 500 oncospheres versus 44.4% (16/36) with 1000 oncospheres), but the difference was not statistically significant (P = 0.44). The number of cysticerci found in the muscle was similar (5 versus 6), disregarding the dose of oncospheres inoculated; all cysts recovered from the muscle were in colloidal/caseous stage. The difference in the proportion of vesicular cysticerci found in brain (34%) and muscle (0%) was significant (P = 0.01).
The results of the histopathological examination of brain tissue samples and parasites are shown in Table 2 (Fig 2). Most cysticerci were in a degenerating stage and were surrounded by an exacerbated inflammatory reaction (grade 4). No difference was observed in the inflammatory grade in the two groups of pigs (500 versus 1000 oncospheres inoculated). When comparing grade-2 and grade-3 inflammation versus grade-4 and grade-5 inflammation, 33.3% of animals (8/24) in the 500 oncospheres group versus 34.6% of animals (9/26) in the 1000 oncospheres group showed a low inflammatory reaction (grades-2,3; P = 1).
The immunohistochemical analysis of the brain tissues surrounding cysticerci showed a high expression of CD80, CD106, CD69, and IL-10; a moderate expression of IL4, TNF-α, and CD54, and a low expression of IL-17 and IL-6 with respect to brain tissues from sham-operated pigs (Figs 3 and 4). GFAP and vimentin (indicators of immature astrocytes) were highly expressed in tissues proximal to the parasite, but expression lowered in tissues distal to the parasite (Figs 5 and 6). In general, the expression of NeuN (viable neurons) and nestin (immature neurons) was low, particularly in tissues proximal to the parasite, and increased slightly in tissues distal to the parasite (Fig 5).
The level of anti-cysticercal antibodies and of the HP10 antigen in serum were measured at seven times during the experiment (Fig 7A and 7B, respectively). Samples were positive to antibodies at day 53 post-infection (PI) in all the animal groups (intramuscular and subarachnoid inoculation). Although a decrease was observed at the time of sacrifice, all sera remained positive during the experiment (Fig 7A). Antibody levels were significantly higher in the groups inoculated in the subarachnoid space (SA) (500 oncospheres, R = 0.75, P = 0.05; 1000 oncospheres, R = 0.80, P = 0.03), while only a tendency was observed in the group inoculated intramuscularly (IM) (R = 0.73, P = 0.06). The animals intramuscularly challenged exhibited higher antibody levels. Significant differences in OD values between groups of animals were observed at different times: at day 53 PI, control versus inoculation of 500 oncospheres in SA (P = 0.01); at day 81 PI, control versus IM inoculation (P = 0.02), controls versus inoculation of 1000 oncospheres in SA (P = 0.03), and inoculation of 500 oncospheres in SA versus IM (P = 0.02); at day 123 PI, control and IM inoculation (P = 0.001). Antibody levels did not increase in control pigs during the experiment (R = 0.48, P = 0.27).
With respect to HP10 serum levels (Fig 7B), while the three groups of pigs challenged with oncospheres were positive at the time of sacrifice, no significant increase was observed during the four months of the experiment (R = 0.62, P = 0.14; and R = 0.31, P = 0.49 when 500 and 1000 oncospheres were inoculated in SA respectively). Only in the IM inoculated group a significant increase was observed (R = 0.88, P = 0.008). OD values at each sampling time were not significantly different between groups.
The feasibility of infecting pigs by surgical implantation of activated oncospheres in the central nervous system is explored in this study.
Two individual experiments were performed, with important differences with regard to the efficiency of establishment. While only three parasites developed successfully in the first one, the implantation of multiple cysticerci in the central nervous system was observed in the second. These discordant results could be due to differences on the viability of the oncospheres used in each experiment. In the first experiment, the tapeworm was kept in refrigeration for 4 weeks before use, and only 5% of the oncospheres were activated when evaluated before implantation. In contrast, the tapeworm was kept for less than one week in refrigeration before use in the second experiment, and 85% of the oncospheres were activated before implantation. On the other hand, slight variations in the activation protocol of the oncospheres in both experiments could have had an impact on infection efficiency. For instance, it is likely that the trypsin added to pancreatin in the second experiment played a role in the higher oncosphere activation rate. These results point to the need of using oncospheres with high activation rates to achieve a higher infection efficiency. However, since both variables (the time elapsed between adult taenia recovery and infection, and the activation protocol) were changed in the same experiment, the main factor underlying these differences could not be determined in this work.
At necropsy time, 4 months after inoculation, most cysticerci were degenerating, with an exacerbated inflammatory reaction surrounding them. However, the proportion of vesicular cysticerci was significantly higher in the brain with respect to the cysts in muscle (P = 0.01). This observation is in accordance with previous works in naturally infected pigs [24,25]. It is likely that the higher destruction rates of muscle cysticerci are related to a more effective inflammatory response in the periphery. In contrast, the magnitude of the inmmunoinflammatory response is tightly regulated in the central nervous system, and cysticerci may remain for longer periods without damage [31]. However, most parasites located in the central nervous system were also in a degenerative stage. In humans, parasites reaching the brain are also mostly calcified at diagnosis, as shown in CT-scan-based epidemiological studies [32,33]. However, it is also known that brain cysticerci may persist vesicular for months or even years in humans [34]. It is thus probable that neurosurgery in our model generate additional inflammation, which could be promoting parasite degeneration. It is also possible that the hatching/activation process in the gastrointestinal track during natural infection confers some additional property to the oncospheres that will let them survive for longer periods in the central nervous system. Disregarding the factors involved in cysticercal destruction, this observation will be important to take into account when further experiments are performed to evaluate potential treatments. Indeed, the time elapsed between oncosphere implantation, treatment, and sacrifice should be shortened when comparing different therapeutic schemes.
In this study, we found that the infection efficiency was higher when lower doses of oncospheres were used in the challenge (3.6% using 1000 oncospheres and 5.4% when 500 oncospheres were inoculated, P = 0.1). On the other hand, the proportion of vesicular parasites was higher when higher oncospheres doses were implanted in the brain (44.4% versus 33.3%, P = 0.44). Similar tendencies were previously reported in a dose-response study performed in pigs [25]. These findings could reflect the fact that the immune mechanisms involved in the control of oncosphere establishment and in cysticerci destruction are different.
The inflammatory reaction surrounding most of the parasites was evident and most cysticerci were at grade 3 and 4 in the grading system by Aluja & Vargas [30]. Cells were mostly lymphocytes, plasma cells, macrophages, and eosinophils, a finding in accordance with the result of natural infections [35]. Immunohistochemical analysis showed a high lymphocyte activation, as revealed by the high expression of the CD69 activation marker and by the activation of astrocytes in parasite proximity. The high expression of CD106 (the vascular cell adhesion molecule VCAM-1) may result from the presence of TNF-α, which is known to increase the expression of adhesion molecules in endothelial cells, favoring the adhesion of peripheral leukocytes to enter the brain and therefore promoting brain inflammation [36]. It is noteworthy that the expression of proinflammatory cytokines (IL-6, IL-17) was not especially high, contrasting with the high expression of the regulatory cytokine IL-10. It is possible that regulatory factors secreted by the parasite accounted for this observation. Indeed, several molecules with potentially immunomodulatory functions have been found in the recently reported genome of Taenia solium [37]. The expression of the neuronal markers NeuN and nestin was low, although it increased slightly in brain tissues distal to the parasite with respect to tissues proximal to the parasite; this finding could reflect the neuronal damage in the cysticercus proximity. The high expression of mature and immature astrocyte markers in areas proximal to the parasite demonstrates the inflammatory reaction surrounding the parasite, which includes astrocyte activation.
Even though an assessment of the specificity of these changes was not the purpose of this work, it would be interesting to evaluate them by implanting other parasites in future experiments.
Serum specific anti-cysticercal antibodies increased in all groups of infected pigs, an additional result that demonstrates the close connection between the central and the peripheral immune system, as it has been extensively shown [38]. However, it is interesting to note that IM infection in experiment 2 elicited significantly higher antibody levels, than those observed in SA infected pigs at day 81 post-infection. The higher systemic antibody levels induced by peripheral infections may be due to the presence of multiple B cell follicles able to detect the presence of parasite antigens and promoting differentiation to plasmatic B cells, with the ensuing production of specific antibodies. On the other hand, the specific antibodies detected in the central nervous system may be produced in the periphery and then be centrally recruited due to some disruption in the blood brain barrier of cysticerci-infected pigs. The local intrathecal production of anti-cysticercal antibodies can also contribute to the detected central antibody levels.
None of the infected animals in experiment 2 exhibited neurological manifestations. This aspect, previously reported in other works [39] is noteworthy. In humans, several epidemiological studies have showed that an important proportion and in some cases most of the infected subjects are asymptomatic [32,33]. This same phenomenon could occur in pigs, and the fact that no neurological signs were detected could be due to the small number of pigs included in this study. However, it is interesting to note the high expression of the regulatory cytokine IL-10, while the frequency of IL-10-producing cells was relatively low [40]. Thus, it is possible that the absence of neurological symptoms in pigs be also due to a more active immunomodulatory process in this species. Disregarding the cause underlying the absence of clinical signs in pigs, this situation will not allow us to use this model to study the neurological aspects of the disease.
It should be noted that the model reproducibility was not evaluated in this study. Ten animals were used in the successful second experiment, and the issue of reproducibility should be addressed soon.
Finally, it is important to note the limitations of this model, mainly due to logistical requirements. Since adult pigs are not easy to manipulate, piglets were employed instead. This aspect is important, as our results could not reflect the infection in adult humans. Additionally, it has been shown that the clinical picture in human neurocysticercosis is quite different between children and adults [41]. On the other side, obtaining a Taenia solium tapeworm is not effortless, and the requirement of implanting eggs shortly after acquiring the taenia specimen limits its use to endemic countries. Furthermore, the model requires a technical team to care for animals after surgery, and the coordinated efforts of specialists (surgeons, veterinary personnel, parasitologists, immunologists, and medics) who might not be accustomed to work together. Nevertheless, we consider that all these difficulties need to be solved, since the results of studies using naturally infected pigs are often difficult to interpret, with the animals being not only infected in the central nervous system but also in the muscles, a fact that can change their immunological status.
In spite of its limitations, the new model for neurocysticercosis that we are proposing will allow for evaluating different therapeutic approaches that eventually could be employed to treat human neurocysticercosis.
|
10.1371/journal.pntd.0005414 | Seeking the environmental source of Leptospirosis reveals durable bacterial viability in river soils | Leptospirosis is an important re-emerging infectious disease that affects humans worldwide. Infection occurs from indirect environment-mediated exposure to pathogenic leptospires through contaminated watered environments. The ability of pathogenic leptospires to persist in the aqueous environment is a key factor in transmission to new hosts. Hence, an effort was made to detect pathogenic leptospires in complex environmental samples, to genotype positive samples and to assess leptospiral viability over time.
We focused our study on human leptospirosis cases infected with the New Caledonian Leptospira interrogans serovar Pyrogenes. Epidemiologically related to freshwater contaminations, this strain is responsible for ca. 25% of human cases in New Caledonia. We screened soil and water samples retrieved from suspected environmental infection sites for the pathogen-specific leptospiral gene lipL-32. Soil samples from all suspected infection sites tested showed detectable levels of pathogenic leptospiral DNA. More importantly, we demonstrated by viability qPCR that those pathogenic leptospires were viable and persisted in infection sites for several weeks after the index contamination event. Further, molecular phylogenetic analyses of the leptospiral lfb-1 gene successfully linked the identity of environmental Leptospira to the corresponding human-infecting strain.
Altogether, this study illustrates the potential of quantitative viability-PCR assay for the rapid detection of viable leptospires in environmental samples, which might open avenues to strategies aimed at assessing environmental risk.
| Leptospirosis is an emerging zoonotic disease caused by infection with pathogenic strains of Leptospira. Most human infections arise from environmental exposure to contaminated freshwater environments or watered soils where pathogenic Leptospira are considered as able to survive for prolonged periods. Therefore, a good understanding of Leptospira survival strategy in the environment is a key step to identifying crucial factors amenable to interventions and public health actions to lower leptospirosis burden. In this study, we investigated the environmental presence and survival of pathogenic leptospires in areas where recent human leptospirosis cases had been reported. Although detection of Leptospira from complex environmental samples is difficult, we successfully detected the presence of pathogenic Leptospira in soils of suspected infection sites. In addition, we showed that these pathogenic leptospires were alive and present in soils several weeks after the infecting event. Typing of leptospiral DNA retrieved from the environment revealed identities between environmental pathogenic Leptospira and the causative strains involved in human leptospirosis index cases. Interestingly, we also identified yet unreported genotypes. Altogether, our work illustrates the potential of quantitative molecular assays for the rapid detection and typing of viable leptospires in environmental samples, which could prove useful to assess the risk of environmental exposure.
| Leptospirosis is an acute febrile disease caused by pathogenic spirochetes of the genus Leptospira. It is considered an important re-emerging infectious disease that affects more than 1 million humans worldwide [1]. The spectrum of human disease caused by leptospires is extremely wide, ranging from subclinical infection to a severe syndrome of multiorgan infection with high mortality. Leptospira transmission from the urine of reservoir hosts to incidental hosts, including humans, usually occurs through the contamination of skin lesions or mucosae with contaminated surface water or soil [2]. The incidence of such infections depends on several factors including the density of the reservoir species and its Leptospira carriage prevalence, the dilution into watered environment and the survival time of the leptospires into possibly nutrient-poor and adverse environmental conditions. Estimation of survival time and virulence preservation of pathogenic Leptospira spp. after excretion into the environment is becoming a crucial challenge to determine the environmental risk and to adopt preventive measures. The duration of Leptospira survival in natural habitat is affected by many factors including abiotic and biotic factors. The persistence of pathogenic Leptospira in moist soil and freshwater for long periods of time is thought to depend on a slightly alkaline pH, high oxygen, and low salt concentrations [3–5]. The classical assumption is that slightly higher alkalinity (up to pH 8.0) allows for longer survival. Under laboratory conditions, a strain of serovar Javanica was reported to survive in distilled water (pH 7.8) for 152 days [6]. More recently, Andre-Fontaine et al. [7] showed that pathogenic Leptospira can survive for months in mineral water. Interestingly, Leptospira were reported to survive as long as 10 months in adverse conditions (4°C) and up to 20 months when stored at 30°C.
Interactions of Leptospira spp. with the environmental microbiota also begin to be examined. Environmental microbial blooms alter the concentration of oxygen, minerals, and other nutrients in the water and favor either multiplication or destruction of some species of pathogenic Leptospira [8]. Several common bacterial genera including Azospirillum and Sphingomonas were found along with pathogenic and saprophytic Leptospira spp. in biofilms formed in freshwater or in dental water unit systems [9,10]. Co-incubation with a Sphingomonas spp. increased Leptospira growth rate [8], suggesting possible syntrophic interactions. When incubated with Azospirillum brasilense, viability of pathogenic Leptospira was enhanced at high temperature and extended under UV radiation or exposure to penicillin G, tetracycline or ampicillin. In addition, soil adsorption, thought to be an important step that favors leptospire persistence in the environment, was greatly increased in the presence of A. brasilense [8].
A major impediment to assess environmental risk for leptospirosis has been the difficulty to isolate pathogenic Leptospira from environmental samples, attributable in part to the fact that non-pathogenic leptospires outgrow pathogenic strains in culture. Other methods including direct animal inoculation are time-consuming, ethically questionable and have a low analytical sensitivity. However, the increasing use of molecular methods overcomes some limitations inherent to culture- and animal-based methods and provides quantitative information about the concentration of leptospires in contaminated waters [11–13].
New Caledonia provides an ideal location for studying environmental risk factors of leptospirosis because of its high leptospirosis incidence, on average 45 cases per 100,000 inhabitants, and the presence of known hot spots where annual incidence reaches up to 500 cases/100,000 population. Based on data of leptospirosis surveillance in New Caledonia, serogroup Icterohaemorrhagiae is the dominant serogroup involved in ca. 60% of human cases. Other serogroups involved in human leptospirosis include Pyrogenes (18–25%), Ballum, Australis and Pomona. Interestingly, the New Caledonian L. interrogans serovar Pyrogenes was formerly shown to be epidemiologically related to freshwater contaminations. Therefore, human leptospirosis cases infected with this strain provide opportunities to investigate the persistence and survival of pathogenic Leptospira in natural habitats.
The purpose of the present study was to assess the presence of pathogenic leptospires in environmental samples and to estimate their viability over time. Using a TaqMan-based real time quantitative polymerase chain reaction, we screened 73 environmental samples retrieved from 4 suspected environmental infection sites for the pathogen-specific leptospiral gene lipL-32. This study found that a large proportion of soil samples were positive for pathogenic leptospiral DNA, suggesting that repeated exposure to Leptospira may be occurring in these high-risk areas. Herein, we report findings from retrospective investigations of environmental contaminated areas to assess the presence of pathogenic Leptospira in order to better delineate and monitor high risk areas.
Four sites were identified according to the infectious strain and the good acceptance of the project by the patients and custom chiefdom (Kaala-Gomen, Koné, Touho (2 sites), Fig 1). All four study sites were within Melanesian tribes, where many outdoor activities are part of the everyday life, including fishing and bathing in freshwater streams, maintenance of backyard pig pens, hunting (deer and wild hogs). In addition, two extra sites where L. interrogans Pyrogenes was known to have been involved in former cases but where no recent contamination were reported were chosen as control sites and investigated according to the same sampling procedure. Most of the investigated sites were located in the North province of the main island where climate is sub-tropical and oceanic with a hot and rainy season from December to March (average temperature 28°C) and a cooler season from June to September (average temperature 20°C). Annual cumulative rainfall is 2400 mm on average but can range from 1460 mm to 3550 mm. Daily rainfall data for each site were obtained from the Météo France free online public database, using the nearest meteorological station for each study site.
Environmental investigations were started 6 to 10 weeks after the supposed infection date. Between March and June 2016, a total of 73 environmental samples were collected: 10 water samples, 52 soil samples and 11 other samples (vegetal floating debris, algae) were analyzed. Water and soil sample collections were carried out as follow: For water samples, 10 mL of subsurface water (stream or river) were collected at a 10–30 cm depth every 10 meters, alongside the water body directly into 15-mL sterile Falcon tubes, stored on ice and transported to the laboratory. For soil samples, approximately 50 g topsoil was collected from river banks (from 10 cm below to 1 meter above water level) in shaded areas using a core drilling (3 cm large by 5–7 cm height). Each soil sample was immediately placed into a 50-mL sterile Falcon tube. Water quality and environmental parameters were collected at the time of sampling (apparent meteorological and hydrological conditions, presence of iridescences, debris, foam or stagnant fludge, water color, clarity, turbidity, salinity, temperature, dissolved oxygen, pH, UV radiation, altitude). The location of sampling sites was taken with a Garmin GPS. All samples were transported to the laboratory and processed within 48 hours of collection.
Each water sample (10 mL) was centrifuged at 8000 × g for 10 min. The pellets were resuspended to a total volume of 200 μL in the original water and immediately lysed to begin the extraction process using a commercial kit (QIAamp DNA Mini Kit, Qiagen, Australia) according to the manufacturer’s instructions. DNA elution was performed with 50 μL of buffer AE. The quantity of DNA was measured by NanoDrop (Thermo Fisher Scientific). Soil samples were submitted to DNA extraction using the PowerSoil DNA Isolation kit (MO BIO), shown in preliminary experiments to be the most efficient to extract leptospiral DNA from New Caledonian soils.
Briefly, 250 mg of soil is poured in a PowerSoil bead tube before addition of 60μL suspension Buffer C1. This suspension is shaken for 5 minutes at 2,000 rpm using a MagNA Lyser (Roche). The supernatant is lysed at 4°C for 5 min with 250 μL lysis buffer C2. Up to 600 μL of supernatant is transferred in a new tube before addition of 200 μL of Inhibitory Removal Technology solution C3 before incubation at 4°C for 5 min. This step is essential for the final DNA quality as it allows the cationic flocculation of humic substances which usually account for low DNA recovery and qPCR inhibition. Up to 750 μL of the supernatant is transferred in a new tube and gently mixed with 1,200 μL of DNA binding solution C4 prior to be loaded into a Spin Filter and centrifuged at 10,000x g for 1 minute at room temperature. After washing the precipitated DNA with 500 μL of wash buffer C5 through the spin filter membrane, the DNA is eluted with 100 μL elution buffer C6.
Soil and water samples were tested for the presence of pathogenic Leptospira DNA using the real-time PCR targeting lipL-32 [15]. The reactions were performed in a final volume of 20 μL containing 1X LightCycler 480 probes Master (Roche Applied Science, New Zealand), 0.4 μM each primer and 0.13 μM probe, and 2 μL template DNA. The cycling conditions were as described in the original publication in a LightCycler 480 (Roche Applied Science) [15].
Samples with a positive lipL-32 qPCR were investigated with BLU-V Viability PMA Kit (Qiagen) to evaluate the presence of viable pathogenic leptospires, except for site 1. Briefly, 5 g of soil were gently resuspended in 5 mL of 1X Phosphate Buffered Saline and let to settle down for 1 hour. Then 100 μL of this soil suspension supernatant was mixed with 2 μL of propidium monoazide (PMA; 50 μM final concentration) in light-transparent 1.5 mL microcentrifuge tubes. Following a 10 min incubation in the dark, samples were exposed for 10 min to a 3-watt LED light (460-470 nm) with gentle homogenization every 2 minutes. The sample tubes were laid horizontally under the light source to ensure optimal PMA/DNA cross-linking, thus avoiding false positive results. In order to test the efficiency of PMA treatment of membrane-compromised bacterial cells, duplicate tubes of the same soil solution supernatant were heated at 80°C for 10 min. The heat-treated samples were then cooled to room temperature before PMA addition, incubation and photoactivation. In addition, a control tube without PMA was included to determine the presence of total pathogenic Leptospira (both dead and live) in the soil sample. After photoinduced cross-linking, samples were treated for DNA isolation using QIAamp DNA Mini kit (Qiagen). The corresponding DNA extracts were used as templates for qPCR targeting the lfb-1 gene [17] in order to subsequently phylogenetically identify the viable pathogenic Leptospira present in the sample. This qPCR was run on a LightCycler LC 2.0 using the LightCycler FastStart DNA Master SYBR Green I kit (Roche Applied Science, New Zealand) as described before [17].
The lfb-1 sequence polymorphism was used as a molecular phylogenetic target to link the identity of environmental leptospiral sequences to the corresponding human infecting strain. Amplified lfb-1 DNA products obtained from environmental samples were identified by DNA sequencing. The amplicons were purified using a DNA purification kit (Qiagen, Australia) and sequenced directly as described before [16]. The resulting DNA sequence data were compared with sequences retrieved from the patient’s sample and with the GenBank database using the BLAST algorithm. The sequences obtained in this research were deposited in GenBank under the Accession Numbers: KY052025; KY052026; KY052027; KY052028; KY052029; KY052030; KY052031; KY052032; KY052033; KY052034; KY052035; KY052036; KY052037; KY052038; KY052039; KY052040.
Located in the north of New Caledonia, in places were leptospirosis is endemic, 4 different sites distributed over 3 municipalities were investigated (Fig 1). For sites 1, 2 and 4, the infections supposedly occurred on February 2nd, during the same heavy rain event which hit New Caledonia the same day (S1 Fig). All patients were swimming in a freshwater stream when the rain started to fall and all 3 reported an increase of the water flow and a change in water color and turbidity in their respective bathing sites. For site 3, the patient probably got infected on February 9th when fishing freshwater shrimp using a mask and snorkel. For site 1, only one investigation was performed 6 weeks after the contamination event. Sites 2, 3 and 4 were investigated twice with 7 weeks (site 3 and 4) or 10 weeks (site 2) between investigations (S1 Fig).
The overall results for detection of pathogenic leptospires from these 4 sites are summarized in Table 1.
Interestingly, of the 10 water samples collected, none were positive for the presence of pathogenic Leptospira DNA by qPCR, either at the early or late investigation time point. Contrarily, soil samples were mostly positive: 58% of soil samples (30/52) were positive using lipL-32 qPCR. It is worth to note that among soil samples investigated, we were able to amplify pathogenic Leptospira DNA from the river bank up to 1 meter above the water level. In such a core soil sample, DNA from pathogenic Leptospira was amplified from all 1-cm thick slices down to a 5-cm depth. In addition, 2 samples mostly made of benthic algae collected on the bottom of the streams were also positive using lipL-32-qPCR.
Despite a decreasing number of leptospiral DNA-positive soil samples in sites 3 and 4, we still successfully detected pathogenic Leptospira DNA in the late samples collected 4 months after the index infection event, although the qPCR Cycle Thresholds (Ct) slightly increased (S1 Fig). In contrast, in the two control sites where L. interrogans Pyrogenes was known to have been involved in human cases in previous years but without recent contamination reported (> 1 year), none of the samples collected was positive.
For each soil sample positive for lipL-32 by qPCR, we further investigated the genotype of these pathogenic leptospires using the lfb-1 phylogenetic scheme used for patients. Viability-PCR (v-PCR) and qPCR targeting lfb-1 using the v-PCR treated DNA as a matrix were performed subsequently when possible (for sites 2–4). The overall results for v-PCR and lfb-1-derived phylogenetic analysis from these 6 sites are summarized in Table 2.
In all 4 sites investigated, we were able to amplify DNA from L. interrogans Pyrogenes, respectively at 6; 9; 9 and 10 weeks post-infection (WPI) for site 1, 2, 3 and 4. To further assess if this L. interrogans Pyrogenes DNA derived from live cells, we performed a viability-PCR (for sites 2–4). Two sites (2 and 3) out of the 3 investigated were positive for v-PCR and phylogenetic analyses of the amplified DNA matched to L. interrogans Pyrogenes. Interestingly, a lfb-1 sequence identical to a L. interrogans from serogroup Australis, involved in other human cases in New Caledonia, was also detected in site 2, concomitantly with Pyrogenes. v-PCR was also positive in site 4, but the phylogenetic analysis of the amplified lfb-1 sequences did not match the infecting strain nor any other reported isolate (except one sequence displaying 96% identity with L. kmetyi; Fig 2).
To clarify whether the pathogenic leptospires could be detected over a longer period, soil samples were collected again 19 (site 2), 16 (site 3) or 17 WPI (site 4). All the samples investigated from these 3 sites were negative using v-PCR. However, we were still able to amplify a few lfb-1 sequences using direct qPCR for site 3 and 4. Phylogenetic analysis of these lfb-1 sequences appeared to differ from any known strain or species, though some were similar to those amplified during our first investigation (Fig 2).
Finally, it is interesting to note that temporal analysis of our results seems to highlight dynamic changes of the pathogenic leptospires in environmental sites. Indeed, when sequences identical to L. interrogans Pyrogenes or Australis were found during our first investigation, they were either not re-detected (site 2 and 4) or substituted by other unknown pathogenic leptospires upon our 2nd investigation (site 3).
Infected mammals by shedding large amounts of virulent leptospires in their urine, massively contaminate their surrounding environment [5,18–20]. These pathogens eventually get drained in freshwater systems upon heavy rain episodes. This dispersion through freshwater not only participates to substantial contamination of large areas but also brings the threat right in human influence area. As environmental contamination is the major source of human leptospirosis, we attempted in this study, to evidence the presence of virulent leptospires in natural habitats in New Caledonia. Using molecular-based methods, we investigated the presence and viability of pathogenic leptospires in area believed to be contamination sites. Although the use of qPCR is becoming frequent for diagnostic purpose [14], the use of this technique on environmental samples is not commonplace, mainly because of the presence of inhibitors impairing qPCR efficiency [13].
We applied this methodology to complex environmental samples from places selected as putatively involved in human cases and we successfully amplified pathogenic Leptospira DNA in all the sites investigated. Phylogenetic analysis based on the lfb-1 gene sequence successfully linked the identity of environmental leptospiral sequences to the corresponding human cases. More importantly, we demonstrated by viability qPCR that these pathogenic leptospires were viable and probably persisted in infection sites weeks after the contamination event. Notwithstanding that little is known regarding the mechanisms regulating the persistence of pathogenic leptospires in natural habitats, outside a mammalian host, general agreement in the scientific community agrees on the fact that pathogenic Leptospira can survive for long periods in freshwater [7] and a few studies exploring the survival of Leptospira in soils successfully reported re-isolation of the same Leptospira isolate 5 months later [21]. Our results show, that when performing retrospective investigations, pathogenic leptospires could be evidenced only in soils samples, up to 4 months after the index contamination. Though repeated contaminations from an animal source might occur, our results strongly suggest a prolonged survival in river banks and soils. Moreover, to our knowledge, this study reports the first successful viability-PCR performed on Leptospira from complex soil samples in combination with molecular-based typing to identify environmental leptospires involved in human cases. We formally established that up to 9 weeks after infection, the pathogenic Leptospira strain involved in human cases was viable in environmental soil samples, and thus potentially infectious, suggesting an ongoing risk for humans. Using samples collected 4 months after the contamination event, we were not able to evidence the presence of this particular virulent strain, therefore suggesting a decrease in Leptospira viability over time, a hypothesis supported by higher Ct values in qPCR from late samples (S1 Fig). These 2nd investigations were performed during the cool season, also assumed to be detrimental for Leptospira survival. Whether this change in temperature contributed to the decrease in environmental contamination remains unknown but would be in good agreement with empirical knowledge as well as the experimental results reported by Andre-Fontaine and collaborators [7].
Interestingly, we have not evidenced pathogenic leptospires in any of our water samples, contrasting with previous observations [22,23]. Following patients’ interviews, we have investigated the suspected flowing water bodies (streams and rivers) at their normal flow rate and weeks after the index contamination event. Oppositely, stagnant water sources (gutters, wells, puddles, reservoirs) considered in other studies [22,23] were not mentioned in the interviews and therefore were not investigated. In addition, we have processed a relatively small volume of water for DNA extraction (10 mL), contrasting with larger volumes (50–1,000 mL) in other studies. Lastly, we also have investigated a smaller number of water samples compared with soils. Taken together, these facts may explain the differences observed.
The detection (or absence of detection) of live L. interrogans Pyrogenes after long periods should be interpreted with caution because of possible limitations of the v-PCR methodology, especially in environmental samples [24]. Although selective nucleic acid intercalating dyes, like propidium monoazide used in this study, represent one of the most successful recent approaches to detect viable cells (as defined by an intact cell membrane) by PCR and have been effectively evaluated in different microorganisms, major drawbacks have also been reported [24]. When applied to complex environmental samples the dye may be able to penetrate into viable or reversibly damaged cells, leading to false negative results. Further, bacteria might not all be transferred from their substratum to the supernatant or be damaged during the initial steps of the v-PCR protocol. Considering that Leptospira concentrations in our samples were low, v-PCR most probably underestimated bacterial viability in our experimental procedure.
Interestingly, the control sites, defined by the former presence of L. interrogans Pyrogenes but without human contamination reported over more than one year, showed no positive sample for lipL-32 qPCR detection. This not only suggests that the index patients actually got infected in the sites investigated, but also that targeting human contamination areas is a valuable strategy to properly identify Leptospira-contaminated areas, notably for research purpose.
Dynamic changes in Leptospira population in environmental samples seem to have occurred over the time course of this study. While the infecting L. interrogans Pyrogenes could only be detected by qPCR during the first investigation, our study revealed the presence of other pathogenic Leptospira DNA not associated to any known species (site 4). In sites 2 and 3, L. interrogans Pyrogenes was detected alongside with other pathogenic Leptospira spp. It is well known that microorganisms can cooperate in complex assemblages to better exploit nutritional resources and resist to stressful environmental conditions. Because leptospires are thought to be highly susceptible to adverse environmental stresses, they could have promoted a unique microbial interaction, by which leptospires would successfully survive and persist in the environment. This emerging idea has been highlighted by the recent discovery of biofilm produced by pathogenic leptospires [25,26]. Whether multispecific biofilms either produced by Leptospira spp. or formed by other environmental bacteria and providing shelter to leptospires, might be present in natural habitats, contribute to the persistence and allow long-term survival of pathogenic leptospires in nutrient-poor or adverse aqueous environments deserves consideration. Recent work in other settings where leptospirosis is highly endemic supports this hypothesis [9].
Interestingly, during the flooding event which occurred on February 2nd, people who got infected at sites 1, 2 and 3 were bathing with at least 2 other persons who were exposed similarly to the Leptospira environmental risk, but did not develop a clinical form of leptospirosis. This observed low attack rate raises many questions including asymptomatic leptospirosis. A recent sero-incidence study in Brazil has shown that only a very small proportion of infections actually leads to clinical disease [27].
Overall, this study revealed that pathogenic Leptospira are widespread in river soils in places associated with recent human cases. The infecting strain was evidenced in all the investigated sites and viable leptospires were still detected 9 weeks after the contamination event. These observations are particularly interesting especially if they are analyzed in regards of daily rainfall data (S1 Fig). Analysis of the daily rainfall shows that in all 4 study sites, several episodes of heavy rain occurred over the 6-month period when this study was performed. But consistently with our qPCR results these rain events have probably not been a major source of re-contamination with human threatening strains, as supported by (i) the fact that a similar sampling strategy failed to evidence an environmental contamination with L. interrogans Pyrogenes in late samples and (ii) an increase in qPCR Ct values (S1 Fig), suggesting a decrease of the environmental Leptospira load over time. Therefore, it is likely that the Leptospira DNA that was detected over this 4-month study corresponded to leptospires deposited by the flooding event of February 2nd.
Still, temporal investigations evidenced changes in leptospiral diversity and revealed the presence of yet unreported strains in soil samples and never evidenced in any mammal in New Caledonia despite long research, suggesting that soil might act as an environmental reservoir of pathogenic leptospires offering them a protective atmosphere. v-PCR coupled to molecular-based typing on soil samples proved effective at confirming infection sites and investigating the leptospiral risk over time. Because soil DNA extraction only uses small amounts of soil, the use of this approach for risk evaluation should consider the possibility of false negative results. Still, the assessment and quantification of the leptospiral burden in environmental samples might prove valuable to guide public health interventions. To help expand the current knowledge about the leptospirosis environmental cycle and the spatial and temporal distribution of leptospires in the environment, further studies will also characterize the physicochemical characteristics of soils shown to support or oppositely compromise the survival of pathogenic leptospires. Furthermore, determination of the environmental burden may help inform health authorities before adopting preventive measures such as access restrictions to contaminated areas during heavy rainfall events. Finally, evaluation of the environmental leptospiral load through quantitative methods can be a useful method to monitor high risk areas and help protect local populations, but also to discover an unexplored biodiversity of pathogenic leptospires.
|
10.1371/journal.pbio.1002572 | Relative Contributions of Specific Activity Histories and Spontaneous Processes to Size Remodeling of Glutamatergic Synapses | The idea that synaptic properties are defined by specific pre- and postsynaptic activity histories is one of the oldest and most influential tenets of contemporary neuroscience. Recent studies also indicate, however, that synaptic properties often change spontaneously, even in the absence of specific activity patterns or any activity whatsoever. What, then, are the relative contributions of activity history-dependent and activity history-independent processes to changes synapses undergo? To compare the relative contributions of these processes, we imaged, in spontaneously active networks of cortical neurons, glutamatergic synapses formed between the same axons and neurons or dendrites under the assumption that their similar activity histories should result in similar size changes over timescales of days. The size covariance of such commonly innervated (CI) synapses was then compared to that of synapses formed by different axons (non-CI synapses) that differed in their activity histories. We found that the size covariance of CI synapses was greater than that of non-CI synapses; yet overall size covariance of CI synapses was rather modest. Moreover, momentary and time-averaged sizes of CI synapses correlated rather poorly, in perfect agreement with published electron microscopy-based measurements of mouse cortex synapses. A conservative estimate suggested that ~40% of the observed size remodeling was attributable to specific activity histories, whereas ~10% and ~50% were attributable to cell-wide and spontaneous, synapse-autonomous processes, respectively. These findings demonstrate that histories of naturally occurring activity patterns can direct glutamatergic synapse remodeling but also suggest that the contributions of spontaneous, possibly stochastic, processes are at least as great.
| The modification of synaptic connections by specific activity histories (a phenomenon known as synaptic plasticity) is widely believed to represent a major substrate of processes collectively referred to as learning and memory. Recent studies indicate, however, that synapses also change spontaneously, even in the absence of specific activity histories—or, for that matter, any activity whatsoever. This raises a fundamental question: how do changes directed by specific activity histories quantitatively compare to spontaneous changes in synaptic properties? Put differently—what is the “signal-to-noise ratio” of synaptic plasticity at individual synapses? To address this question we followed—over several days—pairs of synapses formed between the same neurons under the assumption that their common activity histories should drive similar changes in their sizes. Indeed, sizes of such synapses tended to change in a correlated manner; yet the extent of this correlation was surprisingly modest, accounting for less than half of the changes that such synapses exhibited. Moreover, sizes of synapses with apparently common activity histories tended to be quite different. Our findings thus indicate that the “signal-to-noise ratio” of synapse remodeling might be rather poor, on the order of 1:1 or less.
| Activity-induced modification of synaptic connections (synaptic plasticity) is widely believed to represent a major mechanism for modifying the functional properties of neuronal networks. Indeed, overwhelming experimental evidence supports the idea that synaptic properties are affected by the history of their activation. What is less established and often ignored is the "flip side" of synaptic plasticity: that is, the implicit supposition that synapses, when not driven to change their characteristics, will retain these over time. This assumption would seem to be an essential complement of the synaptic plasticity concept; without it, spontaneous changes occurring independently of physiologically relevant input would cause spurious changes in network function or undo physiologically relevant ones.
The validity of this assumption has been called into question by recent studies, in which sizes and contents of individual synapses—both excitatory and inhibitory—were observed to fluctuate considerably over timescales of hours and days (e.g., [1–17]); notably, such fluctuations persisted even in the absence of specific activity patterns or any activity at all (e.g., [5,6,9,12,17]). Finally, it was shown that these fluctuations could be described remarkably well by statistical processes that are essentially stochastic [5,6,8,16,17]. Given the emerging view of the synapse as a complex assembly of dynamical components [1,2], the presence of such fluctuations might not be very surprising. Nevertheless, they would seem to imply that synaptic tenacity, which we define as the capacity of individual synapses to maintain their properties over behaviorally relevant time scales [6,9,11,17], is inherently limited, and that synapses exhibit a non-negligible degree of spontaneous size remodeling.
Although these conclusions were derived mainly from studies in reduced systems (cell and organotypic cultures), they are not limited to these settings [4,8,14,15]. Thus, for example, it has recently been shown that synapse size fluctuations in the cerebral cortex of adult mice are at least as large as those observed in culture ([15]; see also [4]); in fact, the degree of such size fluctuations is comparable to the magnitude of size changes induced by experimental stimulation paradigms that induce long-term potentiation (e.g., [18,19]). Thus, when considering changes in synaptic sizes, it remains to be asked what the relative contributions of specific activity histories to such changes are and how these compare to size changes driven by other, possibly stochastic, processes.
In the rodent cerebral cortex, two neurons are often connected by more than one excitatory synapse (reviewed in [20]). This situation provides an excellent opportunity to examine the relative contributions of specific activity histories to changes in synaptic sizes and then compare these to the contributions of other processes. This claim is based on the reasonable premise that, to a first approximation, all synapses connecting two specific neurons (commonly innervated [CI] synapses) will have similar activation histories when these are integrated over many days [21,22]. Assuming that changes in synaptic properties are driven primarily by activation histories, changes in the sizes of such CI synapses might be expected to co-vary significantly. In contrast, synapses formed on the same neuron or dendrite by two different upstream neurons (non-commonly innervated [non-CI] synapses), would have somewhat different activation histories, and thus their sizes would not be expected to co-vary to the same degree. Moreover, the remodeling covariance would be expected to be even greater for nearby synapses formed between the same axonal and dendritic segments, as regional differences in axonal/dendritic properties would minimally affect activity histories and their biological consequences. Finally, this approach provides an opportunity to examine how synaptic sizes are affected by more natural activation histories, spanning hours and days, as compared to the brief and rather artificial stimulation paradigms typically used in experimental settings (reviewed in [23]).
In the current study we measured and compared the remodeling of CI and non-CI synapses in monolithic and modular networks of cortical neurons in primary culture by using genetically encoded fluorescent reporters combined with multielectrode array (MEA) recordings, automated confocal microscopy, and pharmacological manipulations. Although cortical networks in culture differ in many ways from their in vivo counterparts, in the current context, they are advantageous in the sense that they provide a generic, isolated, and well-controlled system for studying the net effects of activation histories, free from potential confounds inherent to in vivo settings such as behavioral states, stress, neuromodulatory input, and circadian rhythms. Moreover, as shown below, this system allows for excellent long-term and continuous monitoring of synaptic sizes, the presynaptic origins of individual synapses, and experimental differentiation of activation histories. Our findings are presented next.
The rationale of the experiments described below is depicted in Fig 1A. In this scheme, a single postsynaptic neuron is innervated by multiple axons belonging to different “upstream” excitatory neurons. A subset of synapses formed on this postsynaptic cell represents connections formed with a particular upstream axon, and these are hereafter referred to as CI synapses. Some CI synapses are located on the same dendrite, and these are hereafter referred to as Commonly Innervated Same Dendrite (CISD) synapses. For each CI synapse, a nearby synapse is selected, which represents a connection between the postsynaptic neuron and another axon. These are hereafter referred to as reference (Ref) synapses. As explained above, it might be expected that CI synapses will have very similar activation histories (even more so, perhaps, for CISD synapses). If activation history is the major force that drives changes in synaptic size, then CI synapses should change in a similar manner, resulting in a strong covariance of their sizes over time (as illustrated schematically in Fig 1B). Similarly, given that CI and Ref synapses are activated by different upstream neurons and assuming that the activity histories of these neurons differ significantly (a matter we will return to later), sizes of CI and Ref synapses (non-CI synapses) would not be expected to co-vary to the same extent (Fig 1C), with the residual covariance mainly representing the combined contributions of (postsynaptic) neuron- (or dendrite-) wide, non-synapse-specific processes. The overall goals were therefore to (1) quantify the covariance of CI synapses, (2) compare it to the covariance of non-CI synapses, and (3) use these data to estimate the specific contributions of particular activity histories to the remodeling of glutamatergic synapses.
The experiments were carried out in a system based on networks of rat cortical neurons growing on thin glass MEA substrates, automated confocal microscopy, and genetically encoded fluorescent reporters. This system, which we have previously used to explore relationships between activity and remodeling of excitatory [6,16,24] and inhibitory synapses [17], allows for chronic recordings of network activity from up to 59 electrodes while simultaneously imaging synapses by automated confocal microscopy for many days, even weeks. For these experiments, we used cortical networks maintained in culture for 18–21 d, as at this time, synaptogenesis is mostly complete and synapses are relatively mature. To estimate changes in synaptic sizes, we expressed fluorescently tagged variants of the postsynaptic density (PSD) protein PSD-95 (/Dlg4/SAP90) and followed changes in its fluorescence at individual synapses. PSD-95 is a core postsynaptic scaffolding protein of glutamatergic synapses that is thought to control the number of glutamate receptors at the postsynaptic membrane through direct and indirect interactions (reviewed in [25]; see also [26]). Importantly, a recent in vivo correlative light and electron microscopy study [15] demonstrated excellent correlations between tagged PSD-95 fluorescence and PSD area when these are measured for the same synapses, and thus fluorescently tagged PSD-95 can be used to record changes in PSD area and, by extension, in synaptic size. To locate CI synapses, we expressed spectrally separable fluorescently tagged variants of presynaptic molecules, namely SV2 (a conserved, highly specific synaptic vesicle integral membrane protein; [27,28]) or Synapsin I (a synaptic vesicle-associated phosphoprotein [29]; experiments described later on). Expression of all fluorescent reporters was carried out using lentiviral vectors, resulting in minimal overexpression levels of exogenous proteins and very sparse labeling of individual neurons. In spite of the sparse labeling, postsynaptic sites (labeled with fluorescently tagged PSD-95) juxtaposed against fluorescent presynaptic sites (labeled with fluorescently tagged SV2 or Synapsin I) were often observed. Careful examination then allowed us to locate pairs (and sometimes triplets or more) of CI synapses, that is, postsynaptic sites connected to the same axon. As axonal shafts were often barely discernable, the selection of CI synapses for subsequent analyses was limited to short axonal stretches for which a common axonal origin could be determined unambiguously (see Materials and Methods for further details). Fluorescently tagged CI synapses were then followed over time to verify that presynaptic and postsynaptic compartments remained juxtaposed at all time points. A Ref synapse was then chosen near each synapse connected to the common axon, and tagged PSD-95 fluorescence was measured at all synapses—CI and Ref alike—at each time point for the duration of the experiments (all measurements were made in maximum intensity projections of all sections). To minimize the potential effects of measurement noise, fluorescence measures of each synapse were first smoothed with a 2.5- to 3-h low-pass filter [16]. The fluorescence covariance of all CI and non-CI synapse pairs was then calculated using Pearson’s correlation (a linear measure) as well as Spearman’s rank correlation (a measure that quantifies monotonic, but not necessarily linear, relationships between two variables).
We first compared the covariance of CI and non-CI synapses in monolithic cortical networks. In these experiments, individual postsynaptic sites were visualized using PSD-95 tagged with enhanced green fluorescent protein (EGFP) (PSD-95:EGFP; [6,24]), whereas presynaptic sites were visualized using SV2 tagged with Cerulean (a cyan fluorescent protein variant; Cer:SV2; [30]). As shown in Fig 2A, dendrites, individual postsynaptic sites, and presynaptic boutons were readily discernable, allowing us to locate and follow CI and Ref synapses (Fig 2B). To compare the size covariance of CI and non-CI synapses, the networks were mounted on the combined MEA recording/imaging system described above and provided with optimal environmental conditions (a sterile atmospheric environment of 5% CO2 and 95% air, slow perfusion with fresh feeding medium, and a temperature of 37°C), allowing us to carry out experiments lasting one week or longer with no signs of deterioration or cell death (Fig 2B). Stacks of images (at 10 focal planes) were collected automatically from 6–12 fields of view (or sites). Images were collected at 30-min intervals for several days concomitantly with recordings of network activity (action potentials) from the 59 electrodes of the MEA dish. As we noted in preliminary experiments that Cerulean exhibited significant photobleaching, axons were imaged at longer intervals (once every 7.5 h). Imaging was started only 2–3 d after mounting the preparations, as we noted here and elsewhere [6,17] that the first 24–36 h of such experiments are invariably associated with increases in spontaneous activity levels related to the introduction of slow perfusion. Imaging in spontaneously active networks was then carried out for at least two further days. Finally, the Na+ channel blocker tetrodotoxin (TTX) was added to the MEA and perfusion media to suppress spontaneous network activity, and imaging was continued for additional 1–2 d.
In agreement with prior cell culture [3,6,11,13,24,31] and in vivo [4,15] studies, the fluorescence of individual PSD-95:EGFP puncta often changed considerably over timescales of many hours. This is exemplified for two CI and two non-CI synapses in Fig 2C. The synapse size covariance of CI and non-CI synapses was then compared by calculating the correlation between the changes in PSD-95:EGFP fluorescence for each CI and non-CI pair over periods of 48 h. This is illustrated for one CI synapse pair (Fig 2C) and respective non-CI pairs (Fig 2D). In this example, the covariance of the CI pair is much greater than that of the non-CI pairs; this difference, however, was not nearly as obvious in all such comparisons (92 pairs from 24 neurons in 6 experiments). In fact, distributions of both Pearson’s correlation coefficients (r) and Spearman’s rank correlation coefficients (ρ) measured for both CI and non-CI pairs were quite broad (Fig 3A and S1A Fig, respectively). Nevertheless, the average covariance measured for all 92 CI pairs was somewhat greater than that measured for all non-CI pairs: (Fig 3B, S1B Fig; CI pairs: r = 0.17 ± 0.05, ρ = 0.15 ± 0.05; non-CI pairs: r = 0.06 ± 0.02, ρ = 0.05 ± 0.02; average ± SEM; p = 0.04, p = 0.04, Pearson’s and Spearman’s correlation respectively, two-tailed Mann-Whitney U test). This difference was also observed when data were pooled by experiment (Fig 3D, S1D Fig; CI pairs: r = 0.19 ± 0.05, ρ = 0.18 ± 0.05; non-CI pairs: r = 0.05 ± 0.03, ρ = 0.04 ± 0.03; average ± SEM; p = 0.04, p = 0.04, Pearson’s and Spearman’s correlation, respectively, Mann-Whitney U test).
If the greater covariance observed for CI synapses is due to the commonality of their activation histories, blocking network activity might be expected to reduce CI synapse covariance to levels observed for non-CI synapses. Somewhat surprisingly, however, suppressing spontaneous network activity as described above, resulted in substantial increases in covariance values for both CI and non-CI synapses (Fig 3C and 3E, S1C and S1E Fig; CI pairs: r = 0.25 ± 0.06, ρ = 0.25 ± 0.06; non-CI pairs: r = 0.17 ± 0.04, ρ = 0.17 ± 0.03; average ± SEM). We attribute this general increase in remodeling covariance to the nonspecific growth of glutamatergic synapses associated with the suppression of network activity (S2A Fig, see also [6,24]). A small difference between the covariance of CI and non-CI pairs was still apparent; this difference, however, was not statistically significant (p = 0.33, p = 0.41, Pearson’s and Spearman’s correlation, respectively, Mann-Whitney U test). The suppression of network activity is known to evoke and affect numerous cellular processes (collectively referred to as synaptic “homeostatic” processes [32]) and parametrically affect the statistics of stochastic remodeling processes [6,16]. As the effects of “homeostatic” processes are not easily disentangled from activity-dependent remodeling processes in active networks, these apparently straightforward experiments were not as informative as might have been expected, although they hint that CI synapses might change in a slightly more correlated manner even when network activity is blocked (see Discussion).
Although the size covariance observed for CI synapses in active networks was somewhat greater than that observed for non-CI synapses, the difference was surprisingly modest. We explored several possible reasons for this modest difference.
We first considered the possibility that the overall extent of remodeling exhibited by synapses in these preparations was small, and, thus, the measures of covariance used here might have reflected, for the most part, the (in)coherence of low amplitude noise in fluorescence measurements. To evaluate this possibility, we measured for each synapse its normalized range of change (“range over mean”) defined as
RangeMean=100*Fmax−FminF¯
where Fmax, Fmin, and F¯ are the maximal, minimal, and average PSD-95:EGFP fluorescence intensities, respectively, measured for a given synapse over a period of 48 h. As shown in Fig 3F, distributions of range over mean values were rightward skewed and similar for CI and Ref synapses; about 35% of synapses changed by more than 40% over this period, whereas averages (±SEM) of range over mean values were 37% ± 1.6% (CI) and 37% ± 1.6% (Ref) (Fig 3G). Thus, synapses exhibited substantial changes over these periods, similar in magnitude to changes induced in organotypic slice cultures by paradigms that induce long-term potentiation (33% on average; [18]). This and the fact that all data were low-pass filtered before analysis is thus not in line with the possibility that our covariance measures mainly reflect low amplitude measurement noise (see also [16]). Interestingly, the suppression of network activity reduced, but did not eliminate, synaptic remodeling (Fig 3H; CI: 23% ± 1.2%; Ref: 24% ± 1.5%; average ± SEM). Here too, however, the contributions of “homoeostatic” and other processes to this remodeling are not readily disentangled.
The expected differences in size covariance of CI and non-CI synapses are based on the assumption that activity histories of CI synapses are much more similar than activity histories of non-CI synapses. If, however, all synapses—regardless of their presynaptic origin—share similar activation histories, the size covariance of CI and non-CI synapses might not be expected to differ much. This possibility cannot be ignored, as activity in the preparations used here tends to occur as synchronized bursts that encompass a large fraction of neurons within the network (Fig 4A; e.g., [6,24,33–36]). To increase the “contrast” between the activity histories of synapses belonging to different neurons, we desynchronized network activity by exposing the neurons to Carbachol [24], a non-hydrolysable cholinergic agonist. As shown in Fig 4B, Carbachol (20 μM) greatly diversified the spontaneous activity characteristics, causing some neurons to fire continuously, others to fire more sporadically, and others to fire only occasionally. Furthermore, the tendency of the network to generate network-wide, synchronous bursts was suppressed. Somewhat unexpectedly, this manipulation eliminated the differences between CI and non-CI synapses while elevating their absolute size covariance values (Fig 4C and 4D; CI pairs: r = 0.26 ± 0.09, ρ = 0.21 ± 0.11; non-CI pairs: r = 0.25 ± 0.05, ρ = 0.22 ± 0.06; average ± SEM). Here too, the increased covariance reflects the generalized synaptic growth that follows prolonged exposure to cholinergic agonists (S2B Fig) [24].
The experiments described so far indicated that synapses with similar activity histories changed in a somewhat more correlated manner in comparison to synapses with apparently different activity histories, but the difference between the two groups was rather modest. The possibility that this might have been due to the limited diversity of activity histories in these networks was not supported by pharmacological network desynchronization, but the interpretation of the latter experiments was complicated by the global effects of cholinergic agonists on synaptic properties. Moreover, due to the tendency of synchrony to reemerge after ~12 h in such experiments [24], the duration of such experiments was inherently limited. We thus sought to diversify the activity histories of CI and non-CI synapses by different means. To that end, we turned to modular network architectures.
As mentioned above, large groups of neurons in the networks used here tend to fire in synchronized bursts, indicating that the activity histories of neurons in such networks might be quite similar. Previous studies have shown, however, that when such networks are divided into modules separated by barriers partially restrictive to axonal extension, activities in the two modules become more disparate ([37]; see [38] for a comprehensive analysis). We thus set out to compare the size covariance of synapse pairs innervated by axons originating in the same module with the size covariance of synapse pairs in which each synapse is innervated by axons originating in two different modules. To that end, we labeled neurons in one module with a postsynaptic reporter (referred to here as the “postsynaptic” module) and labeled cells in the other module with a presynaptic reporter (the “presynaptic” module; see Fig 5A for a schematic illustration of this “presynaptic/postsynaptic” arrangement). We then searched for pairs of synapses on neurons in the postsynaptic module formed by axons that crossed over from the presynaptic module. The assumption here was that the activity history of these synapses will be similar yet substantially different from the histories of most other synapses in the postsynaptic module, the axons of which were much more likely to have local origins.
In practice, networks were divided into two subnetworks by fabricating polydimethylsiloxane (PDMS) inserts with two compartments and sealing them onto special MEA dishes whose electrodes were arranged in a modular fashion (four-quadrant [4Q] MEAs). The two modules were connected through 6–12 very narrow channels (~400 μm long, ~13 μm wide, and ~3 μm high), which allowed some axons to grow across the barrier and innervate neurons in the other module (Fig 5B), yet were restrictive to the migration of entire cells. Neurons in the presynaptic module were labeled with GFP-tagged Synapsin-Ia (EGFP:SynI) instead of Cer:SV2 due to its much greater photostability, its high endogenous expression levels, and its very high fidelity as a presynaptic marker [39]. Neurons in the postsynaptic module were labeled with PSD-95 tagged with mTurquoise2 (PSD-95:mTurq2), a very bright and relatively photostable variant of cyan fluorescent protein ([40]; see also [13]). The expression of each reporter was fully restricted to its respective module, ensuring that EGFP:SynI-labeled axons observed in the postsynaptic module originated in the presynaptic module. A limited number of EGFP:SynI-labeled axons crossed over to the postsynaptic module (Fig 5B) and formed synapses with neurons in that module (Fig 5C); based on comparisons of axon labeling density in the presynaptic and postsynaptic modules, axons from the presynaptic module represented a tiny fraction (less than 1%) of the total number of axons in the postsynaptic module. Thus, the vast majority of PSD-95:mTurq2 puncta in the postsynaptic module was innervated by axons originating within that module.
The presence of extracellular electrodes in both modules allowed us to examine the disparity of activity in the two modules. As shown in Fig 5D, some network-wide bursts spread from one module to the other (with some delay), but many network-wide bursts remained confined to one module and did not spread to the other module. To verify that axons traversing the barrier indeed carried the activity patterns of the presynaptic module into the postsynaptic module, we expressed the genetically encoded calcium indicator GCaMP6s [41] in presynaptic module neurons and used an electron-multiplying charged couple device (EMCCD) camera to measure Ca2+ transients in the presynaptic boutons of axons that crossed into the postsynaptic module (S3 Fig). To that end, sequences of 600 frames were captured at rates of ~7 Hz, allowing us to compare the timing of Ca2+ transients with network activities of the presynaptic and postsynaptic modules. As illustrated in S3E Fig, Ca2+ transients measured in such axons corresponded extremely well with bursts of activity recorded from the electrodes in the presynaptic module, but not nearly as well with bursts recorded from the postsynaptic module. This analysis also confirmed that Ca2+ transients in boutons distributed along the labeled axonal segments correlated almost perfectly, as might be expected. (S3C and S3D Fig; see also [42]). Collectively, these observations show that activity histories of CI synapses are very similar, insofar as action potentials are concerned, whereas those of non-CI synapses differ significantly in both patterning and timing.
We then carried out long-term combined imaging and electrophysiological recordings of neurons expressing PSD-95:mTurq2 and of axons expressing EGFP:SynI as described above. PSD-95:mTurq2 was imaged at 1-h intervals (and EGFP:SynI at 1–3-h intervals) for 2 d. Here too, imaging was initiated only 2–3 d after mounting the preparations on the microscope. After the experiments CI, CISD, and Ref synapses were located, tracked, and their fluorescence values measured. The covariance of CI, CISD, and non-CI pairs (i.e., pairs in which one synapse was innervated by an axon from the presynaptic module and the other by a local axon) was then calculated and compared.
As in the experiments performed in monolithic networks (Fig 3), distributions of correlation values measured for CI, CISD, and non-CI pairs were quite broad (Fig 6A). Yet, in agreement with the aforementioned experiments, the average covariance measured for all CI pairs was greater than that measured for all non-CI pairs (Fig 6B and S4B Fig; CI pairs: r = 0.28 ± 0.03, ρ = 0.28 ± 0.03; non-CI pairs: r = 0.11 ± 0.02, ρ = 0.11 ± 0.02; average ± SEM; p = 1*10−6, p = 4*10−7, Pearson’s and Spearman’s correlation, respectively, Mann-Whitney U test; 271 CI pairs from 29 neurons from 8 experiments). This difference was also observed when data were pooled by experiment (Fig 6C, S4C Fig; CI pairs: r = 0.22 ± 0.07, ρ = 0.24 ± 0.07; non-CI pairs: r = 0.02 ± 0.06, ρ = 0.02 ± 0.06, average ± SEM; p = 0.05, p = 0.04, Pearson’s and Spearman’s correlation, respectively, two-tailed Mann-Whitney U test). A similar observation was made for CISD pairs, i.e., nearby synapses innervated by the same axon and formed on the same dendrite (Fig 6D and 6E, S4D and S4E Fig; CI pairs: r = 0.34 ± 0.05, ρ = 0.34 ± 0.05; non-CI pairs: r = 0.18 ± 0.03, ρ = 0.16 ± 0.03; average ± SEM; p = 0.0036, p = 0.0008, Pearson’s and Spearman’s correlation, respectively, two-tailed Mann-Whitney U test; 91 CISD pairs from 29 neurons from 8 experiments). This difference was also observed when data were pooled by experiment (Fig 6F, S4F Fig; CI pairs: r = 0.35 ± 0.08, ρ = 0.36 ± 0.08; non-CI pairs: r = 0.08 ± 0.07, ρ = 0.06 ± 0.07, average ± SEM; p = 0.04, p = 0.04, Pearson’s and Spearman’s correlation, respectively, two-tailed Mann-Whitney U test).
Although the introduction of a barrier diversified the activity histories of synapses belonging to non-CI pairs, some network-wide bursts did spread from one module to the other, suggesting that activity histories of synapses belonging to non-CI pairs were not entirely dissimilar. The degree to which the two modules were coupled in terms of their bursting activity varied from one experiment to another, ranging from 0.20 to 0.91 (0.64 ± 0.26 average ± standard deviation; see Materials and Methods for further details on this measure). Comparing this coupling with non-CI synapse covariance on an experiment-by-experiment basis revealed a positive correlation (r = 0.62) between these two measures, although this correlation was not statistically significant (p = 0.09). In contrast, and as might be expected, no correlation was observed for CI synapses (r = 0.04, p = 0.99). It should be noted that the measure used here to quantify coupling only considered the fraction of bursts that propagated from one module to another, ignoring functionally important features such as propagation delays and burst durations (see [38] for a comprehensive analysis). Nevertheless, these findings indicate that even in modular networks, the size covariance of non-CI synapses might be influenced somewhat by partial similarities in activity histories, although this influence is at most very small (Fig 6B, 6C, 6E and 6F, S4B, S4C, S4E and S4F Fig; see Discussion).
Although the remodeling covariance of CI and non-CI pairs differed in a statistically significant manner, the actual differences were rather modest. We wondered if this might be due to the inclusion of relatively small synapses, which are more prevalent than large synapses in these preparations [6,16] and in the intact brain [43], as these would be most sensitive to minor fluctuations in background fluorescence or measurement noise. To examine this possibility, we increased the stringency of selection criteria of CI pairs, removing small synapses from the analyses. Even with these stringent selection conditions, however, differences between the covariance of CI and non-CI pairs remained quite modest (S5 and S6 Figs; 103 CI pairs: r = 0.26 ± 0.05, ρ = 0.26 ± 0.04 non-CI pairs: r = 0.08 ± 0.03, ρ = 0.08 ± 0.03; p = 0.003, p = 0.002, Pearson’s and Spearman’s correlation, respectively; 40 CISD pairs: r = 0.33 ± 0.07, ρ = 0.35 ± 0.06; non-CI pairs: r = 0.17 ± 0.04, ρ = 0.14 ± 0.04; p = 0.04, p = 0.009, Pearson’s and Spearman’s correlation, respectively; Mann-Whitney U test).
Not only were the differences in size covariance for CI and non-CI synapses rather modest; the absolute covariance values for CI synapses were surprisingly small, with the highest average values observed in any of the experiments described above being r = 0.35 and ρ = 0.36 (CISD synapses in modular networks; Fig 6F, S4F Fig, respectively). This would seem to suggest that, in addition to joint remodeling, each synapse within a CI pair exhibits significant change that occurs independently of its counterpart. Assuming that CI synapses and, in particular, CISD synapses share common activity histories, the residual remodeling would seem to represent spontaneous, activity-independent synaptic remodeling. Yet it remained possible that at least some of the imperfect size covariance of CI synapses stems from measurement limitations, such as fluorescence measurement inaccuracies. We therefore set out to determine what would have been the average size covariance measured in our system had CI synapse sizes co-varied perfectly. To that end, we introduced artificial correlations between PSD-95:mTurq2 puncta synaptic fluorescence levels by modulating excitation light intensities from one time point to the next (Fig 7A and 7B); we then measured the fluorescence of PSD-95:mTurq2 puncta (Fig 7C) and calculated the correlations for all pairs of synapses in the fields of view (Fig 7D). The depths and temporal profiles of excitation laser light intensity modulation were based on changes in fluorescence levels measured for particular synapses during the long-term experiments described above (Fig 7A), selecting for this purpose synapses whose range/mean ratios were similar to average range/mean ratios measured during those experiments (e.g., Fig 3F and 3G). The experiments were carried out in exactly the same way all experiments described so far were performed, except that here, 48 images were collected in rapid succession to minimize the effects of true synaptic remodeling. As shown in Fig 7E, average correlation values measured here were all positive and rather high (r = 0.78, ρ = 0.76; 100 synapses from 4 neurons, 1,223 pairwise comparisons). These experiments thus suggest that the modest size covariance observed for CI synapses cannot be solely attributed to measurement inaccuracies.
So far, the analyses presented concerned the degree to which sizes of synapses with common activity histories changed together over time. But how similar were the absolute sizes of such synapses? It might be expected that, given their common activity history, their sizes should be similar [21,22]. To examine the degree to which sizes of synapses with identical activity histories were similar, we plotted for each synapse in a CISD synapse pair its PSD-95:mTurq2 fluorescence against the fluorescence of its counterpart in the same pair. For this analysis we used the most stringent data set in which the smallest synapses were omitted (see S5 and S6 Figs), using the measures of PSD-95:mTurq2 fluorescence obtained from the first image stack of each time-lapse series. As shown in Fig 8A, the correlation between the sizes of synapses belonging to the same CISD pair was rather poor (r = 0.23). We then repeated the same analysis for the same synapses, but now using PSD-95:mTurq2 fluorescence values averaged for each synapse over a period of 24 h. Here too, however, the correlation was still quite poor (Fig 8B; r = 0.25).
The degree to which synapses formed between the same axon onto the same dendrite have similar sizes was explored as part of a recent study in which a small volume of mouse neocortex was reconstructed in full by serial section electron microscopy [21]. All data obtained in that study were made publicly accessible, allowing us to compare our findings, obtained in living cells, in cell culture, by light microscopy, to data obtained in fixed tissue, in vivo, by means of state-of-the-art electron microscopy. To that end, we identified in the aforementioned data set groups of spine synapses made by a particular axon onto a particular dendrite (CISD synapses) and plotted, for each synapse in each CISD pair, its PSD size and spine volume against the PSD size and spine volume of its CISD counterpart (124 pairs; Fig 8C and 8D, respectively). As these figures show, the size similarity of in vivo CISD synapses was no greater than the similarity of CISD synapses in culture. In fact, the correlation (r = 0.23) between PSD sizes for synapses belonging to the same CISD pair was identical to the correlation observed in our study for PSD-95:mTurq2 fluorescence at synapses belonging to the same CISD pair (Fig 8A and 8B).
The data presented here suggest that the covariance of size changes for synapses that share similar activity histories is greater than that of synapses formed on the same neurons or dendrites that differ in their activity histories. At the same time, the data suggest that the covariance of size changes for synapses that share similar activity histories is significantly smaller than what might have been expected if synaptic remodeling was solely determined by activity histories. Our data thus allow for a conservative estimation of the maximal relative contribution of activity history-dependent processes to glutamatergic synapse remodeling in our system (Fig 8E). For this estimation, we used (1) the highest average covariance values obtained here for CISD synapses (pooled data, r = 0.34 and ρ = 0.34; Fig 6E and S4E Fig, respectively), as these represent synapses whose activity histories are probably the most similar in our data sets; (2) the lowest average covariance values obtained here for non-CI synapses (pooled data, r = 0.06 and ρ = 0.05; Fig 3B and S1B Fig, respectively), as these represent the lowest possible contributions of (postsynaptic) neuron/dendrite-wide, nonspecific processes (and possibly of some residual shared activity); and (3) the maximal average correlation measurable in our system (r = 0.78, ρ = 0.76; Fig 7E). Using Pearson’s and Spearman’s correlations, respectively, the relative contributions of specific activity histories might thus be estimated as follows:
0.34−0.060.78≈0.36 (Pearson’s), and 0.34−0.050.76≈0.38 (Spearman’s)
The contributions of spontaneous processes that occur autonomously at each synapse can then be estimated as follows:
0.78−0.340.78≈0.56 (Pearson’s), and 0.76−0.340.76≈0.55 (Spearman’s)
This analysis suggests that under our experimental conditions, the ratio of contributions by activity history-dependent and -independent processes to synaptic remodeling is, at most, about 2:3. Put differently, the “signal-to-noise ratio” of activity history-dependent synapse remodeling is approximately 0.7, i.e., less than one.
In the current study, we set out to compare the contributions of specific activity histories to the size remodeling of glutamatergic synapses, with the contributions of processes occurring irrespective of specific histories. To that end, we examined the covariance of changes in the sizes of CI synapses, that is, synapses formed between the same axon and the same neuron or dendrite, under the assumption that the specific activity histories of such synapses will be very similar; we then compared this covariance to that of synapses formed between different axons (non-CI synapses), which presumably differ in their activity histories, and to the maximal covariance measurable in our experimental system. We found that the size covariance of CI synapses was higher than that of non-CI synapses in both monolithic and modular networks; yet the average covariance of CI synapses was rather modest in comparison to what might have been expected had remodeling been dictated exclusively by specific activity histories. Indeed, comparisons of the momentary and time-averaged sizes of CI synapses revealed that the sizes of synapses with nearly identical activity histories correlated rather poorly, in perfect agreement with electron microscopy-based measurements of CI synapse sizes in the intact mouse cortex. A conservative compilation of covariance data for CI and non-CI synapses, and comparisons with maximal covariance values measurable in our system, suggests that only about 40% of glutamatergic synapse size remodeling could have been attributed to specific activity histories, and thus the contributions of other processes, including neuron-wide, nonspecific processes and other, possibly stochastic, synapse-autonomous processes, were at least as great.
The interpretation of the experiments described here is based on several key assumptions that warrant some discussion.
The first assumption concerns the identification of CI synapses as such. This identification was based on the juxtaposition of pre- and postsynaptic synaptic proteins tagged with fluorescent groups and, thus, ultimately on light microscopy. Several studies (e.g., [21,44]) have suggested, however, that the proximity of an axon to a dendrite or a spine observed by light microscopy might not reliably predict the presence of a synapse, not even in a statistical sense (e.g., [45]). Yet it should be noted that this conclusion pertained to proximities of axons and spines, visualized by volume-filling dyes or electron microscopy, whereas, here, the presence of a synapse was deduced from the juxtaposition of fluorescent foci, originating in proteins that cluster almost exclusively at pre- and postsynaptic sites; thus, the presence of a synapse was deduced here not only from physical proximity but also from the juxtaposition (in three dimensions) of pre- and postsynaptic specializations. Furthermore, unlike most proximity-based assignments, which are typically performed in fixed tissue, assignments here were based on multiple observations of the same clusters over at least 2 d, and thus small movements of axons and dendrites (which are quite common in these preparations; see [46,47]) allowed us to exclude juxtaposed pre- and postsynaptic protein clusters that did not move in unison. It is also worth noting that when we restricted the analysis to the most cleanly identifiable, bright CI synapses, the results were practically identical (compare S5 and S6 Figs with Fig 6, S4 Fig). Thus, although we cannot exclude the possibility that some CI synapses were not really innervated by the same axon, it is unlikely that our conclusions were significantly affected by erroneous assignments.
The second assumption concerns the relationships between tagged PSD-95 fluorescence, synapse size, and synaptic strength. A good correspondence between spine volume and synaptic strength has been established in multiple studies (e.g., [48–51]). Similarly, good correspondences between spine volume, PSD size, as well as α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA)-type glutamate receptor content have been shown repeatedly (e.g., [18,22,52–54]; reviewed in [23,55]). Finally, an excellent correspondence between tagged PSD-95 fluorescence, measured by light microscopy, and PSD area, measured for the same synapses by electron microscopy, was recently reported [15]. As PSD area is thought to correlate with the number of synaptic glutamate receptors, [56] tagged PSD-95 fluorescence might represent an acceptable surrogate of synaptic strength [55]. Yet, under some circumstances, for example after strong stimuli that drive spine enlargement, spine volume is temporarily decoupled from PSD size and postsynaptic scaffold molecule contents, which “catch up” on a slower timescale (2–3 h; [18,19]); this uncoupling might indicate that synaptic strength is not always predicted correctly by PSD size; by extension, it remains possible that synaptic strength is more stable than measurements of PSD-95 content would seem to suggest, because other processes (for example, changes in glutamate receptor numbers) acting over faster timescales maintain synaptic function within precise limits. How such processes might achieve this over a background of varying scaffold size, however, is not clear. Furthermore, repeated electrophysiological measurements of the same synaptic connections point to significant spontaneous fluctuations in connection strengths over comparable timescales (12 h or less), even when activity or synaptic transmission is blocked pharmacologically (e.g., [57–59]). It thus seems more likely that tagged PSD-95 fluorescence measurements such as those used here and elsewhere provide low-pass filtered estimates of synaptic strength, which underestimate, rather than overestimate, fluctuations in synaptic strength. In this respect, it is worth noting that synaptic contents of AMPA-type glutamate receptors seem to fluctuate at least as much as synaptic PSD-95 contents do ([11]; see also [60]).
The third assumption concerns the premise that CI synapses have similar activation histories when these are integrated over many days. Although this is a very reasonable premise [21,22], it is not perfect. Ignoring for the moment the statistical nature of neurotransmitter release (which would probably average out over these long timescales; [22]), presynaptic sites of cultured hippocampal neurons have been shown to exhibit significant functional variability even for sites located along the same axons (e.g., [61–65]; reviewed in [20]). As a result, activity histories of CI synapses might not be as similar as presumed, which might partially explain the modest covariance of their remodeling. We note, however, that this is an unlikely explanation. First, it was shown that presynaptic functional properties of nearby synapses formed by the same axons on the same dendrites are much more similar than those formed on different dendrites [61,63]. Consequently, had presynaptic variability been at the source of differential remodeling, the covariance of CISD synapse remodeling should have been much higher than that observed for CI synapses. Such a difference, however, was not apparent (compare Fig 6B, 6C, 6E and 6F; S4B, S4C, S4E and S4F Fig). Along this line, measurements made in cortical neurons suggested that release properties of presynaptic sites formed by the same axons on the same neurons are remarkably similar, even when such synapses are formed on different dendrites of target neurons (a phenomenon referred to as Normalization of Release Probability; [42]). Perhaps more importantly, however, here (Fig 3) and elsewhere (e.g., [5,6,11,13,66]), it was shown that synaptic remodeling continues at significant rates even when activity and/or synaptic transmission is blocked. It therefore seems more likely that the modest remodeling covariance observed for synapses belonging to the same CI pairs is due to spontaneous remodeling processes occurring autonomously at each synapse, independent of activity, specific or otherwise. In fact, we suspect that the variability of presynaptic functional parameters observed in studies such as those mentioned above might be the outcome, rather than the cause, of such spontaneous remodeling processes. This would not be surprising, given the strong coupling between PSD and active zone remodeling (e.g., [13,18,67,68]; see also [52,59,69]). Moreover, the skewed (heavy tailed) distributions of presynaptic properties [61–65,69] as well as previously reported features of active zone remodeling dynamics [13,16] are archetypical hallmarks of a stochastic process known as the Kesten process, which was previously shown to capture the spontaneous remodeling processes of excitatory [16] and inhibitory [17] synapses.
The fourth concerns the assumption that the differences between CI and non-CI synapse covariance stemmed entirely from differences in commonality of activation histories. Even in modular networks, however, some network-wide bursts did spread from one module to the other, suggesting (as elaborated on above) that activity histories of non-CI synapses were not entirely dissimilar. This might have led us to overestimate the covariance contributed by (postsynaptic) neuron/dendrite-wide, nonspecific processes (non-CI synapse remodeling; gray regions in Fig 8E). We note, however, that the very low values of these estimates (7% to 8%) leaves little room for lowering them further. Conversely, the experiments described in Fig 3 and S1 Fig hint that, even in inactive networks, remodeling covariance of CI synapses might still slightly exceed that of non-CI synapses. It thus remains possible that some of the covariance exhibited by CI synapses reflects contributions of factors other than common activity histories. Thus, for example, spontaneous neurotransmitter release from presynaptic boutons belonging to the same axons could be coordinated by a variety of intra-axonal processes, such as molecular and synaptic vesicle interchange [2,70]. Similarly, processes acting extracellularly (such as receptor spillover and retrograde messengers) might act preferably on CI synapses even in the absence of overt activity. Consequently, we may have overestimated the contributions of specific activity history-dependent processes to synaptic remodeling (Fig 8E, green sectors). It is important to note, however, that neither of these deviations from the underlying assumptions would affect our estimations of the largest component, that is, synapse-autonomous, spontaneous remodeling processes (Fig 8E, blue sectors), and, thus, these deviations were unlikely to significantly impact the main conclusions of this study.
Finally, it should be noted that our experiments were carried out in networks of dissociated rat cortical neurons in primary culture. In the context of this study, the system was advantageous not only because of the experimental access it provided but also because it allowed us to focus on activity history dependence of synaptic remodeling in a manner free from other influences such as neuromodulation. Yet, it might be asked to what degree the conclusions reached here apply beyond our experimental system. As described in the introduction, the observation that spine volumes and PSD sizes fluctuate in the intact brain is well documented. Where such observations are concerned, however, it remains unknown what fraction of these fluctuations represents bona fide history-dependent synaptic remodeling and which represents other, possibly stochastic, processes. Nevertheless, we note that history-dependent remodeling processes should ultimately control synaptic size, and when sizes of synapses with apparently identical histories are compared, their sizes correlate quite poorly, not only in our data but in the intact mouse neocortex as well (Fig 8; [21]). A similar finding emerges from another recent electron microscopy study for a much smaller data set (17 pairs of CI synapses from adult rat hippocampi [22]). Here, it was found that, on average, within each CI pair, spine volumes and PSD areas differed by factors of ~2 and ~3, respectively. We thus cautiously suggest that our findings, obtained in culture, might apply to the intact brain as well, although the actual fraction of activity history-dependent remodeling might differ somewhat and vary, perhaps, according to behavioral state.
The assertion that synaptic strength is defined by the history of pre- and postsynaptic activity is one of the oldest, yet widely accepted tenets of contemporary neuroscience [71,72]. One facet of this assertion concerns activity-dependent structural plasticity of synaptic connections, including changes in the sizes of existing synapses [73]. Indeed, the capacity of particular activity paradigms to drive excitatory synapse enlargement (and shrinkage) is now well established (reviewed in [23]). The stimulation paradigms used in such studies, however, are typically brief and artificial (e.g., tetanic or theta burst stimulation, glutamate uncaging in low extracellular Mg2+, sometimes in the presence of various pharmacological agonists). In contrast, relationships between histories of protracted, more natural activity forms and synapse remodeling are less established. In the current study, we show that sizes of synapses with shared activity histories co-vary more than sizes of synapses with different activity histories (Figs 3 and 6, S1 and S4 Figs), thus demonstrating that histories of spontaneously occurring activity forms can significantly affect synaptic remodeling as well.
Somewhat surprisingly, our data also suggest that the contributions of shared activity histories and the contributions of spontaneous processes to synaptic size remodeling are of comparable magnitudes. It might be argued that other activity regimes, which differ from those present in the networks studied here, might be more effective in controlling synaptic sizes. Had this been the case, however, it might have been expected that sizes of CI synapses would be more similar in vivo as compared to the situation in culture, given that activity regimes in vivo are richer and more physiologically relevant. Contrary to these expectations, however, sizes of CI synapses in vivo were no more similar than the sizes of CI synapses in culture (r = 0.23; Fig 8). It might be further argued that this poor correlation is attributable to local details such as bouton to bouton variability, as described above. We note, however, that such sensitivity to local details would further undermine the notion that predictable relationships exist between synaptic remodeling and particular histories of pre- and postsynaptic activities. We thus suspect that, regardless of activity regime, the governance of synapse remodeling by particular activity histories is partial at best. In this respect, it is worth noting that the magnitude of PSD enlargement induced by the aforementioned experimental stimuli (50% or less; e.g., [18,19,74]) is not that different from the magnitude of spontaneously occurring changes in PSD sizes observed in the intact cortex of mice merely maintained in their home cages (44% on average; [15]). It is also worth noting that both covariance measures used here (i.e., Pearson’s correlation and, even more so, Spearman’s rank correlation) only quantify similarities in the trends of synaptic remodeling but are quite indifferent to the similarities in the absolute magnitudes of such remodeling. Thus, the governance of synapse remodeling by particular activity histories might be even more limited than our estimates indicate.
How then can one reconcile the overwhelming evidence for activity history-dependent synapse remodeling with such a significant degree of spontaneous remodeling? How can persistent functions be embedded in neuronal networks if directed and spontaneous changes in synaptic sizes are of similar magnitude? This experimental and conceptual gap might be partially bridged by considering the following matters.
The first matter concerns the growing appreciation that synaptic plasticity is affected or even gated by various neuromodulatory systems [72,75,76]. The absence of neuromodulatory systems in the networks used here was useful to examine the net contributions of particular activity histories; yet, in the intact brain, timed neuromodulator release might significantly enhance the contributions of specific activity histories and thus minimize the relative contributions of spontaneous remodeling, at least during behaviorally important time windows. We note once again, however, that this entails an expectation that CI synapse sizes would be quite similar in vivo, an expectation that is not matched.
A second matter concerns the fact that functional connections between neurons are often based on multiple synapses (reviewed in [20]), and, thus, activity history-independent fluctuations at individual synapses might average out at the level of neuron-to-neuron connections. Furthermore, changes in connection strength might be most reliably modified by increasing or decreasing the number of synapses connecting two neurons (e.g. [21]; see also [77]). Indeed, it has recently been shown that numbers of synapses formed between particular axons and dendrites are very different from what might be expected by chance [21,44]. It remains to be seen, however, if the time course over which synapses are added/removed, the actual numbers of such synapses, and the signal-to-noise ratios of multiple synapse connections can satisfactorily address the discordance described above. In this regard, it worth noting a recent in vivo (mouse) study in which basal rates of spine formation and loss were found to be almost unaffected by chronic blockade of calcium channels and N-methyl-D-aspartate (NMDA) receptors [66].
A third matter to consider is the possibility that persistent changes in network function involve vast numbers of synapses and neurons such that fluctuations at the individual synapse level are mitigated by massive redundancy [78] or rendered insignificant by the sheer numbers of synapses involved. Indeed, a recent study provided evidence suggesting that the acquisition of a new motor skill in mice involves about 4,700 motor cortex neurons and about 410,000 synapses [79]. In this regard, it is interesting to note that in his influential monograph, Hebb [73] considered this matter and suggested that, although stochastic processes might preclude predictable actions in small parts of the system, statistical constancies might emerge in larger systems. Indeed, when large numbers of synapses are followed over time, their remodeling dynamics do seem to obey certain well-defined statistical rules [5,8,16].
A final matter to consider is the possibility that stochastic changes in synaptic properties are crucially important components in the organization of network learning, as they enable networks to explore and sample synaptic configurations for those most congruent with input from the external world or with desired functions [80]. This recent study suggests that changes in synaptic weights are driven not only by deterministic, activity-dependent rules (and biological constraints) but also by stochastic processes, which dramatically improve the ability of networks to generalize and compensate for unforeseen changes. Within this context, our finding that the magnitudes of deterministic and stochastic components are comparable would seem to suggest that the contribution of exploratory processes is at least as significant as the contribution of deterministic processes, lending further support to this emerging view of synaptic plasticity.
All experiments were performed in primary cultures of rat neurons prepared according to a protocol approved by the "Technion, Israel Institute of Technology Committee for the Supervision of Animal Experiments" (ethics approval number IL-019-01-13).
The thin glass MEAs (MultiChannelSystems—MCS, Germany) used here for monolithic networks contain 59 flat, round electrodes made of titanium nitride arranged in an 8 x 8 array with an inter-electrode spacing of 200 μm. In this arrangement, the corner electrodes are missing, and one of the leads is connected to a large reference (ground) electrode. Although the recording and reference electrodes are opaque, the very thin glass (180 μm) substrate and the Indium Tin Oxide leads are fully transparent, allowing excellent optical access to the cells growing on the array.
Modular MEAs were prepared using 4Q, commercially available MEAs (MultiChannelSystems) fabricated to our request on thin glass. Apart from their layout, 4Q MEAs used here were identical to the thin glass MEAs described above. A PDMS insert was sealed onto the MEA surface, effectively dividing the MEA into two modules, separated by a number of thin channels (similar to the method described in [38]). The PDMS inserts were made using a silicon mold microfabricated using standard, single-layered SU8 photolithography techniques [81,82]. Briefly, SU-8 2002 (Microchem, Inc.) was spun on a 4-inch silicon wafer at a nominal thickness of 3 μm, baked, exposed with a dark-field transparency channel mask, baked again, and developed. Each mold had multiple barrier patterns with channel numbers ranging from 6 to 12 (3 μm x 13 μm x 400 μm; H x W x L). The mold was silanized ([tridecafluoro-1,1,2,2-tetrahydroocytl]-1-trichlorosilane evaporated for 1h in vacuum) to allow easier release and slowly filled with PDMS silicone rubber (Sylgard 184; 10:1 ratio of pre-polymer [base]: cross-linker [curing agent]; Dow-Corning, Midland, Michigan), to 2-mm height and de-gassed in a vacuum desiccator. Once the PDMS spread over the entire wafer, it was cured for 3h at 65°C. Following curing, 17-mm diameter circular barriers were cut, and two 5-mm-wide wells were punched on each side of the channels; the finalized inserts were then stored for future use. On the day of cell culture preparation, each barrier was aligned to the electrodes of pre-coated 4Q MEA dishes (see below) using a drop of 70% ethanol and heated for 2h at 54°C to allow ethanol evaporation and PDMS sealing. Finally, the dishes were cooled to 37°C in a cell culture incubator.
Primary cultures of rat cortical neurons were prepared as described previously [6]. Briefly, cortices of 1–2-d-old Wistar rats of either sex were dissected and dissociated by trypsin treatment followed by trituration using a siliconized Pasteur pipette. For monolithic cultures, a total of 1–1.5 x 106 cells were plated onto thin-glass MEA dishes, the surfaces of which had been pretreated with polyethylenimine (PEI, Sigma) to facilitate cell adherence. Modular cultures were prepared on 4Q thin glass MEAs described above (see also [37]) as follows: 100 μl aliquots of cells in suspension (at 1–1.5 x 106/ml cells) were infected with predetermined amounts of viruses and incubated for 2 h in a tissue culture incubator at 37°C. Following the incubation, the infected cells were spun down for 60 s at 2,000 g, and 60 μl of the supernatant were replaced with pre-warmed culture medium, and the cells were resuspended by gentle pipetteation. The process was repeated two more times (three washes in total). After the third spin down and resuspension, the cells were pipetted thoroughly, and 20–25 μl of cells in suspension were seeded in their respective module. No contact was allowed between the two droplets. 160 μl of uninfected cells at similar concentrations were seeded dropwise at the dish perimeter (outside the PDMS barrier) to enrich the environment with diffusive nutritional factors. Dishes with droplets were put in 10-cm petri dishes containing small vessels with water (to maximize humidity) and incubated overnight in a humidified tissue culture incubator at 37°C in a gas mixture of 5% CO2, 95% air. The next morning, 2 ml of culture medium were added to each dish.
Both uniform and modular preparations were kept in a humidified tissue culture incubator and grown in medium containing minimal essential medium (MEM, Sigma), 25 mg/l insulin (Sigma), 20 mM glucose (Sigma), 2 mM L-glutamine (Sigma), 5 mg/ml gentamycin sulfate (Sigma), and 10% NuSerum (Becton Dickinson Labware). Half of the volume was replaced three times a week with feeding medium similar to the medium described above but devoid of NuSerum, containing a lower L-glutamine concentration (0.5 mM) and 2% B27 supplement (Invitrogen).
All DNA constructs (except GCaMP6s; see below) were introduced into neurons using third generation lentiviral expression vectors based on the FUGW backbone [83]. The construct used for expressing PSD-95:EGFP (FU-PSD-95:EGFP-W) was described in detail in [6]. The construct used to express Cer:SV2 (FU-Cer:SV2a-Wm) was made as follows: FUGW was modified to FUGWm by moving the XhoI site from the 3’ to the 5’ side of the woodchuck hepatitis post-transcriptional regulatory element (WPRE). Cerulean [30], flanked with AgeI (5') and BsrGI (3') sites, was synthesized de novo and inserted into FUGWm instead of EGFP using the AgeI and BsrGI sites, resulting in the interim construct FUCWm. SV2a was then cut out of FU-EGFP:SV2a (a generous gift by Craig C. Garner; [84]) using BsrGI (5’) and XhoI (3’) sites and inserted into FUCWm, resulting in FU-Cer:SV2a-Wm. Sequencing confirmed 100% identity with Rattus norvegicus SV2A (GenBank accession: L01788.1). The construct used to express PSD-95:mTurq2 was made as follows: Large-scale gene synthesis was used to synthesize a fusion of PSD-95 and mTurquoise [85] flanked by of AgeI (5’) and EcoRI (3’) as detailed in [13], and this segment was inserted into FUGWm instead of EGFP using the AgeI and EcoRI sites. A point mutation was then inserted to convert mTurquoise into mTurquoise2 (Isoleucine to Phenylalanine; [40]). Sequencing confirmed 100% identity with Rattus norvegicus discs large homolog 4 (NM_019621.1). All cloning and gene synthesis was done by Genscript (Piscataway NJ, US). The construct used to express EGFP:SynI (FU-Syn:EGFP-W) was provided as a generous gift by Craig C. Garner [86].
GCaMP6s [41] was expressed using an Adeno Associated Viral (AAV) vector obtained from the Penn Vector Core (University of Pennsylvania).
Lentiviral particles were produced in house as previously described [17]. Briefly, HEK293T cells were transfected using Lipofectamine 2000 (Invitrogen), a mixture of the three ViraPower kit packaging plasmids (Invitrogen), and the expression vector. Lentiviral stocks were prepared by collecting the supernatant after 48 h, filtering it using 0.45-μm filters, and storing it as small aliquots at -80°C.
Transduction of monolithic cortical cultures was performed at 5 d in vitro by adding 10–20 μl of lentiviral stock solution to each MEA dish. Transduction of modular cultures was performed as described above.
MEA network activity was recorded using a commercial 60-channel headstage (Inverted A1060, MCS). Signals were first amplified by the internal headstage amplifier (1024x), multiplexed into 16 channels, amplified further (x10) by a 16-channel amplifier (Alligator technologies, US), and then digitized by an A/D converter (Microstar Laboratories, US) at 12 KSamples/sec per channel. Software used for data acquisition and display was based on AlphaMap (Alpha-Omega, Israel). Spiking activity data were stored as threshold crossing events (threshold = -40 μV) and analyzed offline using custom scripts written within the Matlab (MathWorks, US) programming environment.
Fluorescence and brightfield images were acquired using a “homemade” confocal laser scanning microscope built around a Zeiss Axio Observer Z1. All imaging was carried out using a 40×, 1.3 N.A. Fluar objective (Zeiss). The system, controlled by software written by one of us (NEZ), allows for automated, multisite time-lapse microscopy. The MEA headstage described above was attached to the system’s motorized stage (Märzhäuser Wetzlar, Germany), and the MEA dishes were placed firmly within it.
PSD-95:mTurq2 and Cer:SV2 were excited using a 457-nm solid state laser (Cobolt, Sweden). PSD-95:EGFP, EGFP:SynI, and GCaMP6s were excited using a 488-nm solid state laser (Coherent, US). Fluorescence emissions were filtered through 467–493-nm and 500–550-nm bandpass filters (Semrock, US and Chroma Technology, US). Laser intensity modulation of the 457-nm solid state laser in experiments such as those described in Fig 7 was performed using the digital interface and software provided by the manufacturer.
Time lapse recordings were typically performed by averaging five frames collected at 10–11 focal planes (0.9 μm apart). Images were collected at a resolution of 640 x 480 pixels, 12 bits/pixel. The confocal aperture was kept fully open to minimize illumination intensities. The software-controlled motorized stage was used to collect data sequentially from up to 12 predefined locations. PSD-95:EGFP was imaged at 30–60-min intervals and Cer:SV2 at 7.5-h intervals. PSD-95:mTurq2 was imaged at 1-h intervals and EGFP:SynI at 1–3-h intervals. Focal drift was corrected before collecting data from each location by automatically locating the glass/medium interface plane and moving the focal position to a user-defined offset above this plane.
GCaMP6s transduced axons were imaged for ~1 min (600 frames, ~130 msec per frame) using a cooled EMCCD (Andor) controlled by custom written software.
To maintain neuronal network viability, the MEA dishes were covered with a “cap” equipped with ports through which sterile air mixtures and perfusion media were introduced and removed [6,17,24]. In addition, the cap was equipped with a dipping reference electrode made of thin platinum wire and a removable transparent glass window. The preparations were continuously perfused with feeding media at a rate of 2 ml/day using silicone tubes connected to the cap through the aforementioned ports and an ultra-slow peristaltic pump (Instech Laboratories Inc., US). In addition, a 95% air/5% CO2 sterile mixture was streamed continuously into the dish at rates regulated by a high-precision flow meter (Gilmont Instruments, US). The MEA dishes were heated to 36–37°C by the heating base at the bottom of the headstage/amplifier and by a custom objective heater as previously described [17].
To minimize perturbations, all pharmacological agents were added to 100 μl media drawn from the MEA dish by temporarily removing the aforementioned caps glass window. The media was then returned and mixed gently using a sterile pipette, followed by returning the removable glass window. The same reagents were then added to the perfusion media at identical final concentrations, which were 1 μM for TTX; (Alomone Labs) and 20 μM for CCh (Sigma).
Analysis of imaging data was performed using an application (“OpenView”) written by one of us (NEZ). This application provides features for automated tracking of punctate fluorescent spots in time series of multiple images and the quantification of their fluorescence over time (see [24] for further details). 9 × 9 pixel (~1.3 x 1.3 μm) regions of interest (“boxes”) were centered on postsynaptic puncta, and average pixel intensities within these boxes were obtained from maximal intensity projections of all focal (Z) sections. As the reliability of automatic tracking was not absolutely perfect, all tracking was verified and, whenever necessary, corrected manually. Puncta for which tracking was ambiguous were excluded. Concomitant juxtaposition of marked presynaptic puncta was verified at every relevant time point and Z section. Whenever presynaptic puncta disappeared (even for a single time point) or became separated from their putative postsynaptic counterpart (both in XY or Z plane), the data for this synapse were excluded. Identification of CI synapses as such was limited to short, relatively straight axonal stretches, which did not intersect with other axons within the short stretch. To further facilitate CI disambiguation, low-magnification images of the imaged areas were collected for the purpose of resolving the branching structure of labeled axons and determining if axonal segments could be traced back to common origins. By keeping axonal labeling as sparse as possible, these procedures allowed for high-confidence CI synapse identification. Analysis of GCaMP6 time series was performed by first averaging four frames obtained between network bursts and thereafter subtracting these images from all images in the time series. GCaMP6 fluorescence was then quantified using OpenView as described above.
Covariance of CI synapses was calculated after smoothing PSD-95:EGFP or PSD-95:mTurq2 data with a 2.5- to 3-h low-pass filter (depending on imaging frequency). For CI synapse pairs (Figs 3 and 4 and 6D–6F, S4D–S4F, S5C and S5D and S6C and S6D Figs), covariance of CI synapses was calculated for the two synapses belonging to each pair, whereas covariance for non-CI synapses was calculated for CI1 to Ref1, CI2 to Ref2, CI1 to Ref2, and CI2 to Ref1 (four comparisons) to minimize potential effects of inter-synaptic distance. For multiple CI synapses formed between one axon and any dendrite (Fig 6A–6C, S4A–S4C, S5A and S5B and S6A and S6B Figs), covariance values for all possible CI pairs and CI-Ref pairs were calculated. Pearson’s and Spearman’s covariance values were calculated using Matlab and Microsoft Excel (using the Real Statistics Resource Pack; http://www.real-statistics.com). Data compilation, statistical testing, and plotting were performed using Microsoft Excel (and Real Statistics). Image examples (Figs 2, 5 and 7 and S2 Fig) were prepared using OpenView and Adobe Photoshop. Final figures were prepared using Microsoft PowerPoint.
Sampled raw activity measurements were analyzed using custom written scripts in Matlab. Briefly, specific algorithms were used to identify bursting activity in each module (defined as activity in at least 25% of active electrodes in the module during 300-msec windows). A successful propagation of a burst from one module to the other was defined as a burst initiated in one of the modules followed by the appearance of a burst in the second module with a delay of no more than 50 msec between first spikes in each burst. The inter-modular synchronization measure S was calculated as
S= (BjB1+B2+Bj)
where Bj is the number of joint bursts, and B1 and B2 are the number of bursts in modules #1 and #2, respectively, which did not propagate into the second module.
|
10.1371/journal.pntd.0000072 | Genetic Selection of Low Fertile Onchocerca volvulus
by Ivermectin Treatment | Onchocerca volvulus is the causative agent of
onchocerciasis, or “river blindness”. Ivermectin has
been used for mass treatment of onchocerciasis for up to 18 years, and
recently there have been reports of poor parasitological responses to the
drug. Should ivermectin resistance be developing, it would have a genetic
basis. We monitored genetic changes in parasites obtained from the same
patients before use of ivermectin and following different levels of
ivermectin exposure.
O. volvulus adult worms were obtained from 73 patients
before exposure to ivermectin and in the same patients following three years
of annual or three-monthly treatment at 150 µg/kg or 800
µg/kg. Genotype frequencies were determined in
β-tubulin, a gene previously found to be linked
to ivermectin selection and resistance in parasitic nematodes. Such
frequencies were also determined in two other genes, heat shock
protein 60 and acidic ribosomal protein, not
known to be linked to ivermectin effects. In addition, we investigated the
relationship between β-tubulin genotype and female
parasite fertility. We found a significant selection for
β-tubulin heterozygotes in female worms. There
was no significant selection for the two other genes. Quarterly ivermectin
treatment over three years reduced the frequency of the
β-tubulin “aa” homozygotes
from 68.6% to 25.6%, while the
“ab” heterozygotes increased from 20.9% to
69.2% in the female parasites. The female worms that were
homozygous at the β-tubulin locus were more fertile
than the heterozygous female worms before treatment (67% versus
37%; p = 0.003)
and twelve months after the last dose of ivermectin in the groups treated
annually (60% versus 17%;
p<0.001). Differences in fertility between
heterozygous and homozygous worms were less apparent three months after the
last treatment in the groups treated three-monthly.
The results indicate that ivermectin is causing genetic selection on
O. volvulus. This genetic selection is associated with
a lower reproductive rate in the female parasites. We hypothesize that this
genetic selection indicates that a population of O.
volvulus, which is more tolerant to ivermectin, is being selected.
This selection could have implications for the development of ivermectin
resistance in O. volvulus and for the ongoing
onchocerciasis control programmes.
| Onchocerca volvulus is the causative agent of onchocerciasis, or
“river blindness”. Ivermectin has been used for mass
treatment of onchocerciasis for up to 18 years, and recently there have been
reports of poor parasitological responses to the drug and evidence of drug
resistance. Drug resistance has a genetic basis. In this study, genetic changes
in β-tubulin, a gene associated with ivermectin
resistance in nematodes, were seen in parasites obtained from the patients
exposed to repeated ivermectin treatment compared with parasites obtained from
the same patients before any exposure to ivermectin. Furthermore, the extent of
the genetic changes was dependent on the level of ivermectin treatment exposure.
This genetic selection was associated with a lower reproductive rate in the
female parasites. The data indicates that this genetic selection is for a
population of O. volvulus that is more tolerant to ivermectin.
This selection could have implications for the development of ivermectin
resistance in O. volvulus and for the ongoing onchocerciasis
control programmes. Monitoring for the possible development and spread of
ivermectin resistance, as part of the control programmes, should be implemented
so that any foci of resistant parasites can be treated by alternative control
measures.
| Onchocerca volvulus is the filarial nematode, transmitted by
Simulium flies, that causes human onchocerciasis, or
“river blindness”. It is estimated that 37 million people,
mostly in Africa, are infected with this worm [1]. At present, ivermectin
(IVM, Mectizan) is the only safe drug available for mass treatment of
onchocerciasis. IVM, administered at the standard dose of 150 µg/kg, has a
rapid effect on the embryonic stage of the parasite, the microfilariae (mf), which
cause most of the ocular and cutaneous manifestations of the disease. As a result of
this microfilaricidal effect, the skin microfilarial loads decrease by
95–99% within one month after treatment. The drug also blocks
the production of new mf by the adult female worms, who only resume mf release
3–6 months after treatment. This “embryostatic effect”
of IVM explains why the mf loads remain at very low levels for up to one year.
Furthermore, IVM treatments repeated at 1- to 3-monthly intervals have some, though
moderate, effect on the longevity of the adult worms (“macrofilaricidal
effect”) [2,3].
The drug, when given repeatedly, is therefore acting on at least three components of
parasite fitness: reproduction, microfilarial survival and adult parasite lifespan,
which together affect morbidity and the intensity of transmission. Due to the
limited macrofilaricidal effect of the drug, treatments must be repeated and
sustained. Endemic communities in Africa receive annual IVM treatment, while those
of Latin America receive semi-annual treatments. To date, more than 400 million
treatments have been distributed in Africa [4], with some individuals
having received up to 18 annual treatments.
Due to this enormous drug pressure on the parasite, there is a risk of resistance of
O. volvulus to the drug [5–7]. This concern is
justified by reports of suboptimal responses to IVM from Sudan [8] and Ghana [9,10], although in the
former report reduced immune responsiveness in some of the treated people has been
suggested as a possible explanation for the suboptimal responses to IVM. And in the
study in Ghana the poor responses have been attributed to the parasites, with adult
female worms resuming microfilarial production earlier after treatment than
classically described. More recently, another report in Ghana [11] shows the
first unequivocal parasitological and epidemiological evidence of ivermectin
resistance in O. volvulus populations.
In addition to this evidence of IVM resistance, changes in the genetic structure of
O. volvulus populations, associated with IVM treatments, have
been observed in parasites from Ghana [12–16]. These changes occurred particularly
on the β-tubulin gene [16,17], which has been associated with IVM
resistance in the sheep intestinal nematode Haemonchus contortus
[17].
However, in these previous studies, O. volvulus from
IVM-naïve and -treated human populations were collected from different
individuals in different communities.
It is important to assess whether the genetic changes reported in O.
volvulus are associated with a reduced response to IVM in any of the three
effects of IVM on parasite fitness, described above. Furthermore, to eliminate the
possibility that differences in genotype frequencies between IVM-naïve and
-treated populations could be due to geographical effects, due to separate
individuals and communities being sampled, it is important to assess whether changes
in genetic frequency could occur in parasites collected from the same individuals
before and after exposure to IVM. Genetic changes clearly associated with treatment,
which could not possibly be associated with other covariates, would provide
unequivocal evidence of genetic selection by IVM on O. volvulus.
Such treatment-induced selection would be heritable. Heritable genetic changes that
could reduce the susceptibility of O. volvulus to any of the
effects of IVM on the parasite could have long-term consequences for the control of
onchocerciasis because there is currently no alternative drug available for mass
treatment of this disease.
In a previous study [18], we reported that in an IVM-naïve
O. volvulus population from Cameroon, adult female worms
presenting a homozygous genotype for β-tubulin were more fertile than adult
worms that were heterozygous at this locus. In the present study, we have analyzed
genetic characteristics (β-tubulin gene and two control
genes, heat shock protein 60 (hsp60) and acidic ribosomal
protein (ARP)) and phenotypic characteristics (female worm fertility)
of parasites collected, in the same individuals, before and after 4 or 13 IVM
treatments over a three-year period. These treatments were administered as part of a
clinical trial conducted in Central Cameroon. The main objective of this trial was
to assess the effects of different regimens of IVM treatment on the mortality of
O. volvulus adult worms, and the results of this phase have
been published elsewhere [3]. In the second phase, results of which are
presented in this paper, we evaluated whether repeated treatment with IVM led to (a)
genetic changes in the adult worm population and (b) any modification of the
relationship between β-tubulin genotype of the female worms
and their reproductive status.
The study was carried out in the Mbam Valley, a region hyper-endemic for
onchocerciasis, located in the Central province of Cameroon, where no IVM had
been distributed at the beginning of the study and where no vector control
activities have ever been performed. In this area, before the introduction of
IVM, the intensity of infection in the population, as expressed by the Community
Microfilarial Loads (CMFL) [19] ranged between 10 and 114 mf per skin snip
(mf/ss) [3]. The full details of the clinical trial, which was
approved by the Cameroonian Ministry of Public Health and by Merck and Co., the
manufacturer of IVM (Mectizan), have been published elsewhere [3]. The
study also subsequently received approval from the institutional review board of
McGill University. Briefly, 657 individuals were selected using the following
inclusion criteria: men between 18 and 60 years old, with at least two palpable
nodules during the preliminary examination but otherwise in good health, who had
not received any filaricidal treatment within the five previous years, and who
agreed to participate in the trial by signing an informed consent form. These
patients were randomly allocated to one of the four following IVM treatment
groups: 150 µg/kg body weight/year (standard group; group 1); 150
µg/kg/three-monthly (group 2); 800 µg/kg/year (group 3); and
800 µg/kg/three-monthly (group 4). Over the three-year study period,
patients received either 4 or 13 IVM treatments.
In order to assess the macrofilaricidal effect of IVM on O.
volvulus, adult worms were collected, by nodulectomy, at the outset of
the trial (before the first IVM dose was administered) and once again after
three years of treatment in the four different treatment groups described above.
The protocol used for parasite collection was identical for the two rounds of
nodulectomy. Just before each round of nodulectomy, each person was carefully
examined and all the sites on their body where a nodule or a group of nodules
was palpable were recorded on a body chart. Subsequently, one of the sites was
selected at random and all the nodules located at this site were removed from
each person. The site selected for the second nodulectomy was one of those
recorded at the outset of the study so that the worms collected at that time had
probably been subjected to the IVM treatments administered over the previous
three years. Just after the nodulectomy, all the nodules collected were immersed
in fixative (70% ethanol, 20% water, 10%
glycerol). One of the nodules was used for histological examination, as
previously described [3], to evaluate the status of the worms. Any
additional nodules (“extra nodule”) from the excision site
were stored in the fixative at room temperature and available for genotyping and
phenotyping.
Of the 657 individuals selected before treatment, 290 had more than one nodule at
the first nodulectomy site, and thus at least one “extra
nodule” available after the histological examination. Similarly, of
the 541 patients present at the second round of nodulectomy (following three
years of treatment), 156 had at least one extra nodule available. Patients
included in the present study were selected taking into account our objectives,
which were to assess the genotypes of three polymorphic genes, including
β-tubulin, in the adult worms, and any relationship
between the genotype of female parasites and their reproductive status, before
and after IVM treatment. To make the comparison more sensitive, we performed the
genotyping and the phenotyping only on parasites obtained from those people for
whom “extra nodules”, containing at least one adult worm,
had been collected at both nodulectomy rounds (pre-treatment and after three
years of repeated treatments). The total numbers of individuals who met these
inclusion criteria were 18 in group 1, 16 in group 2, 22 in group 3 and 17 in
group 4. Thus, the analyses were performed on the nodules collected from 73
individuals.
This procedure has been described previously [18]. In 2002, the
nodules were washed with phosphate buffered saline (PBS) for 24 h with regular
changes of medium in order to remove all residues of fixative. The nodules were
then digested in collagenase [20]. Worms were collected and stored individually
in labelled Eppendorf tubes, which were frozen at −80°C. Each
female worm was phenotyped by microscopical examination of its reproductive
status in terms of the presence of mf and embryos. Three phenotypes were
defined: (a) non-fertile females, i.e. worms with empty reproductive organs, (b)
females with low fertility, in which the reproductive organs contained only a
few embryos, but no mf, and (c) fully fertile females, in which the reproductive
organs were full of mf and embryos.
After the phenotyping, each worm was disrupted and its DNA was extracted using a
Dneasy kit (Qiagen Inc., Mississauga, Canada). Heat shock protein
60 (hsp60) (GenBank, AF121264), which is a molecular
chaperone that participates in the folding of proteins, was chosen as a control
gene. It was known to be polymorphic and previously found not to be selected by
IVM treatment in O. volvulus
[16]. Two
polymorphs (“A” and “G”) were found in
the hsp60 gene partial sequence analyzed. The region analyzed
started at position 214 on the cDNA and included 100 bp in the exon, followed by
276 bp in the intron. The A/G polymorphism was located in the intron region. The
fragment of 376 bp was amplified by PCR from individual adult worms with the
primers 5′CAA TCA TGG GGA AGT CCA AAG 3′ and 5′CTC
AAA ACC TTC CTT TGC AAT 3′ at
Tm = 53°C. PCR products were sequenced
with the hsp60 anti-sense primer using the 3730XL DNA Analyzer
system (McGill University/Genome Quebec Innovation Centre). Platinum Taq DNA
polymerase High Fidelity (Invitrogen) was used in the PCR reaction to avoid
introduction of error during amplification. Each individual chromatogram was
analyzed with Sequencher 4.7 software (Gene Codes Corporation, Ann Arbor, MI,
USA), to detect the homozygotes AA and GG and the heterozygotes AG.
The acidic ribosomal protein (ARP) gene
(GenBank, AI130565), which is involved in protein synthesis, was chosen as a
second control because it was expected to be polymorphic [21] and not known
to be sensitive to IVM treatment. Two polymorphs (“C” and
“T”) were found in the acidic ribosomal protein gene partial
sequence analyzed. The region of interest was from 1270 bp to 1488 bp of the
complete gene. It was amplified by PCR from individual adult worms with the
primers 5′ TGA AAA ACT GCT ACC GCA TA 3′ and 5′
AAA TTT TCG TTG GAA TTT GC 3′ at
Tm = 54°C. PCR products were analyzed
by restriction fragment length polymorphism, based on C/T polymorphism apparent
in the EST database, using the restriction enzyme Mnl 1 for 2 hours, and
subjected to electrophoresis on a 12% polyacrylamide gel
(39∶1) for 2 hours at 130 V, stained with ethidium bromide and
visualized using an ABI Imager (Bio-Rad, Hercules, CA, USA).
Two alleles (“a” and “b”) have been
described for β-tubulin
[16].
These two alleles have three single nucleotide polymorphisms in an exon region.
These differences lead to changes in three amino acids in the putative protein
sequence. The worms were genotyped individually for
β-tubulin (GenBank, F019886) by PCR amplification
followed by amplicon length analysis [17].
The aim of the analysis was to assess whether a variety of covariates related to
the worm, nodule or patient characteristics were associated with three different
dependent variables: (a) the inability to genotype some of the worms from the
preserved nodules; (b) the frequency of the various polymorphs analyzed; and (c)
the degree of fertility of the worm. We considered the five following
covariates: the age of the patient at the outset of the trial (continuous
variable); the CMFL in the village of residence of the patient, defined in four
categories: 10–40, 41–60, 61–70, and
71–114 mf/ss; the treatment group (for analysis of the worms collected
post-treatment: 150 µg/kg/year, 150 µg/kg/three-monthly, 800
µg/kg/year, and 800 µg/kg/three-monthly); the total number
of females in the nodule; and the total number of palpable nodules on the
patient at the outset of the trial. In addition, we also assessed the degree of
fertility in relation to the genotype of the worms and to the total number of
males in the nodule.
The procedure for genotyping the worms failed with a significant number of worms
obtained from the nodules that had been preserved at room temperature for 5 to 8
years. To test whether this inability to genotype some worms could be explained
by sampling biases, we assessed, using multivariate logistic regression, whether
the success in genotyping the worm (genotyped vs. non-genotyped status) was
associated with one or the other of the possible covariates quoted above. All
regressions analyses were performed using Stata v9.0 (Stata Corporation, TX,
USA), where parameters were estimated using the cluster option [22]
accounting for intra-nodular correlation.
Hardy-Weinberg equilibrium was tested using the χ2 test,
unless the sample size was small. In this case, Fisher's exact test was
used. The genotypic frequencies before and after treatment were compared using
Fisher's exact test. To evaluate whether some host covariates or
village characteristics may have influenced the heterozygosity of the worms, the
association between heterozygous status and the five main possible covariates
quoted above was assessed separately on pre- and post-treatment data, by
multivariate logistic regressions. Potential intra-nodule clustering was
accounted in the regression models.
Logistic regression models were used to analyze the independent variables
associated with the fertility of the female worms before and after treatment.
The dependent variable “fertility” was defined, for this
analysis, using two categories: no or low fertility versus high fertility [18]. This choice is based on the fact that only worms
with mf have the possibility of having their progeny transmitted, at the time of
sampling, and this may be relevant to the possible transmission of any
“resistant” genotypes. However, any treatment group effect
on fertility status could be due to either treatment frequency or to the fact
that the worms were collected three months after the last treatment in the
three-monthly treated groups (groups 2 and 4) and twelve months post-treatment
in the annual groups (groups 1 and 3). The possible covariates in the model
included the five host-related independent variables defined above, and two
other independent variables: the genotype of the worm at the
β-tubulin locus (homozygous versus heterozygous),
and the total number of males present in the nodule. Here again, the
intra-nodule clustering was considered in the logistic regressions. The
χ2 and Fisher's exact test analyses were
performed using VassarStats (http://faculty.vassar.edu/lowry/VassarStats.html).
A total of 73 patients provided one nodule at the outset of the trial and one nodule
after treatment. A total of 367 worms (248 females, 119 males) were isolated from
the 73 nodules collected before treatment, and 224 worms (153 females, 71 males)
were extracted from the 73 nodules provided by the same hosts after three years of
repeated treatment. Details on the numbers of worms analyzed in the different
treatment groups are given in Tables
1 and 2.
We previously showed, in a sample of 320 female worms collected before treatment
as part of the same trial, that the 90 worms that could not be genotyped for
β-tubulin did not differ significantly, with regard
to several host independent variables, from the 230 worms that could be
genotyped [18]. Similar results were obtained when comparing
the 65 non-genotyped females to the 183 genotyped ones, and the 63 non-genotyped
males to the 56 genotyped males, collected before treatment, from the 73 people
from whom nodules could be analyzed both before and after treatment.
The proportion of female worms that could not be genotyped for
β-tubulin was significantly higher after treatment
(respectively 26.2% and 35.9% before and after treatment;
p = 0.043). Among the 153
female worms collected after treatment, 55 could not be genotyped. According to
multivariate logistic regression, the “genotyped” status was
not associated with any of the five covariates included in the analysis. The
proportion of non-genotyped male worms did not differ significantly before and
after treatment (respectively 52.9% and 63.4%;
p = 0.17). After treatment, we
observed that a significantly higher proportion of male worms could be genotyped
in the 800 µg/kg/year treatment group
(OR = 3.97 (95% CI,
1.05–15.08);
p = 0.043) compared to the
standard group. None of the other independent variables differed significantly
between the genotyped and the non-genotyped male worms. Taken together, these
results do not provide evidence of bias between the 45 non-genotyped and 26
genotyped male worms according to the tested covariates.
Before treatment, β-tubulin heterozygous status was not
influenced by age of host, total number of females in the nodule, total number
of palpable nodules or CMFL in the village of residence (Table 3). After treatment, none of the tested
covariates was significantly associated with the
β-tubulin heterozygous status, except the fact of
living in a village with a CMFL between 41 and 60 mf/ss. This weak association
(OR = 6.24 (95% CI,
1.09–35.59);
p = 0.039) might indicate that
the probability of being heterozygous was higher in the villages where infection
rates were rather high (Table
4). The probability of being heterozygous tended to be higher in the
groups treated three-monthly. Even if this was not significant in the analysis
taking into account the various groups separately, this trend is consistent with
the results presented above comparing the pooled annual treatment groups and the
pooled three-monthly treatment groups.
Before treatment, the homozygote genotype was the only independent variable
associated with a high fertility phenotype (p<0.002)
(Table 5). After
treatment, high fertility of the worms was still associated with the homozygous
genotype (p = 0.035). In
addition, high fertility of the worms was more likely to be observed amongst
younger patients (p = 0.018).
Finally, high fertility in the worms was more apparent in nodules containing
higher numbers of male worms
(p = 0.030) (Table 6).
As the intervals between the last IVM treatment and nodulectomy in the
three-monthly groups and the annual treatment groups were different, they have
also been considered separately (Figure 2). Twelve months after the last IVM treatment, analysis of
the fertility (non- and low fertility versus full fertility) in relation to
genotype showed that the β-tubulin homozygous worms
remained more fertile than the heterozygous worms
(χ2 = 11.06,
p<0.001; Figure 2). Because the sample size was small, we did not perform an
analysis on the data collected on the three-monthly groups (samples collected
three months after the last IVM treatment; groups 2 and 4). However, the figure
shows that the proportion of fully fertile worms was higher in the homozygous
worms (42% compared with 24%), but both genotype groups
showed a similar proportion of non-fertile worms (50% and
48%, respectively, for the homozygous and heterozygous parasites).
With the implementation of the onchocerciasis control programmes, an increasing
proportion of people in endemic areas have received community-directed treatment
with IVM on a regular basis. Even though children under 5 years of age and pregnant
women are excluded from mass treatment, a high proportion of the parasite
population, in control areas, is under treatment. As a consequence, and because most
of the parasite population is in the human hosts rather than in the vector, only a
relatively small proportion of the O. volvulus population is likely
to be in refugia (not exposed to the drug) at the time of
treatment. Thus, selection pressure for any IVM resistance alleles is expected to be
high in O. volvulus
[7].
Parasitological and epidemiological evidence of ivermectin resistance in O.
volvulus populations has been reported in Ghana [11]. Selection on
the β-tubulin gene following repeated IVM treatment of
people infected with O. volvulus, compared with parasites from
IVM-naïve people has been found in Ghana [16,17]. Because gene selection is the first
step in the development of drug resistance, it is important to assess genetic change
in a population of parasites exposed to selection pressure.
Our study of the possible effects of IVM on the genetic structure of an O.
volvulus population is unique in several respects. We analyzed nodules
from the same patients before and after three years of IVM treatment at different
treatment frequencies, dose rates, and with complete knowledge of the number of IVM
treatments. To our knowledge, this has never been done in past investigations of the
effect of IVM on genetic selection in human parasitic nematodes. In the study area,
because there was no vector control and only a small proportion of the population
living in the area was treated during the trial, the force of infection probably did
not decrease during the trial [3].
The main finding, that IVM treatment selected for heterozygotes at the
β-tubulin locus and that this selection was dependent on the number of
doses, raises interesting questions in view of the fact that this gene has been
linked with IVM resistance in another parasitic nematode [17] and the recent evidence
that IVM resistance is occurring in O. volvulus
[11]. The period over which the IVM treatment-associated
genetic change in β-tubulin occurred was short (1994 to
1997). It takes about 1 year from microfilarial birth until an adult worm commences
production of the next generation of microfilaria. The observed genetic changes are
dramatic, given the time period and the generation interval, and could result from
differential mortality of existing adult worms and possibly a differential
establishment of new worms, dependent on the worm genotype and tolerance to the drug
pressure. These possible selective events will require further investigation. IVM
might be more toxic to the more fertile female worms, which were previously found to
be homozygous at the β-tubulin locus [18], as a result of the
effect of IVM in preventing the release of microfilariae from the uterus and the
subsequent degeneration of these trapped microfilariae.
The main limitation of the present study was that some samples could not be
genotyped. As suggested in the previous paper [18] with samples from
the same trial [3], there are several explanations for the difficulty
in genotyping some of the adult worms. The nodules had been stored in a dessicating
fixative, at room temperature, for several years before they were digested and the
DNA extracted. It is likely that some of the worms that could not be genotyped were
dead or moribund at the time that the nodules were harvested. The DNA of dead or
moribund parasites may have been degraded or fragmented, and difficult to amplify.
Dead O. volvulus are not rapidly resorbed and can be readily found
in nodules [20,23]. In the study area before treatment, 15.2% of
adult female worms found in nodules were moribund or dead [24]. After treatment,
the proportion of non-genotyped parasites was higher
(p = 0.024), which could be due to
the macrofilaricidal effect of IVM [3,25,26].
Very similar numbers of worms were genotyped for β-tubulin
and for hsp60 before and after treatment. Most of the worms that
could not be genotyped for the acidic ribosomal protein gene could also not be
genotyped for β-tubulin. However, a higher proportion of
the acidic ribosomal protein gene was genotyped compared with the
β-tubulin and the hsp60 genes. The
fact that acidic ribosomal protein has many gene copies in
eukaryotic genomes [21], whereas β-tubulin and
hsp60 typically do not occur with multiple gene copies, could
explain the difference in the ease of genotyping the acidic ribosomal
protein gene compared with the β-tubulin and
hsp60 genes.
Notwithstanding these likely reasons why some worms could not be genotyped, we have
performed several analyses to evaluate whether the population of those worms that
could not be genotyped was similar to the population of worms that were genotyped.
We showed that the two populations did not differ according to various external
covariates (host age, CMFL level, etc.).
The female worms were not dissected into different tissues, but each worm was
analyzed as a single entity. Any microfilarial DNA present within the female worm
would be included in the DNA analysis of the female worm, and so the
β-tubulin genotype frequencies of the female worms
could have resulted from microfilarial/embryo DNA contamination within the female
worm. Such daughter microfilariae/embryos would have half of their DNA derived from
that mother worm and thus reflect the genome of the mother worm. Any contribution
from male worms to the assessed genotype of a female worm via the daughter
microfilaria/embryos contained within the female worm, would, if a significant
effect, tend to increase the probability that female worm genotypes would be seen as
heterozygotes. In fact, in the female worms before treatment there was an excess of
homozygotes, indicating that any contribution from male worms to the genotype
assessed in the female worms was not significant. Given the low frequency of the
“b” allele in the male worms before and after treatment, the
increase in the frequency of “ab” heterozygotes in the female
worms, following treatment, could not be accounted for by male worm contamination
via microfilaria/embryos contained within the female worm.
Finally, the frequency of “ab” heterozygotes was greater in the
three-monthly treatment groups (69%) than in the annual treatment groups
(56%) and the pre-treatment sample (21%). IVM treatment
reduces fertility so that if microfilarial/embryonic DNA were contributing to the
assessment of DNA of the female worms, treatment should have decreased, rather than
increased, the proportion of female worms that appeared as heterozygotes. For these
reasons we do not believe that microfilarial/embryonic DNA within the female worms
had any significant influence on the observed genotype frequencies for
β-tubulin observed in the female worms.
One possible explanation for the apparent IVM selection on
β-tubulin in the female worms is the existence of a null
allele (an allele that is not being detected by the method used) that could be
distorting the genotype frequencies determined. Previously, we have found that
freshly frozen O. volvulus samples always amplified readily with
the same β-tubulin primers as have been used in this study
[17]
suggesting that no null allele exists, using the procedures followed, for
β-tubulin in O. volvulus. We do not
believe that a null allele exists to affect the genotype frequencies. Furthermore,
even if it did exist, it could not account for the change in genotype frequencies
resulting from IVM treatment. We can conclude that the observed change in
β-tubulin genotype frequency with treatment is genetic
selection as has been seen previously in O. volvulus that have been
obtained from people who have been repeatedly treated with IVM and also observed in
IVM-resistant H. contortus
[16,17].
One of the main findings of the study is that female worms, homozygous for
β-tubulin, were more fertile than heterozygote female
worms. A similar difference in fertility before treatment was reported previously
[18]. Homozygous worms, in the patients treated annually
and collected twelve months after the last IVM treatment, also had higher fertility
than heterozygotes. However, this difference decreased if the worms were collected
three months after 13 three-monthly IVM treatments, at a time when the embryostatic
effect of IVM is normally still apparent. The results observed three months after
IVM (Figure 2) could be due to a
combination of innate differences in fertility between homozygous and heterozygous
worms and the relative effects of IVM on embryostasis in the different genotypes. If
the fertility disadvantage of heterozygotes tends to disappear when the parasite is
under strong IVM pressure, this could have implications for parasite transmission
and possible resistance selection. It would be interesting to study, perhaps in
Caenorhabditis elegans, how polymorphism in
β-tubulin may affect the fertility of nematodes.
Before treatment, hsp60 gene was in Hardy-Weinberg equilibrium in
the female and male worm populations as well as for the
β-tubulin gene in the male worm population. The
acidic ribosomal protein gene normally has multiple gene copies
[21], which means that there are multiple loci in the
genome. For this reason, it is inappropriate to apply a Hardy-Weinberg equilibrium
test, as used on a single locus. The β-tubulin gene was not
in Hardy-Weinberg equilibrium in the female worms, as there was an excess of
homozygotes in this population. Various assumptions are required for Hardy-Weinberg
equilibrium including random mating. Non-random mating could occur in the O.
volvulus population and be due to inbreeding, positive assortative
mating or a subpopulation structure [27]. The inbreeding coefficient F,
which represents the proportional loss of heterozygosity due to inbreeding, was
calculated for the population of 73 patients, based on the β-tubulin data,
to be 0.41. This index of inbreeding is moderately high and has implications not
only for Hardy-Weinberg equilibrium, but also for the possible rate of selection of
a resistant population [28]. Inbreeding could be explained by the fact that
vectors transmitting to the study population were living in the local area and tend
to bite people from the same community.
It is of interest that the FIT (inbreeding coefficient
within a subpopulation) for Wuchereria bancrofti microfilariae, a
closely related human filarial nematode, was calculated on the basis of
β-tubulin genotype frequencies to be 0.44 [29].
However, other processes may be involved. An assortative mating coefficient R (which
could include inbreeding) was calculated [27] to be 0.54. The higher
fertility of the female β-tubulin homozygote
“aa” worms, when compared with the female heterozygote
“ab” worms, could be consistent with positive assortative mating
if the likelihood that a worm will mate is associated with the female
worm's fecundity. It is known that male O. volvulus
migrate between nodules and may be attracted by the fully fertile females
(predominantly homozygous). Subpopulation structure could also explain the
Hardy-Weinberg disequilibrium. Because the samples were collected in a
forest/savannah transition zone, the worms could belong either to a savannah, or to
forest or mixed forest/savannah strains. Subpopulations can result from
environmental segregation, inbreeding and/or positive assortative mating. Finally,
it is possible that in the absence of IVM treatment, the
β-tubulin heterozygote female worms die faster than the
homozygote female worms. Differential mortality between the genotypes could also
affect Hardy-Weinberg equilibrium.
The selection for the β-tubulin heterozygote
“ab”, found in female worms after IVM treatment, was more
important in the worms exposed to IVM every three months compared with the worms
exposed to IVM annually. This difference could be due to either the total number of
treatments or to the interval between treatments. These are important points for the
onchocerciasis control programmes because semi-annual or more frequent treatments
are ongoing in some areas and under consideration in other areas. An increase in
treatment frequency might increase the selection pressure.
No selection was demonstrated in the male worm population after treatment, during the
three-year study period. This result suggests that the female worms are more
susceptible to IVM pressure than the males and could reflect an increased
IVM-induced mortality in female worms that are homozygous at the
β-tubulin locus. From the Gardon et al. trial in
Cameroon [3], it is possible to estimate the relative proportional
decrease in male and female parasites due to IVM treatment over the study period.
This amounts to 13.5% and 27.5%, respectively. If the male
parasites are less affected by the treatment, the genotype frequencies in the male
population would only change significantly following transmission and subsequent
infection by progeny of the parasites under selection. Given that it takes about one
year to complete the life cycle of O. volvulus, in this situation
it would take longer for IVM selection to influence male worm genotype frequencies.
It is interesting in this regard that in Brugia malayi,
sex-dependent expression in a possible IVM receptor has been demonstrated [30]. The
putative glutamate-gated chloride (GluCl) channel gene, AF118554,
was estimated to be expressed at a 24.3-fold higher level in female worms compared
with male worms. As GluCl is thought to be the main target of IVM [31], one
could speculate that the effect of IVM might be greater in female than male worms.
This might explain why O. volvulus male worms seem to be less
susceptible to IVM and thus less rapidly selected than female worms, so that in the
short three-year time interval of this study no significant change in
β-tubulin genome frequency was seen in the male worms.
However, selection on β-tubulin in male O.
volvulus has been found when parasite populations were exposed over 6 or
more years to IVM [16,17].
There may also be an IVM-induced mortality of the incoming larvae or pre-adult stages
of the parasites, with the homozygous worms being more sensitive to repeated
treatment than the heterozygote parasites. An effect of IVM on the pre-adult stages
of O. volvulus has been suggested [32] and demonstrated in
results obtained on the bovine parasite Onchocerca ochengi
[33].
Today, intestinal trichostrongylid nematodes of livestock are commonly strongly
resistant to IVM, and the development of IVM resistance can occur rapidly in these
nematode parasites, sometimes in less than three years [34]. However,
trichostrongylid nematodes and filarial nematodes have different biology. As the
generation time is shorter in trichostrongylid nematodes than in filarial nematodes,
resistance selection would be expected to take longer to be manifested in filarial
worms. In contrast to soil transmitted nematodes, filariae have no free-living
stages and most of the population of the nematode occurs in the human host, so that
refugia is likely to be low in a community under treatment. As
a result, selection pressure for resistance to develop could be high in human
filarial nematodes under intensive drug treatment [7].
IVM has been used since the late 1980s, and more than 400 million doses have been
distributed in Africa [4]. It remains the only safe drug for community
treatment of onchocerciasis. Our results clearly show a genetic selection in
O. volvulus caused by repeated IVM treatment. Since the
parasites were collected before and after treatment from the same patients,
these results cannot be explained as differences arising from different host
populations being sampled. These results, together with other evidence of
genetic selection and reports of sub-optimal responses to IVM, provide a warning
that selection for IVM resistance could be occurring in some populations of
O. volvulus.
In view of these results, it is imperative that field studies be undertaken to
characterize all treatment responses to IVM in O. volvulus,
coupled to further genetic analysis, in order to confirm or not the possible
emergence of IVM resistance. Such longitudinal studies, which would look at the
repopulation of the skin of treated people by mf, should be undertaken without
delay if the benefits that have been achieved by the onchocerciasis control
programmes are not to be lost as a result of the spread of IVM resistance in
O. volvulus. |
10.1371/journal.ppat.1006252 | Genomic and phenotypic characterization of myxoma virus from Great Britain reveals multiple evolutionary pathways distinct from those in Australia | The co-evolution of myxoma virus (MYXV) and the European rabbit occurred independently in Australia and Europe from different progenitor viruses. Although this is the canonical study of the evolution of virulence, whether the genomic and phenotypic outcomes of MYXV evolution in Europe mirror those observed in Australia is unknown. We addressed this question using viruses isolated in the United Kingdom early in the MYXV epizootic (1954–1955) and between 2008–2013. The later UK viruses fell into three distinct lineages indicative of a long period of separation and independent evolution. Although rates of evolutionary change were almost identical to those previously described for MYXV in Australia and strongly clock-like, genome evolution in the UK and Australia showed little convergence. The phenotypes of eight UK viruses from three lineages were characterized in laboratory rabbits and compared to the progenitor (release) Lausanne strain. Inferred virulence ranged from highly virulent (grade 1) to highly attenuated (grade 5). Two broad disease types were seen: cutaneous nodular myxomatosis characterized by multiple raised secondary cutaneous lesions, or an amyxomatous phenotype with few or no secondary lesions. A novel clinical outcome was acute death with pulmonary oedema and haemorrhage, often associated with bacteria in many tissues but an absence of inflammatory cells. Notably, reading frame disruptions in genes defined as essential for virulence in the progenitor Lausanne strain were compatible with the acquisition of high virulence. Combined, these data support a model of ongoing host-pathogen co-evolution in which multiple genetic pathways can produce successful outcomes in the field that involve both different virulence grades and disease phenotypes, with alterations in tissue tropism and disease mechanisms.
| Species jumps and subsequent pathogen evolution are of increasing importance in a globally connected world. The co-evolution of myxoma virus and the European rabbit following the introduction of the virus into Australia in 1950 is the canonical case of host jumping and host-pathogen co-evolution on a continental scale. This natural experiment was repeated with the release of a separate strain of myxoma virus in Europe. On both continents moderately attenuated strains of virus became dominant while rabbits were selected for resistance to myxomatosis. Here we examine the genotypic and phenotypic evolution of myxoma virus in Great Britain compared to Australia and show that despite ecological convergence and equivalent evolutionary rates, the virus has followed distinct evolutionary pathways on both continents with few shared mutations. Furthermore, we reveal novel mechanisms of pathogenesis and tissue tropism compared to the progenitor virus, and that the disruption of virulence genes is compatible with high virulence. This suggests that mutations have occurred that can compensate for the loss of virulence genes driven by the nexus between virulence and transmission in an ongoing host-pathogen arms race.
| The establishment and spread of Myxoma virus (MYXV; genus Leporipoxvirus; family Poxviridae) in the wild European rabbit (Oryctolagus cuniculus) population of Australia in 1950 initiated the textbook case study of host-pathogen co-evolution on a continental scale [1, 2]. The virus was novel to the European rabbit having evolved in the Brazilian tapeti (Sylvilagus brasiliensis). In the tapeti MYXV induces an innocuous, localized cutaneous fibroma from which the virus is mechanically transmitted by mosquitoes or fleas. However, MYXV proteins that had evolved to suppress immune clearance and facilitate virus persistence in the natural host overwhelmed the immune system of the European rabbit causing the disseminated, lethal disease myxomatosis [2, 3].
In Australia MYXV was released into naïve rabbit populations as a biocontrol agent. The initial virus, a strain known as SLS with a case fatality rate (CFR) estimated at 99.8% [4], was rapidly replaced by moderately attenuated viruses, which by permitting longer survival of the infected rabbit were more likely to be transmitted by mosquitoes. The majority of these attenuated viruses still maintained relatively high CFRs of 70–95% [5, 6]. Simultaneously, there was very strong selection pressure for the evolution of genetically resistant rabbits [7, 8]. It is likely that the increased resistance in the rabbit population also drove selection for increased virulence in the virus to maintain transmissibility, as highly attenuated viruses transmitted poorly [9, 10, 11].
This large-scale evolutionary “experiment” is especially informative because it was repeated on a continental scale as MYXV was subsequently released in Europe. In June 1952, a landholder in France inoculated two wild rabbits with a strain of MYXV (Brazil Campinas/1949), now termed the Lausanne (Lu) strain. From this starting point, MYXV spread through the wild and domestic rabbit populations of Europe [12]. Myxomatosis was detected in wild rabbits in Britain in October 1953, probably due to the illegal release of an infected rabbit from France [13]. Despite attempts at control, the virus became established and spread throughout the wild rabbit population [14], which was eventually reduced to perhaps 1% of the pre-myxomatosis level. Strikingly, although the European release involved a different starting strain, with different insect vectors and ecological conditions, it resulted in essentially the same outcome in terms of virulence evolution [1, 12].
To facilitate evolutionary studies, field isolates of MYXV were classified into virulence grades from 1 to 5 based on average survival times (AST) in small groups of laboratory rabbits. The progenitor type viruses, killing 100% of infected rabbits, were of grade 1 virulence, while grade 5 viruses were highly attenuated with CFRs <50%. Most field isolates collected following the initial radiation in Australia were of grade 3 virulence with CFRs of 70–95% [5, 6]. The grade 3 classification was later split into grade 3A and 3B to provide greater resolution [15]. Although the initial virus isolates in Britain were of grade 1 virulence [5], attenuated viruses were detected within 12 months [16, 5].
A large scale study of the virulence of UK MYXV isolates from 1962 revealed a similar evolutionary pattern to Australia, with the majority of isolates being of grade 3 virulence [15]. Studies of UK MYXV isolates from 1975 and 1981 confirmed the predominance of grade 3 viruses, but also showed that grade 2 viruses (with CFRs of >95%) had become much more common than in Australia; over 90% of viruses tested in 1981 were grade 3A or grade 2, implying CFRs of >90% [17]. Genetic resistance to MYXV was documented much later in Britain than in Australia, but then rapidly increased in the wild rabbit population [18, 19] and may again have driven selection for higher virulence.
Although there have been detailed studies of the ecology, transmission, virulence and resistance of MYXV in Britain, little is known about the genetic and phenotypic basis of MYXV evolution and whether and how it parallels the evolutionary process seen in Australia. Indeed, previous studies have largely focused on early virus isolates sampled between 1954 and 1955 [20, 21]. To address this central question in viral evolution we determined the genome sequences of 21 MYXV isolates sampled between 2008 and 2014 in Scotland and England. Importantly, we characterise the phenotype of a number of these viruses in laboratory rabbits compared to the progenitor Lu strain and reveal major changes in disease pathogenesis.
The prototype Lu sequence [22, 23] consists of 161,777 nucleotides of double-stranded DNA with closed single stranded hairpin loops at the termini and duplicated terminal inverted repeats (TIRs) of 11,577 bp. The virus encodes 158 unique open reading frames (ORFs), 12 of which are duplicated in the TIRs.
The UK viruses descend from the Lu strain that was released into Europe as a biological control (Fig 1). The earliest sequences are from the grade 1 virulence Cornwall strain (England/Cornwall/4-54/1) isolated in April 1954 and the grade 3 Sussex strain (England/Sussex/9-54/1) from September 1954 and which quickly diverged from the introduced virus [20, 21]. This divergence is captured in a phylogenetic analysis of these viruses along with an additional early isolate (Belfast/1955) sequenced here, 21 viruses from 2008–2013 (Table 1), and a number of other European viruses (Fig 1). Notably, the viruses from Perthshire, Scotland can be divided into two lineages, with those sampled in 2008 (lineage 1) phylogenetically distinct from those present in 2010–2013 (lineage 2). In 2009, both lineages were present in the Perthshire population and it is possible that our limited sampling has not detected other examples of co-circulation. Within lineage 1, the viruses sampled in 2008 are also distinct from those sampled in 2009, while there is no obvious distinction within the sequences of lineage 2 from 2009–2013. The three viruses sequenced from Yorkshire, sampled between 31/12/2008 and 8/3/2011, represent a third distinct UK lineage.
Despite the difference in progenitor viruses in Australia and Europe the subsequent evolution of these viruses is strongly clock-like. Using a Bayesian approach and a strict molecular clock the mean evolutionary rate for the 32 European viruses was estimated to be 0.99 x 10−5 nucleotide substitutions per site, per year (subs/site/year) (95% HPD values of 0.90–1.09 x 10−5 subs/site/year), while the equivalent value for the 25 Australian viruses was 1.03 x 10−5 subs/site/year (95% HPD values = 0.86–1.21 x 10−5 subs/site/year). Very similar rates were obtained using a variety of data sets and nucleotide substitution, molecular clock and demographic models (Fig 1). In addition, a regression of root-to-tip genetic distance against year of sampling for the combined Australian and European data set revealed strong temporal structure (R2 = 0.93), with a mean evolutionary rate of 1.04 x 10−5 subs/site/year that was very close to that estimated using the Bayesian approach for the entire data set at 1.02 x 10−5 subs/site/year (95% HPD values = 0.94–1.10 x 10−5 subs/site/year) (Fig 1). The similarly of rates among viruses sampled on different continents suggests that their high evolutionary rate is largely a reflection of rapid background mutation as suggested for other pox viruses [25]. Under these evolutionary rates it is estimated that the two MYXV lineages from Perthshire shared a common ancestor between 1956 and 1963, while the lineage leading to the Yorkshire viruses originated between 1953 and 1955 (Fig 1).
Across all the UK viruses there were 162 non-synonymous mutations, 137 synonymous mutations and 26 insertion/deletion events within ORFs compared to Lu; 51 genes had no mutations and a further 23 only possessed synonymous changes (Fig 2A). A comparison with the mutations observed in the Australian isolates (Fig 2B) revealed that different genes tended to show the highest numbers of mutation in each case. Indeed, only the M017L gene exhibited frequent mutation in both data sets (Fig 2C). Overall, 23 genes contained no mutations among both the UK and Australian sequences and a further 23 had only synonymous changes (S1 Table).
As previously reported for MYXV in Australia [20, 21], single or multiple nucleotide insertions/deletions (indels) leading to the predicted disruption of ORFs were relatively common (Table 2). Disruptions of genes previously identified as having major virulence functions and leading to likely loss of function of the encoded protein occurred in M002L/R [26]; M004L/R [27, 28]; M005L/R [29, 30]; M148R [31] and M153R [32, 33]. In addition, there was loss of the M009L ORF in Perthshire lineage 1 and by two independent mutations in the Yorkshire lineage, and of the M036L ORF in Perthshire lineages 1 and 2. There was also an adjacent mutation in M036L in the early Sussex and Nottingham strains, with a possible reversal of this disruptive mutation in the Yorkshire lineage (S1 Fig). Single viruses with gene disruptions were found in all three lineages: M135R (Perthshire 1527) and M008.1L/R (Perthshire 2409) have been shown to have virulence functions [34, 35]. M009L has also been lost in most modern Australian viruses, as well as in some European isolates and in the Californian MSW strain of MYXV [20, 21, 36, 37, 24], suggesting that this gene is not essential.
In addition to indels that disrupted ORFs, there were a number of large and small indels within genes that were not disruptive (S2 Table). Moreover, there were single nucleotide indels in multiple intergenic homopolymer regions and larger deletions in some blocks of intergenic repeat sequence elements. These will not be considered further.
Temporal regulation of most MYXV genes has been predicted on the basis of conserved early, late or intermediate promoter motifs [22, 38]. However, the transcription start sites of most MYXV mRNAs have not been mapped and hence actual expression may differ from that assigned [39, 31, 40]. In the UK sequences, mutations upstream of the M000.5L/R, M001L/R, M008.1L/R, M019L, M033L, and M153R genes were located close to potential promoter sequences and could conceivably alter transcription [41, 42]. However, any effect was likely to be limited, with the possible exception of a mutation in the M153R putative promoter sequence in the Perthshire lineage 2 viruses which could conceivably decrease promoter activity. This mutation was also present in the Australian WS6 1071 virus.
To evaluate how the genetic divergence from the Lu progenitor has affected disease phenotypes in the UK viruses, groups of six laboratory rabbits were infected with representative viruses from Perthshire lineages 1 and 2, and all three Yorkshire lineage viruses, and their virulence and disease phenotypes compared to rabbits infected with the Lu progenitor virus.
The virulence grade of each isolate was estimated using the method of Fenner and Marshall (1957) [5]. These virulence assignments were necessarily inferred since rabbits were euthanized and survival times (ST) estimated rather than using death as an endpoint (Table 3). Kaplan-Meier plots show the actual ST estimates rather than the normalized values (Fig 3). The Lu strain was tested as a control and had a similar AST to previous reports [5]. Notably, the grade 1 Yorkshire 135 isolate had a significantly lower ST than all other viruses tested including Lu.
In our animal experiments the disease caused by Lu was indistinguishable from previous descriptions of Lu as the prototype European virus [5], with the exception that we did not see the copious nasal discharge, likely because of the absence of Pasteurella multocida in the upper respiratory tract of the specific-pathogen-free rabbits. Notable features of Lu compared to the infections with the recent virus isolates were extreme swelling of the eyelids and lips, large size of the primary lesion, large numbers of secondary cutaneous lesions and a precipitous clinical decline between days 10 and 12 (S3 Table; S4 Table).
A striking feature of infection with some viruses from all three recent UK lineages was acute collapse resembling septic shock with relatively mild signs of myxomatosis. This was distinct from the disease caused by Lu. Hemorrhages in multiple tissues, massive pulmonary oedema and swollen, pale or granular livers were also frequently but not universally present, although the degree of pathology may have depended on timing of euthanasia or death. Aggregates of coccoid bacteria were often present in multiple tissues but with no apparent cellular inflammatory response (Fig 4; S5 Table). These rabbits often had higher virus titres in liver and lung compared to rabbits infected with Lu (S6 Table).
Overall, disease phenotypes could be divided into: (i) a nodular cutaneous or “myxomatous” disease with prominent primary lesions at the inoculation site and secondary cutaneous lesions on ears, head, body and legs as seen with Lu, Perthshire 1527 and Yorkshire 127 viruses, or (ii) a disease that resembled the “amyxomatous” phenotype described in Europe [44, 45, 46] and characterized by a poorly defined primary lesion and no or very few secondary cutaneous lesions. This second phenotype was seen with Perthshire 1792, 2082, 2282, Yorkshire col and Yorkshire 135, while Perthshire 1537 had an intermediate phenotype (Fig 5; S4 Table; S5 Table). Acute collapse was only seen with the amyxomatous infections. Other features of myxomatosis such as swollen heads, ears, eyelids and perineum were, to some degree, common to all infections. Prolonged incubation periods described for some amyxomatous viruses [44] were not seen.
Distinctive differences were also present in the pathology of the acute collapse amyxomatous infections compared with Lu and the myxomatous phenotype (S5 Table). Bacteraemia was not a feature of the Lu infections. Although bacteria were observed in a necrotic focus in the liver of one rabbit infected with Lu, these were associated with an acute inflammatory response. The large numbers of neutrophils seen deep in cutaneous tissues and within lymph nodes in the Lu infections (Fig 5H) were absent in rabbits with the acute collapse syndrome and lymph nodes and spleens tended to be more depleted of lymphocytes in these rabbits. Late clinical signs in longer surviving or recovering rabbits were fairly typical of those described for myxomatosis caused by moderately attenuated viruses [5], with the exception that the amyxomatous viruses did not induce secondary lesions (S4 Table; S5 Table).
The prolonged duration of high virus titres in the epidermis of primary or secondary lesions or in sites such as eyelids or ears is critical for transmission by arthropod vectors [9]. In general, longitudinal biopsy samples showed that levels of virus in the primary lesions, measured by qPCR, increased over the first 10 days to > 108 copies/mg and were then reasonably stable, albeit with reduced numbers of rabbits available for biopsy at later time points (S2 Fig). However, two virus infections had consistently lower virus loads: the grade 5 Perthshire 1527 and the grade 2/3 Yorkshire 127 strain. Both viruses had the nodular myxomatous phenotype and the lower loads were probably due to cell destruction in the epidermis. Despite the limited nature of the primary lesion in the amyxomatous phenotypes (Fig 5I) they had very high levels of virus. Similar results were obtained with titres measured by plaque assay on autopsy samples (S6 Table). Titres in the Lu infected rabbits were also relatively low, likely because of the highly scabbed and degenerate nature of the lesion (S6 Table; Fig 5). Biopsies were not collected from rabbits infected with Yorkshire 135 or Lu. Taken together with the histological and gross appearance of the primary lesions, these results indicate that the tissue response to the amyxomatous viruses is entirely different to that induced by Lu, but that this is not due to reduced virus replication.
Despite the observed differences in disease phenotype and virulence, viruses within each lineage exhibit limited sequence divergence. For example, Yorkshire 127 caused the nodular cutaneous phenotype while the closely related Yorkshire 135 and Yorkshire col caused the amyxomatous phenotype (Fig 3). All three Yorkshire viruses have lost the functional domain of the M005L/R gene and have disrupted M009L and M153R genes (Table 2). There are six amino acid differences between Yorkshire 135 and Yorkshire Col and seven between Yorkshire 135 and Yorkshire 127 (S7 Table).
The Perthshire lineage 1 viruses are more complicated, as the 2008 viruses (1527, grade 5 and 1537, grade 3/4) have a disrupted M002L/R gene and Perthshire 1527 has a disrupted M135R gene; both are virulence determinants in Lu [34]. These genes are intact in the amyxomatous 2009 Perthshire 1792 virus (grade 2). As with the Yorkshire lineage, these viruses only differ at a small number of amino acid sites (S8 Table).
Both Perthshire lineage 2 viruses tested had the amyxomatous phenotype and were of grade 3 virulence. Apart from the premature termination of M008L/R in 2082, there are only four amino acid differences between these viruses (S9 Table). Phenotypically, it was difficult to differentiate these grade 3 viruses from the grade 2 Perthshire 1792 and Yorkshire Col.
Overall, these results suggest that single amino acid changes can have a major impact on disease phenotype and virulence gene disruption may be compensated by epistatic mutations or other mechanisms.
Our genome-scale evolutionary analysis reveals that multiple lineages of MYXV have circulated in UK rabbits. In particular, the single lineage of viruses from Yorkshire and the two lineages present in Perthshire clearly diverged relatively early in the epizootic and have evolved independently ever since. This separation of the English and Scottish viruses could reflect a simple biogeographic division and a lack of virus gene flow, particularly since the European rabbit flea (Spilopsyllus cuniculi) is the main arthropod vector in the UK so that virus spread depends on movement of rabbits carrying fleas [47, 48]. However, the phylogenetic separation between the two Scottish lineages is harder to explain as they were sampled within three kilometres of each other. Because these two lineages differ in the range of temporal sampling (2008–2009) and (2009–2013) it is possible that the later sampled lineage is a more recent invader into the study area and has outcompeted the previously existing lineage. Anecdotally, in 2009 this study site experienced a high mortality of rabbits due to myxomatosis, compatible with the possible invasion of a new strain into the area.
Importantly, our comparison of MYXV genome sequences from the UK and Australia confirms previous conclusions that there is no single pathway to attenuation from the progenitor viruses or from attenuation back to virulence [20]. Indeed, it is striking that there are almost no shared mutations between the viruses from the two radiations despite the large number of complete genomes now sequenced. Hence, evolutionary success in these large genome DNA viruses has clearly resulted from the exploration of multiple evolutionary pathways along which different disease phenotypes appear. Indeed, our animal trials reveal that the clinical phenotype of a number of the UK viruses showed dramatic changes compared to the progenitor Lu virus, as well as within and between the modern viral lineages.
Generalized disease seems critical for efficient virus transmission in European rabbits, with rabbits that survive infection (and therefore control virus replication) being poor transmitters [10]. In addition, resistance is manifest as control of virus replication rather than prevention of infection [49, 50, 51], so is likely to select for virus mutations that can overcome this control. The emergence of genetic resistance in the wild rabbit population likely shifted selection towards more virulent viruses (when tested in non-resistant rabbits) to maintain this nexus between virulence and transmission, in turn setting up an arms race between host and virus. As we describe here, this can lead to dramatic changes in the disease phenotype in non-resistant rabbits.
There is an implicit idea that changes in virulence will be due to mutations in genes involved in immunomodulation or host-range functions [40]. The role of many MYXV genes in virulence has been defined by single gene knock-out studies using the Lu strain or an early French derivative, the T1 strain [52]. In particular, the M005L/R and M153R genes have each been shown to have major virulence functions. Rabbits infected with knock-outs of either gene had a much lower CFR: 30% for ΔM153R and 0% for ΔM005L/R compared to 100% for Lu [32, 29]. However, all three Yorkshire viruses have mutations that are predicted to disrupt both these genes causing loss of key functional domains [33, 30] but have CFRs of nearly 100%. This suggests three possible explanations for retained virulence: (i) epistatic mutations compensating for the loss of these genes; (ii) a mechanism for suppressing reading frame disruptions; or (iii) functional activity retained by the truncated protein (potentially in a new role) [53]. Although it seems likely that unique amino acid substitutions are often responsible for alterations in virulence, the number of such amino acid changes evidently makes specific virulence determinants difficult to identify. Similarly, the Californian MSW strain of MYXV, which is found in S. bachmani in North America and is the most virulent strain of MYXV described for European rabbits [5, 54], has disrupted multiple virulence genes, suggesting that multiple epistatic mutations play a role in virulence determination [36].
As well as broad trends in virulence during the early radiation, changes were also observed in the clinical appearance of infected rabbits, with a relatively rapid evolution of a flat lesion morphology in both Australia and Europe rather than the domed SLS and Lu lesions [5, 15]. More recently, the amyxomatous phenotype in European isolates has been distinguished from the nodular type of disease by having few or no cutaneous lesions and, in some cases, apparently prolonged incubation periods [44, 55, 56]. For some Australian isolates the amyxomatous phenotype is seen in laboratory rabbits, although the same virus causes a nodular phenotype when tested in resistant wild rabbits suggesting that changes in the pathogenesis of the disease have occurred due to selection in resistant wild rabbits [57].
Combined, these data strongly suggest that the accumulation of mutations in field strains of MYXV has caused changes in the pathogenesis of myxomatosis, such that we now see a spectrum of disease types that depend on the interactions between the virus genome and the genetics of the rabbit and non-genetic (rabbit) factors such as microbial flora, parasites, and abiotic environmental factors including temperature [58]. As an example, field isolates of European amyxomatous viruses tested in specific pathogen-free laboratory rabbits caused relatively minor disease with few fatalities. However, the same viruses tested in rabbits from commercial rabbitries caused significant disease with severe bacterial bronchopneumonia as the most common cause of death [46, 59]. Different environmental conditions and vectors may therefore facilitate selection of virus strains that are more successful in particular niches. For example, in the farmed domestic rabbit populations in Europe where there has been no selection for resistance, we may expect low virulence strains predominantly transmitted by contact, strains with prolonged incubation periods [60, 61], or high virulence strains that can overcome imperfect vaccination [60, 56, 37].
With the exception of Yorkshire 127, rabbits that died or required euthanasia early in the course of the disease had very different clinical signs from those infected with Lu. Hemorrhage and acute pulmonary oedema were common together with high titres of virus in lungs and liver. In some cases, large numbers of coccoid bacteria were present in multiple tissues, but did not elicit a visible cellular inflammatory response. Lymphocyte depletion from lymph nodes and spleens was relatively common. Despite extremely high virus titres, there was very limited pathology in the epidermis and dermis of the primary inoculation site. This suggests an acute overwhelming of the rabbit immune response triggered by high viral titres in critical tissues. This outcome is also clearly distinct from the secondary gram negative bacterial infections (Pasteurella multocida, Bordetella bronchiseptica) described in the upper respiratory tract for rabbits infected with the progenitor viruses or the bacterial bronchopneumonia described with isolates from rabbit farms [59]. In our study, rabbits that did not die of acute disease developed more typical signs of myxomatosis, although upper respiratory tract occlusion and discharge was relatively mild, possibly reflecting the specific-pathogen free status of the rabbits.
Whether the difference in survival time and clinical disease between the acutely affected animals and the more chronically affected longer term survivors is related to genetic factors in the outbred rabbits or some stochastic factor early in the course of disease is not clear, but these animals clearly have a different form of the disease. Virulence, using the definitions of Fenner and Marshall (1957), essentially meant the AST. However, this raises the question of what virulence means in terms of how a strain of MYXV causes disease? Does a more virulent virus cause a different disease, or are there many pathways to death in an infected rabbit such that the phenotype seen may be due to which particular mechanism occurred in an individual rabbit. Thus, in one animal we see hemorrhage and pulmonary oedema, yet in another we see acute death without pulmonary oedema and hemorrhage, which might have developed if the animal had survived a few hours longer. It is possible that some of the longer-term survivors have a milder form of the disease at this stage and will go on to develop the more typical form of myxomatosis, and this pathway seems to predominate in attenuated viruses such as Perthshire 1527. Clearly, virulence in this case is a more nuanced concept than generally depicted in studies of its evolution.
The parallel evolution of virulence in MYXV in the Australian and British epizootics was evidently not accompanied by the acquisition of similar mutational changes. Our detailed examination of genomics and disease phenotypes of recent isolates of MYXV from the UK radiation reveals that highly virulent and highly attenuated viruses were present in the field, but that disruptions to major virulence genes were not necessarily associated with attenuation. More striking was that the disease caused by many of these viruses was clinically distinct from that caused by the progenitor Lu strain, with alterations in tissue tropism and pathogenesis in acutely affected rabbits, again demonstrating that the virus is able to explore many pathways to evolutionary success.
Sampling was performed according to field procedures approved by the Institutional Animal Care and Use Committee of The Pennsylvania State University (IACUC # 26383 and 34489). Animal experiments were conducted under protocols approved by the Institutional Animal Care and Use Committee, Pennsylvania State University (IACUC # 33615 and 42748). All animal work adhered to the guidelines laid out in the Guide for the Care and Use of Laboratory Animals. 8th ed. National Research Council of the National Academies. National Academies Press Washington DC.
The virus isolates sequenced in this study are listed in Table 1. Samples were taken from rabbits with clinical myxomatosis gathered at multiple locations on two sites, the first located in Perthshire in central-eastern Scotland, and the second in North Yorkshire, England, collected as part of other field studies [62, 63, 64, 65]. An early isolate sampled in Belfast, Northern Ireland in 1955, was also sequenced. All viruses were isolated in RK-13 cells and passaged between 1 and 3 times to prepare seed and working stocks, from which virus DNA was prepared [66]. An aliquot of virus from the DNA preparations was used for rabbit infections.
Virus genomes were sequenced on three different platforms: the Illumina HiSeq 2000 and MiSEq, and the Ion Torrent. For the HiSeq200, template viral DNA was processed using a TruSeq DNA sample preparation kit (Illumina) to produce a multiplex library for sequencing. Briefly, extracted viral genomic DNA (gDNA) was sheared with a Covaris AFA system, creating fragments of 50 to 7,000 bp. After end-repair, purification, and 3′ adenylation, bar-coded sequencing adapters were ligated, and 400- to 500-bp fragments were purified. Fragment enrichment and clean-up were performed with AMPure XP beads. Individual library components were quantitated by quantitative PCR (qPCR), normalized, and pooled into a final sequencing library consisting of eight different viral genomes (this included seven MYXV strains that were analyzed in a separate study), which was run on an Illumina HiSeq2000 to generate 100-bp paired-end reads. For the MiSeq, libraries were produced using the Nextera XT DNA kit (Illumina). Extracted DNA samples were quantified using a Qubit fluorometer and 1ng of each sample was used as input DNA. The standard workflow was followed: duel index barcoding of the tagmented DNA was done according to the low plexity requirements and 1.8x AMPure XP beads were used to purify the library DNA. Library normalization was performed using Illumina beads. Multiplexing of the final library occurred according to Illumina recommendations. Briefly, 5 μl of each of the 14 finished, bead-normalized libraries were combined into a library pool. Next, 24 μl of this mix was transferred to a new tube containing 576 μl HT1 buffer, mixed well, and placed at 96°C for 2 minutes to denature, followed by cooling on ice for at least 5 minutes. Denatured 8pM PhiX was then combined with the denatured library pool in a total volume of 600 μl and a final concentration of 5% to produce the final sequencing pool. Sequencing was performed on an Illumina MiSeq using either 2x75bp V3 or 2X250 V2 paired-end kits, yielding approximately 14.5M paired-end reads for each run. Isolates 1527 and 2282 were sequenced on the Ion Torrent. Genomic DNA was sheared and converted into libraries with the Ion Xpress Plus fragment kit (Ion Torrent) by following the manufacturer’s instructions. Briefly, 200ng of gDNA was sheared for 20 minutes followed by purification, nick repair and adapter/barcode ligation. The DNA libraries were then size selected on the E-Gel SizeSelect (Invitrogen) platform to yield insert sizes of ~200 bp. Libraries were quantitated on the Bioanalyzer (Agilent) and combined in equimolar amounts to make the final sequencing pool. This pool was sequenced on the Ion Torrent with a 316 chip and a 200 base read length target, yielding 2.6M useable reads.
Demultiplexed reads were quality trimmed using the trim.pl perl script (http://wiki.bioinformatics.ucdavis.edu/index.php/Trim.pl) and assembled with the Velvet de novo assembler iterated across a range of k-mers from 45 to 65 for each assembly [67]. Contigs were ordered into a single scaffold for each genome using the Abacas.pl script [68] and the Lu genome as reference (GenBank accession AF170726), and for each assembly the k-mer that generated the most complete coverage of the reference genome was selected for finishing and downstream analysis. The quality of each scaffold was verified by remapping the untrimmed reads to the assembly using Smalt (http://www.sanger.ac.uk/science/tools/smalt-0). One region of ambiguous assembly was amplified by PCR and sequenced using Sanger methodology to confirm the assembly. A nucleotide deletion within a homopolymer run in the M153R gene was also confirmed by Sanger sequencing. In every case, only one complete or near complete copy of the terminal inverted repeat (TIR) was assembled at either the 5’ or the 3’ end. The Belfast 1955 isolate was assembled de novo on a 100K sub-sample of the cleaned, paired-end reads using CLC Genomics (version 8) with a word and bubble size of 30 nt and 150 nt, respectively. This yielded two contigs corresponding to the core genome (~138K) and TIR (~11K). The TIR contig was duplicated and reverse complemented before manually assembling onto the core genome, and then all the cleaned, paired-end data was re-mapped back to confirm final assembly.
Genome annotation was transferred from the Lu strain to the newly sequenced MYXV genomes using the Rapid Annotation Transfer Tool [69]. EMBL flatfiles of transferred gene models were then inspected and compared to the Lu reference using the Artemis Comparison Tool [70]; incorrect models were corrected, and new gene models added where transfer had not occurred.
Nucleotide sequence accession numbers: all genome sequences generated here have been deposited in GenBank (https://www.ncbi.nlm.nih.gov/) under accession numbers KY548792-KY548813 (S10 Table).
The 22 MYXV genome sequences determined here were combined with 35 complete genomes available on GenBank, representing 25 from the Australian outbreak (including the SLS release strain) and 10 from Europe (including the Lu release strain) (S10 Table). These sequences were initially aligned in MUSCLE [71] and adjusted manually, resulting in a final sequence alignment data set of 57 sequences 163,645 bp in length. Because the sequences are highly conserved, the locations of synonymous and non-synonymous mutations in these sequences were determined manually.
An initial phylogenetic tree of these sequences was inferred using the maximum likelihood procedure available in the PhyML package [72]. This analysis utilized the HKY+Γ4 model of nucleotide substitution and NNI+SPR branch-swapping. To test for the presence of recombination we utilized the RDP, Genecov and Bootscan methods (with default settings) available within the RDP4 package [73]. No significant evidence for recombination was found.
To determine the rate of MYXV evolution we first assessed the degree of clock-like structure in the data using a regression of root-to-tip genetic distances on the ML tree inferred above against the year of virus sampling using TempEst [74]. As this analysis revealed strong temporal structure (see Results), we next inferred the rates and dates of viral evolution using the Bayesian Markov chain Monte Carlo (MCMC) approach available in the BEAST package [75]. For this analysis we used a range of nucleotide substitution (HKY+Γ4, GTR+Γ4), molecular clock (strict, relaxed uncorrelated lognormal) and demographic (constant, Bayesian skyride) models. As these gave strongly overlapping results we based our analysis on the simplest model: HKY+Γ4, strict clock, constant population size (Fig 1). All analyses were run twice and for sufficient time (100 million generations) to ensure that convergence was achieved, with statistical uncertainly manifest in values of the 95% highest posterior distribution (HPD). The posterior distribution of trees from the HKY+Γ4, strict clock, constant population size run was also used to infer a maximum clade credibility (MCC) tree (Fig 1). The degree of support of individual nodes is depicted as posterior probability values.
New Zealand White male laboratory rabbits (Oryctolagus cuniculus) of four months of age were purchased from Harlan Laboratories (Oakwood facility). Rabbits were specific-pathogen-free for Pasteurella multocida and Bordetella bronchiseptica. Animals were housed in individual cages on a 12h light regime, fed 125 g of standard pellets per day and allowed 10 days to acclimate in the facility prior to infection.
Groups of six rabbits were inoculated with 100 pfu of virus intradermally in the rump and monitored closely over the course of the infection. Daily clinical examination included: rectal temperature, body weight, primary lesion size and shape at the inoculation site, secondary lesion size and distribution, plus semi-quantitative scoring on a 0 to 3 scale for demeanour, eyelid swelling, ear swelling, anogenital swelling, scrotal oedema, blepharoconjunctivitis, nasal discharge and respiratory difficulty. Food and water intake were recorded and fecal and urinary output monitored by inspection of collecting trays under the cages. Rabbits were euthanized based on the degree of clinical severity using respiratory difficulty, depression, inanition, reluctance to move, weakness on handling, weight loss and failure to eat or drink as indicators; any rabbit exhibiting pain or with a subnormal temperature was immediately euthanized.
To monitor virus replication at the primary inoculation site, 1 mm diameter dermal punch biopsies were collected: in each group, three rabbits were sampled at day 5 post-infection and three at day 7; thereafter, each surviving rabbit was sampled at 5 day intervals. DNA was prepared using the DNeasy kit (Qiagen).
Rabbits were autopsied as soon as possible after death and bodies refrigerated if autopsy was delayed. Samples of the primary lesion and other tissues were collected for virus titration and histology but only from euthanized rabbits or rabbits that died within 1–2 hours prior to autopsy. Blood samples were collected from the marginal ear vein at days 0 and 10, or following euthanasia, by cardiac puncture. Hematology was performed by the Centralised Biological Laboratory Facilities, the Pennsylvania State University.
Because of the unusual virulence of the Yorkshire 135 virus, we tested whether there was any adventitious agent in the virus preparation by challenging immune rabbits with Yorkshire 135. The only reaction was a swelling at the inoculation site, which resolved by day 6. This is typical of what is seen when immune rabbits are challenged. While this does not completely exclude an adventitious agent that was only pathologic in the context of a highly immunosuppressive MYXV infection, it strongly supports the hypothesis that the peracute disease seen with Yorkshire 135 was indeed due to MYXV.
To enable comparison with previous studies of MYXV, survival times (ST) from inoculation to death were estimated for rabbits that were euthanized as follows: (i) moribund rabbits were assigned the time of euthanasia as the ST; (ii) rabbits that were not expected to survive the next 24 hours were assigned an additional ST of +12 hours; and (iii) rabbits euthanized for humanitarian reasons were assigned a ST of +48 hours. Animals found dead were assigned a ST half-way between the time of last observation and finding the body. Average survival times (AST) for each group were calculated from individual ST normalized using the procedure of Fenner and Marshall (1957) [5] as log10(ST-8) and then back-transformed; a survival time of 60 days was assigned to rabbits that recovered or were alive at the end of the trials and considered likely to recover. If more than two rabbits survived, the virulence grade was assigned based on the CFR and clinical severity. Virulence grades were based on Fenner and Marshall (1957) [5] as modified by Fenner and Woodroofe (1965) [43] (Table 4). Data were also analysed using Kaplan-Meier survival plots (using actual inferred survival times rather than the normalized survival times) and tested for statistical significance by log rank test implemented in SigmaPlot.
Quantitative PCR (qPCR) was performed on an ABI 7500-fast machine, using the Quantifast Sybr green kit (Qiagen), by amplification of a 126 bp fragment (nt 584–710) from the M080R gene from DNA extracted from primary lesion biopsies. This was quantified on a standard curve using a linearized control plasmid containing a 642 bp region of the M080R gene (nt 241–883). None of the UK viruses have mutations in this sequence. Virus titres were expressed as genome copy number/mg tissue. The qPCR primers used were: M080 qPCR Forward: 5' TATCAAACAACCTCCGCATACC 3' (M080R 584–605) and M080 qPCR Reverse: 5' CTCCCATAACGCTTCCGAC 3' (M080R 710–692)
Samples of the primary lesion, lung, liver, spleen, and right popliteal lymph node were collected at autopsy from euthanized rabbits. Tissues were homogenized by Tissuelyser (Qiagen). Virus was titrated on RK-13 cell monolayers as previously described [49] with titres expressed as pfu/g of tissue.
|
10.1371/journal.pcbi.1004507 | Exposing Hidden Alternative Backbone Conformations in X-ray Crystallography Using qFit | Proteins must move between different conformations of their native ensemble to perform their functions. Crystal structures obtained from high-resolution X-ray diffraction data reflect this heterogeneity as a spatial and temporal conformational average. Although movement between natively populated alternative conformations can be critical for characterizing molecular mechanisms, it is challenging to identify these conformations within electron density maps. Alternative side chain conformations are generally well separated into distinct rotameric conformations, but alternative backbone conformations can overlap at several atomic positions. Our model building program qFit uses mixed integer quadratic programming (MIQP) to evaluate an extremely large number of combinations of sidechain conformers and backbone fragments to locally explain the electron density. Here, we describe two major modeling enhancements to qFit: peptide flips and alternative glycine conformations. We find that peptide flips fall into four stereotypical clusters and are enriched in glycine residues at the n+1 position. The potential for insights uncovered by new peptide flips and glycine conformations is exemplified by HIV protease, where different inhibitors are associated with peptide flips in the “flap” regions adjacent to the inhibitor binding site. Our results paint a picture of peptide flips as conformational switches, often enabled by glycine flexibility, that result in dramatic local rearrangements. Our results furthermore demonstrate the power of large-scale computational analysis to provide new insights into conformational heterogeneity. Overall, improved modeling of backbone heterogeneity with high-resolution X-ray data will connect dynamics to the structure-function relationship and help drive new design strategies for inhibitors of biomedically important systems.
| Describing the multiple conformations of proteins is important for understanding the relationship between molecular flexibility and function. However, most methods for interpreting data from X-ray crystallography focus on building a single structure of the protein, which limits the potential for biological insights. Here we introduce an improved algorithm for using crystallographic data to model these multiple conformations that addresses two previously overlooked types of protein backbone flexibility: peptide flips and glycine movements. The method successfully models known examples of these types of multiple conformations, and also identifies new cases that were previously unrecognized but are well supported by the experimental data. For example, we discover glycine-driven peptide flips in the inhibitor-gating “flaps” of the drug target HIV protease that were not modeled in the original structures. Automatically modeling “hidden” multiple conformations of proteins using our algorithm may help drive biomedically relevant insights in structural biology pertaining to, e.g., drug discovery for HIV–1 protease and other therapeutic targets.
| Even well-folded globular proteins exhibit significant flexibility in their native state [1]. However, despite advances in nuclear magnetic resonance dynamics experiments and computational simulations, accurately characterizing the nature and extent of biomolecular flexibility remains a formidable challenge [2]. While traditionally X-ray crystallography is associated with characterizing the ground state of a biomolecule, the ensemble nature of diffraction experiments means that precise details of alternative conformations can be accessed when the electron density maps are of sufficient quality and resolution [3]. These maps represent spatiotemporal averaged electron density from conformational heterogeneity across the millions of unit cells within a crystal [4, 5].
Computational methods have made strides toward uncovering and modeling conformational heterogeneity in protein structures from crystallographic data [3]. However, there is currently no automated approach to recognize the features of extensive backbone flexibility in electron density maps, model the constituent alternative conformations, and validate that the incorporation of heterogeneity improves the model. B-factors theoretically model harmonic displacements from the mean position of each atom, but in practice are often convolved with occupancies of discrete alternative positions when multiple backbone conformations partially overlap [5]. Statistical analyses of electron density using Ringer has revealed evidence for a surprising number of “hidden” alternative conformations in electron density maps [6, 7]. The phenix.ensemble_refinement method [8] uses electron density to bias molecular dynamics simulations, then assembles snapshots from this trajectory into a multi-copy ensemble model. However, energy barriers of the simulation may prevent sampling of well separated backbone conformations. Accurately modeling protein conformational heterogeneity, in particular when the mainchain adopts distinct conformations for one or a number of contiguous residues, remains a difficult task. The spatial overlap of electron density of multiple conformations and the relatively similar profiles of branching mainchain and sidechains blur structural features that can guide the human eye to reduce the large number of possible interpretations [9].
We have previously developed qFit [10], a method for automatically disentangling and modeling alternative conformations and their associated occupancies, which are represented by the variable q (for “occupancy”) in standard structure factor equations. The qFit algorithm examines a vast number of alternative interpretations of the electron density map simultaneously. To propitiously explore a high-dimensional search space, conformational sampling is guided by the anisotropy of electron density at the Cβ atom position, the nexus of backbone and sidechain in polypeptides [11]. For each slightly shifted Cβ atom position, qFit samples sidechain conformations with a rotamer library [12] and uses inverse kinematics to maintain backbone closure [9]. Finally, it selects a set of one to four conformations for each residue that, collectively, optimally explain the local electron density in real space.
However, the anisotropy of the Cβ atom limits the exploration radius of qFit to model backbone conformational heterogeneity. While protein backbone motions are often associated with large-amplitude conformational flexibility of surface loop regions, subtle motions can have important ripple effects in closely packed areas via sidechain-backbone coupling. For example, fast (ps-ns) backbone NH and sidechain methyl order parameters from spin relaxation experiments are highly correlated with each other in flexible regions [13], suggesting that mainchain and sidechain motions collectively sample conformational substates. For example, a backbone backrub motion [14] repositions the Cα-Cβ bond vector in a plane perpendicular to the chain direction, enabling the sidechain to access alternative, often sparsely populated rotamers that otherwise would be energetically unfavorable. We previously linked coupled transitions between alternative sidechain conformations, like “falling dominos”, to enzymatic turnover and allostery [15, 16].
Additionally, qFit cannot model discrete conformational substates such as peptide flips, which are >90° rotations of a peptide group while minimally perturbing the flanking residues. Some structure validation methods highlight incorrect peptide orientations [17] and even automate subsequent model rebuilding [18]. However, rebuilding fits a correct, unique conformation rather than multiple well-populated alternative peptide conformations. Peptide flips can have important functional roles in proteins. For example, flavodoxin undergoes peptide rotations between functional states as part of the catalytic cycle [19], and peptide flips that convert β-sheet to α-sheet have been linked to amyloid formation [20]. Furthermore, high-resolution crystal structures have shown that alternative conformations related by a peptide flip may be populated in the same crystal, although not as commonly as backrubs [14].
Modeling alternative conformations of glycine residues, which lack a Cβ atom, is also a current limitation of qFit. The lack of a Cβ atom allows glycine residues to access otherwise forbidden regions of conformational space [11] and thereby fill special structural roles such as capping helix C-termini [21]. In addition, the flexibility of glycines may contribute directly to function at flexible inter-domain linkers or conformationally dynamic enzyme active sites [22]. Automatically modeling such cases as alternative conformations with qFit paves the way toward understanding their contributions to protein function. Increasingly, new experiments are being proposed which, combined with computational analysis, can extract the spatiotemporal ensemble from electron density maps [15, 23, 24].
Adding the capability to model peptide flips and alternative conformations for glycines will increase our power to uncover conformational heterogeneity. While the number of sampled conformations for glycines is modest owing to a missing side-chain, including peptide flips for all amino acids adds significant computational complexity to the qFit algorithm. A powerful quadratic programming algorithm lies at the core of qFit and is necessary to determine non-zero occupancies for up to four conformations from among hundreds or even thousands of candidate conformations for each residue. Even for modest sample sizes, around 500, the number of combinations of candidate conformations is enormous, exceeding 109. As more backbone motion is incorporated into qFit, the computational complexity increases, demanding a parallelized approach to refinement on a residue by residue basis. Although this moves rebuilding away from a single node towards a larger compute cluster, the combination of data-driven sampling and selection has enabled qFit to automatically build multiconformer models that have illuminated intramolecular networks of coupled conformational substates [16] and the effects of cryocooling crystals [25, 26]. Similar hybrid approaches using robotics sampling and selection based on experimental NMR data are also being extended to nucleotide systems such as the excited state of HIV–1 TAR RNA [27].
Here we introduce qFit 2.0, an updated version of the qFit algorithm with new capabilities for modeling near-native backbone conformational heterogeneity in crystal structures. We first describe the quadratic programming procedure that allows selection of a small set of conformations per residue that collectively account for the local electron density, and discuss its extension to fitting backbone atoms in addition to sidechain atoms. We then describe new conformational sampling features of qFit 2.0, in particular glycine shifts and peptide flips. Finally, we validate the updated algorithm with both synthetic and experimental X-ray data. qFit 2.0 is freely available by webserver and source code is available for download at https://simtk.org/home/qfit.
To automatically identify alternative backbone conformations, including peptide flips, we augmented the sample-and-select protocol in qFit (see Fig 1 and Methods). Previously, conformations were sampled based on anisotropy of the Cβ atom and were selected based on the fit between observed and calculated electron density for the sidechain (Cβ atom and beyond) only. Alternative conformations for mainchain atoms were ultimately included in the multiconformer model only because they accommodated the best sidechain fits. In qFit 2.0, we now select conformations based on the fit between observed and calculated electron density for the sidechain atoms and also the backbone O atom. The O atom is an excellent yardstick for identifying backbone conformational heterogeneity for two reasons. First, it is furthest from the Cα-Cα axis so its density profile is somewhat isolated and is displaced most by rotations around that axis [14]. Second, it has more electrons than other backbone heavy atoms, so is most evident in electron density maps. This change allows us to select peptide flips outside of α-helices and β-sheets, where flips are prevented by steric and hydrogen-bonding constraints, then directly select flipped conformations. This procedure is effective because the large movement of the backbone O during a peptide flip leaves a major signature in the electron density.
Incorporating the backbone O atom also enhances the detection of less discrete backbone conformational changes. In particular, we now sample alternative glycine conformations based on anisotropy of the electron density for the O atom, by analogy to the Cβ-driven sampling for all other amino acids. This results in alternative glycine conformations that are dictated by their own local electron density. After sampling, we select combinations of conformers from a pool of candidates based on both sidechain and backbone O atoms for all amino acids, including glycines. This addition results in greater potential to discover alternative conformations throughout the protein and include additional conformational heterogeneity in the final multiconformer model.
The nullspace inverse kinematics procedure of qFit [9] naturally encodes backrub [14], crankshaft [28, 29], and shear [30, 31] motions (S1 Fig) where they are dictated by the anisotropy of the electron density for the Cβ atom. However, this anisotropy cannot identify more discrete substates of the backbone, such as peptide flips. Peptide flips are large, ~180° rotations of a peptide plane in protein backbone with minimal disturbance of adjacent peptide conformations. Enumerating many peptide flip candidate conformations with the nullspace inverse kinematics procedure would quickly lead to prohibitively large sample sizes. We therefore examined common geometries of discrete peptide flips to expedite sampling of discrete backbone substates in qFit 2.0.
Steric interactions prevent arbitrary rotations of the peptide plane, much like sidechains adopt preferred rotamer conformations. To identify plausible geometries for peptides relative to a single input peptide, we examined cases where the peptide rotates by 90–180° around the Cα-Cα axis. We identified 147 peptide flips modeled as alternative conformations in high-quality structures. After filtering this set of peptide flips with structure validation criteria and reserving some examples for a test set, we retained 79 examples that clustered around four geometries (S1 Table, S1 Data). We observed that peptide flips often included rotation and translation within the peptide plane such that the first Cα moves “below” the Cα-Cα axis and the second Cα moves “above” it (from the view in Fig 2A and 2C). These in-plane movements justify sampling geometries found in natural peptide flips in qFit 2.0 rather than, e.g., simply rotating the peptide 180° around the Cα-Cα axis. The first two clusters, “simple down” (Fig 2A and 2C, blue) and “tweaked down” (Fig 2A and 2C, red), feature a very nearly 180° rotation around the Cα-Cα axis, but with different in-plane adjustments. By contrast, the second two clusters, “left” (Fig 2B and 2D, green) and “right” (Fig 2B and 2D, brown), feature rotations closer to 120°, but in opposite directions. Our dataset here is sufficient to propose plausible, well-validated peptide flip geometries for sampling in qFit 2.0, and suggests that the four clusters could also be used to inspire moves in protein design.
We found that the two “down” clusters were more common in tight turns between β-strands: 41–50% of flips in these clusters were found in turns, as compared to 0–14% for the other two flip clusters (with a conservative definition of a turn; see Methods) (Table 1). The flip is nearly always associated with a transition between Type I/I’ and II/II’ turns. The “left”/”right” clusters were dispersed among many irregular structural contexts, but not α-helices or β-sheets. Across the four clusters, the first residue of the peptide was a glycine 7.5% of the time, in line with the general abundance of glycines in proteins (7–8%). However, the second residue of the peptide was a glycine significantly more frequently (50%, p < 10−22). This was true for the “left”/”right” clusters (21%, p < 0.05) and especially the two “down” clusters (Fig 2C) (64%, p < 10−24). This may be in part because a glycine as the second residue of a peptide can lower the flip transition energy [32]. These results generally agree with reports of flip-like conformational differences between the same tight turn in separate homologous structures [33].
To test these advances, we first explored synthetic datasets spanning resolutions from 0.9 to 2.0 Å with increasing B-factors as a function of resolution and Gaussian noise added to structure factors (see Methods). We used the Top8000 peptide flip geometry cluster centroids, with the alternative conformations at 70/30 occupancies for the “tweaked down” cluster and 50/50 occupancies for the other three clusters. Because qFit uses these geometries to sample peptide flips, we expected it would be able to successfully identify each flipped alternative conformation starting from the primary (labeled “A”) conformation at high-to-medium simulated resolution, but less well at lower simulated resolution. Indeed, qFit 2.0 successfully finds the flipped conformations for most peptide flip geometry clusters across resolutions with a 92% success rate overall; this rate drops only slightly with resolution from 0.9 to 2.0 Å (Fig 3). Since we rebuilt the entire protein chain, we also assessed the performance on other residues. By contrast to the true positive peptide flip results, the peptide flip and rotamer false positive rates remain quite low across clusters and resolutions (Fig 3). These results indicate that qFit 2.0 is effective at identifying peptide flip alternative conformations across a wide range of crystallographic resolutions without introducing spurious conformations.
Although tests with synthetic datasets offer insight into resolution dependence, a more direct test of the usefulness of qFit 2.0 involves crystal structures with real data. We combined structures left out of the training set from the Top8000 peptide flip examples with a few more manually curated examples for a total of 15 test cases (Table 2). When comparing qFit 2.0 models to rerefined original structures, Rfree is better for 7/15 cases and Rwork is better for 8/15 cases (S2 Fig). However, after rerefinement with automated removal and addition of water molecules to allow the ordered solvent to respond to the new protein alternative conformations modeled by qFit (see Methods), Rfree is better for the qFit 2.0 model for 10/15 cases and Rwork is better for 13/15 cases (Fig 4). The differences generally are small: the average ΔRfree is ~0.1%. Overall, these results suggest that qFit 2.0 models explain experimental crystallographic data as well as or better than traditional refinement protocols at a global structural level.
While global metrics are important, a major focus of the current work is correctly identifying local alternative backbone conformations. To explore this aspect, we compared results from qFit 2.0 to those from qFit 1.0 and original deposited structures for our test set (Table 2). qFit 2.0 successfully models both flipped conformations in 14/18 (78%) cases. For example, Val539-Gly540 in the Kelch domain of human KLHL7 is modeled with two alternative conformations related by a peptide flip (1.63 Å, PDB ID 3ii7) (Fig 5A). qFit 1.0 fails to discover the flip, resulting in significant difference electron density peaks (Fig 5B). By contrast, qFit 2.0 beautifully recovers both alternative conformations (Fig 5C). In another example, Asn42-Gly43 in carbohydrate binding domain 36 at high resolution (0.8 Å, PDB 1w0n) adopts flipped peptide conformations—yet MolProbity flags geometry errors in the deposited structure that indicate it re-converges too quickly, with alternative conformations for only the Asn42 and not also Gly43 (Fig 5D). qFit 1.0 fails to capture the flip (Fig 5E). However, qFit 2.0 not only identifies both peptide flip conformations for Asn42, but also includes split conformations for Gly43, thereby repairing the covalent backbone geometry (Fig 5F). In both cases, the peptide flip and glycine sampling enhancements in qFit 2.0 combine to model discrete backbone heterogeneity as accurately as or even better than the original structure.
In addition to retrospective positive-control tests, we also looked prospectively for “hidden” peptide flip alternative conformations that are unmodeled in existing structures. One such example is Met519-Thr520 in RNA binding protein 39. In chain A of the room-temperature structure (PDB ID 4j5o), the mFo-DFc difference electron density map around this peptide has significant positive and negative peaks, indicating it is mismodeled as a single conformation (Fig 6A). Other instances of this peptide—including in chain B of the room-temperature structure and both chains of the cryogenic structure—feature conformational diversity, much of which may be related to crystal contacts; however, these conformations fail to account for the room-temperature chain A mFo-DFc peaks (Fig 6B). However, using the room-temperature data, qFit 2.0 identifies a peptide flip in this region, which repositions Met519 and flattens the local difference density (Fig 6C). By contrast, it does not identify a peptide flip for this region in either chain using the cryogenic data, which is in accord with previous reports that cryocooling crystals can conceal or otherwise perturb conformational heterogeneity that is present at room temperature [25, 26].
In addition to selection of conformers based on fit to density for the backbone O atom for all amino acids, qFit 2.0 also adds sampling based on this atom for glycine, enabling density-driven backbone sampling for the most flexible amino acid. This facilitates modeling peptide flips in which one of the constituent residues is a glycine, as seen in the examples above (Fig 5)—but also opens the door to modeling less discrete glycine flexibility. For the 489 glycines across the 15 datasets in the test set (Table 2), qFit 1.0 cannot model more than a single conformation, but qFit 2.0 models alternative conformations for 365/489 (75%) of glycines. The Cα displacements average 0.28 Å and range from <0.01 Å up to 1.70 Å. Only 4 (4%) of these glycines were modeled with alternative conformations in the original PDB structures. These results show that the direct sampling and selection based on electron density for glycine backbone atoms in qFit 2.0 successfully identify conformational heterogeneity that was formerly unrecognized. For example, a small, glycine-rich loop in PDB ID 3ie5 is modeled with a single conformation in the deposited structure and qFit 1.0 model (Fig 7A). By contrast, qFit 2.0 recognizes the anisotropy of the electron density for each of the three glycine O atoms in the loop, so models them with alternative conformations that collectively shift the entire mini-loop region (Fig 7B).
Selecting conformers based on fit to density for the backbone O atom helps find alternative conformations not only for glycines, but also more generally for other amino acids. In many cases, this additional data-driven aspect to conformer selection drives the identification of subtle, non-discrete backbone motions that are coupled to larger, discrete sidechain changes. Indeed, for the 15 proteins in Table 2, qFit 2.0 shifts the Cα more than does qFit 1.0 for 52% of residues, but the reverse is true for only 20% of residues (the remaining residues are not moved by either version) (Fig 8A). Furthermore, for 63% of the residues for which qFit 2.0 finds a new sidechain rotamer that qFit 1.0 does not, qFit 2.0 also moves the Cα more (Fig 8B). These results imply that the backbone sampling by qFit 2.0 not only increases backbone heterogeneity in and of itself, but also drives discovery of sidechain conformational heterogeneity. As one specific example, Thr157 in cyclophilin A is modeled with alternative backbone and rotamer conformations in the deposited structure (Fig 8A). qFit 1.0 fails to find the alternative rotamer because it maintains a single backbone conformation (Fig 8B), but, driven by carbonyl O anisotropy, qFit 2.0 identifies the alternative backbone conformations, allowing it to discover the second rotamer (Fig 8C).
We also observed hidden peptide flips for the Ile50-Gly51 tight turn in the “flap” region of HIV–1 protease. HIV–1 protease is a homodimer, with residue numbers often denoted by 1–99 and 1’-99’. The flap region consisting of residues 46–56 is an antiparallel β-sheet and tight turn at the interface of the dimer (Fig 9A). In most of the hundreds of crystal structures of HIV–1 protease, the two tight turns (Leu50-Gly51 and Leu50’-Gly51’) adopt an asymmetric conformation, with one flap in a single type I conformation and the other in a single type II conformation. However, NMR relaxation data suggest that these flips can undergo chemical exchange on a slow (~10 μs) timescale in solution [35]. Mutational data also linked collective conformational exchanges of these flips to catalytic rates [36]. In line with these solution studies, we noticed that for many HIV–1 protease crystal structures, the electron density maps actually reveal strong evidence for alternative conformations related by dual peptide flips. For example, in one high-resolution inhibitor-bound structure (PDB ID 3qih), the Leu50-Gly51 and Leu50’-Gly51’ flaps are modeled with single asymmetric conformations, but strong positive mFo-DFc electron density coincides with potentially flipped states (Fig 9B). Strikingly, qFit 2.0 automatically identifies dual “flap flips”, suggesting the flaps actually populate two different asymmetric states (green vs. purple in Fig 9C) in this particular inhibitor complex. More generally, this result suggests that these inhibitor-gating flaps in HIV–1 protease sample multiple conformations more often than previously recognized across many inhibitor complexes, which may motivate further investigation of the effects that protein and inhibitor flexibility have on binding affinity, efficiency of catalytic inhibition, and arisal of drug resistance in this biomedically important target.
The ruggedness of protein energy landscapes leads to conformational heterogeneity even in folded globular proteins. Evidence for these alternative conformations is remarkably prevalent in high-resolution (<2 Å) crystallographic electron density maps [6]. However, because these alternative conformations are difficult and/or time-consuming to model manually using existing graphics and refinement tools, they are underrepresented in the PDB [6]. qFit is a computational approach to overcoming these problems, by automatically identifying “hidden” alternative conformations and using quadratic programming to select a parsimonious subset that collectively best explains the diffraction data. Here we have demonstrated a new version of this algorithm, called qFit 2.0, with several enhancements to handling flexible backbone—most notably, automated detection of discrete peptide flips and explicit fitting of backbone atoms for glycines.
qFit has previously captured different types of backbone motion that can occur in secondary structure. For example, it correctly identifies the backrub motion [14] that helps Ser99 transition between sidechain rotamers in the active-site β-sheet network of CypA [15, 16], and also identifies a previously hidden α-helix winding/unwinding or “shear” motion [14, 30] (S1 Fig). However, qFit 2.0 can now model larger backbone motions in which the backbone change itself is discrete, instead of inherently continuous but coupled to discrete sidechain rotamer changes. Specifically, it models peptide flips, which occur outside of helices and sheets and involve discrete jumps over a larger energetic barrier.
Peptide flips have important implications for understanding protein function. For example, our results for HIV–1 protease (Fig 9) strongly suggest that conformational heterogeneity, in particular peptide flips, may play underappreciated roles in protein-inhibitor complexes. Previously, molecular dynamics simulations identified a large-scale “curling” motion of these flaps that is maintained by drug-resistance mutations and therefore seems important for substrate access [37]. Although this motion is more dramatic than the peptide flaps at the tips of the flaps that we observe, it underlines that flap flexibility—potentially across multiple length scales—is central to protease function and viral propagation. The peptide flip acts as a key conformational switch between type I/II turns, rearranging its environment beyond its immediate sequence neighbors and enabling alternative sidechain conformations with implications for function. However, the large number of unmodeled turns in HIV protease structures illustrates the challenge of distinguishing alternative conformations in electron density maps, even at high resolution. As an additional example which unfortunately lacks deposited structure factors, the active-site Gly57-Asp58 peptide in C. beijerinckii flavodoxin adopts distinct peptide flip states in concert with the oxidation state of the FMN prosthetic group [19]. The N137A mutation removes artificial lattice contacts that otherwise influence the conformation of the Gly57-Asp58 peptide, which results in a mixture of these peptide conformations simultaneously populated in the crystal; this suggests these multiple flip states may also coexist in solution [19].
Beyond the specific improvements to peptide flips, qFit 2.0 now fits conformations for each residue based on both sidechain (beyond Cβ) and backbone (carbonyl O) atoms. Although we originally envisioned this change for modeling glycines, we observed that it results in dramatically more extensive backbone conformational heterogeneity across the protein (Fig 8). R-factors are similar or better (Fig 4), suggesting the new models with more heterogeneity are at least as good an explanation of the experimental data. Notably, these new backbone shifts drive discovery of many more alternative sidechain rotamers (Fig 8). Our results suggest that sidechain and backbone degrees of freedom in proteins are tightly coupled, in agreement with previous reports that even subtle backbone motions can facilitate rotamer changes [14], open up breathing room for natural mutations [38], and expand accessible sequence space in computational protein design [31, 39].
Future work will investigate an armamentarium of methods for modeling larger backbone conformational change in qFit, including helix shear motions [30], adjustments of entire α-helices [40, 41], correlated β-sheet flexing [28], automated loop building algorithms such as Xpleo [9], and pre-knowledge of conformational differences between homologous structures. While these future steps will move us closer to capturing the full hierarchy of protein conformational substates [42], they will also dramatically increase the computational cost of automated multiconformer model building. Many aspects of qFit are parallelizable; however, the total computational cost for reproducing the data in this manuscript is approximately 105 CPU hours. As cloud-computing capabilities of 108 CPU hours can now be leveraged for pure simulation data [43], we envision that marshalling similar computational capabilities will become increasingly important for analysis of experimental X-ray data. Such data-driven computational approaches to studying the dynamic relationship between protein structure and function will be especially powerful when applied to series of datasets in which the protein is subjected to perturbations that modulate conformational distributions, such as ligand binding or temperature change [23].
To define possible relative geometries between flipped peptide conformations, we searched for trustworthy peptide flips modeled as alternative conformations in the Top8000 database. This database contains ~8000 (7957) quality-filtered protein chains from high-resolution crystal structures, each with resolution < 2 Å, MolProbity score [34] < 2, nearly ideal covalent geometry, and <70% sequence identity to any other chain in the database [44]. We searched the Top8000 for peptides with carbonyl C-O bonds pointed away from each other (O-O distance > C-C distance + 1 Å) and rotated by at least 90°, and for which both flanking Cα atoms reconverged to < 1.5 Å. Although peptide rotations of < 90° also occur, they occur more often in irregular loop regions, have less well-converged backbone for flanking residues, and are generally more diverse and difficult to simply categorize. By contrast, in this study we investigate the class of localized peptide rotations with well-converged backbone for both flanking residues. These are either very small rotations, or large flips with a rotation nearer to 180° —the latter being the focus here. To identify test cases for qFit 2.0, we curated the resulting dataset by removing examples with more than two alternative peptide conformations; a cis rather than trans conformation for either state; or obvious errors based on steric clashes, strained covalent geometry, or torsional outliers from MolProbity [34]. This resulted in 104 examples, from which we kept a randomly selected 79 for a geometry training set (S1 Table). We combined a subset of the remaining 25 peptide flips with a few other known examples for a test set of 18 examples (Table 1). The resolution range is 0.92–1.95 Å for the training set and 0.80–1.85 Å for the test set.
Next we characterized the geometry of peptide flips by clustering the coordinates of the flipped alternative conformation (labeled “B”) in the training set after superimposing onto a reference peptide. We used the k-means algorithm with RMSD between the five heavy atoms of the peptide backbone (Cα1, C1, O1, N2, and Cα2) for different values of k. We selected k = 4 because we observed cluster centroids with approximately 180°, +120°, and -120° rotations and for k > 4 no other significantly different rotations were identified. Notably, all four cluster centroids featured translations of the flanking Cα atoms of >0.2 Å, and as much as >0.9 Å for one cluster (“tweaked down”, red in Fig 2). The transformation matrices relating the flipped peptide cluster centroids to the reference peptide were used in qFit 2.0 to sample plausible alternative conformations, with subsequent refinement adjusting the atomic positions away from the centroid geometry.
We defined tight turns as having a mainchain-mainchain hydrogen bond between i–1 carbonyl C = O and i+2 amide N-H that was detectable by the program Probe [45]. This definition is somewhat conservative; several more examples also were visually similar to tight turns. Enrichment of glycines at the two positions involved in a peptide flip was assessed for different peptide flip clusters within the training set relative to a large set of 337 randomly selected structures containing 6,092 total glycines out of 78,094 total amino acid residues. The statistical significance of this enrichment was assessed using a one-tailed Fisher’s exact test based on the hypergeometric distribution [46].
Hydrogens were placed at nuclear positions for Label in qFit 1.0 and at electron-cloud positions for Label in qFit 2.0. Correspondingly, for Label in qFit 2.0, hydrogen van der Waals radii were taken from the new values in Reduce [48], which are intended to match those used in PHENIX. Hydrogens were absent for all other steps in qFit, including the final refinement step; however, the user is encouraged to add hydrogens to the final qFit model for their protein of interest and proceed to other analyses. Future work will update programs for downstream analysis of qFit models such as CONTACT [16] to also use electron-cloud instead of nuclear hydrogen positions.
To generate synthetic datasets for testing qFit, we used the protein chains containing the four peptide flip cluster centroids (3mcw B 101–102, 2ior A 159–160, 2g1u A 51–52, 3g6k F 172–173). We first used phenix.pdbtools to convert any anisotropic B-factors to isotropic, added 10 Å2 to each B-factor per Å of resolution worse than the original structure’s resolution to roughly simulate the general rise of B-factors with resolution, and placed the chain in a P1 box that comfortably encompassed it. Next we used phenix.fmodel to calculate structure factors (with the “k_sol = 0.4” and “b_sol = 45” bulk solvent parameters, and also generating 5% R-free flags) and added 10% noise in complex space with the sftools utility in CCP4 [47]. This process was repeated for every simulated resolution from 0.9 to 2.0 Å with a 0.1 Å step size.
qFit uses an input parameter (MC_AMPL) to scale the magnitude of movements of the Cβ (or O for glycines) along the directions dictated by its thermal ellipsoid. As in previous work [10, 16, 26], we explored multiple values for this parameter: 0.1, 0.2, and 0.3. For evaluating results such as true vs. false positive peptide flips and rotamers here, we considered all three resulting qFit models for each dataset. This is sensible because an end user of qFit 2.0 will likely reproduce this same protocol (with a few MC_AMPL values) and thus have a choice of models to use for developing insights into conformational heterogeneity and its connection to function. For other analyses, we used the minimum-Rfree qFit model model unless otherwise noted.
To compare R-factors between the deposited models and qFit 2.0, we finalized both models with phenix.refine for 10 macro-cycles using the same parameters, including the “ordered_solvent = true” flag. The resulting R-factors for qFit 2.0 models are similar or slightly better (Fig 4).
PHENIX version 1.9–1692 (the most recent official release) [49] was used for all steps of both qFit 1.0 and 2.0. Coordinates and structures factors were obtained from the Protein Data Bank [50]. qFit uses the following libraries: IBM’s ILOG CPLEX solver for QP and MIQP, which is available free of charge for academic use, and LoopTK for inverse kinematics calculations [51]. qFit is implemented in parallel; it is capable of sampling and evaluating conformations for each residue as an independent job on a Linux cluster. We have implemented job management for qFit on both Oracle/Sun Grid Engine and LSF Platform.
|
10.1371/journal.pgen.1007783 | Identification of Elg1 interaction partners and effects on post-replication chromatin re-formation | Elg1, the major subunit of a Replication Factor C-like complex, is critical to ensure genomic stability during DNA replication, and is implicated in controlling chromatin structure. We investigated the consequences of Elg1 loss for the dynamics of chromatin re-formation following DNA replication. Measurement of Okazaki fragment length and the micrococcal nuclease sensitivity of newly replicated DNA revealed a defect in nucleosome organization in the absence of Elg1. Using a proteomic approach to identify Elg1 binding partners, we discovered that Elg1 interacts with Rtt106, a histone chaperone implicated in replication-coupled nucleosome assembly that also regulates transcription. A central role for Elg1 is the unloading of PCNA from chromatin following DNA replication, so we examined the relative importance of Rtt106 and PCNA unloading for chromatin reassembly following DNA replication. We find that the major cause of the chromatin organization defects of an ELG1 mutant is PCNA retention on DNA following replication, with Rtt106-Elg1 interaction potentially playing a contributory role.
| DNA replication is the central process that duplicates the genetic information during cell multiplication. Many cellular factors play important roles in the efficient and accurate duplication of DNA, critical for faithful transmission of genetic information. One such factor is Elg1. Elg1 acts to unload PCNA, the ring-shaped processivity factor that holds DNA polymerases on DNA for replication. In this work, we identify an additional role for Elg1 during replication. We show that lack of Elg1 leads to defects in packaging of DNA into chromatin after DNA replication. In addition, we found that Elg1 interacts with histone chaperones, factors which play key role in chromatin formation. Examining causes of the chromatin re-assembly defect, we show that accumulation of PCNA on DNA is the main cause of defective chromatin formation in the absence of Elg1. By uncovering a new route through which Elg1 ensures chromosomes are perfectly copied, our findings advance understanding of how Elg1 contributes to the stability of the genome through its key roles in DNA replication.
| The genetic material in eukaryotes is packaged into chromatin, composed mainly of DNA and nucleosomes. During DNA replication, DNA helicases separate the two parental strands of DNA and nucleosomes are removed from the DNA. Once the nascent DNA strands have been synthesized, the nucleosomal structure must be reassembled to restore the chromatin and permit reinstatement of epigenetic information. Defective chromatin re-assembly leads to improper chromatin formation and loss of epigenetic marks carried on the parental histones, resulting in genomic instability [1].
Various replication-associated factors play a key role in ensuring all the genetic and epigenetic information is efficiently duplicated. A critical component of the replication machinery is PCNA, which serves as the processivity factor for DNA polymerases. Apart from acting as an accessory factor for DNA polymerase, PCNA coordinates replication-associated processes including chromatin re-assembly, cohesion establishment, DNA repair and the damage response [2]. PCNA is loaded onto chromatin during replication by the Replication Factor C (RFC), a pentameric complex consisting of Rfc1-5 [3,4]. During the initiation of each Okazaki fragment, RFC loads PCNA prior to polymerase δ recruitment. On completion of each Okazaki fragment, PCNA must then be unloaded, which requires the Elg1 RFC-Like Complex (also called Elg1-RLC; [5,6]. The Elg1-RLC contains the same Rfc2-5 subunits as RFC, but the largest subunit Rfc1 is replaced by Elg1. Timely removal of PCNA is important, and PCNA accumulation in the absence of Elg1 contributes to genomic instability phenotypes such as elongated telomeres, telomeric silencing, chromosomal rearrangements, cohesion defects, and increased sister chromatin recombination [7–11].
Histone chaperones are crucial auxiliary components of the replication machinery [12,13], which ensure the proper coupling of DNA replication with re-assembly into nucleosomes [14]. FACT complex of the budding yeast S. cerevisiae contains subunits Spt16 and Pob3, and can bind both H2A-H2B and H3-H4. FACT associates with components of the replication machinery including the MCM complex and DNA polymerase δ [15,16] and acts in parental histone recycling and placement on the newly replicated DNA, as well as being implicated in transcription-coupled chromatin control [17,18]. In S. cerevisiae newly synthesized histone H3-H4 dimers are bound by the histone chaperone Asf1, with new histone H3 preferentially acetylated at H3K56. Asf1 binding and H3K56Ac modification promote the interaction of new H3-H4 with further histone chaperones including CAF-1 and Rtt106 [19], and Asf1 additionally interacts with RFC [20]. CAF-1 is a three subunit complex consisting in yeast of subunits Cac1, Cac2, and Cac3. Two CAF-1 complexes associate to assemble an H3-H4 tetrasome in the initial step of nucleosome re-assembly [21]. CAF-1 promotes nucleosome assembly at replication forks through interaction with PCNA and by binding to DNA directly [22–24]. Rtt106 is also implicated in nucleosome reassembly following DNA replication. Containing two Pleckstrin Homology Domains that mediate its preference for K56-acetylated H3 [21], Rtt106 has been shown to dimerize to mediate assembly of an H3-H4 tetrasome [25,26]. Deletion of RTT106 when combined with deletion of CAC1 showed a defect in deposition of H3K56Ac, which is marker of newly deposited histone in yeast [19,27]. Rtt106 is also involved in heterochromatin formation: rtt106Δ mutant cells exhibit loss of silencing at mating type loci and telomeres [19,28]. In addition, Rtt106 is proposed to be important for nucleosome assembly during transcription at highly transcribed genes [29] and in regulation of histone gene expression [30,31]. However, it remains unknown how Rtt106 is recruited to required sites of nucleosome assembly.
Because of the links between PCNA and nucleosome assembly, and the effects on chromatin and genome stability caused by ELG1 deletion [9], we were prompted to investigate whether the PCNA unloading factor Elg1 has a role also in the chromatin re-assembly process. Here we show that Elg1 activity is critical for timely nucleosome organization on nascent DNA. We moreover discovered that Elg1 interacts with histone chaperones, in particular Rtt106 and the FACT complex, with the interaction of Elg1 and Rtt106 not dependent on PCNA. We find however that the most significant cause of defective post-replication nucleosome organization in an elg1Δ mutant is delayed unloading of PCNA, with Elg1-Rtt106 interaction potentially playing a contributory role.
The process of DNA replication and nucleosome re-assembly are tightly coupled. Because it acts at replication forks in PCNA unloading, we examined if Elg1 also affects nucleosome deposition onto newly replicated DNA. Initially, we examined Okazaki fragment length in strains lacking Elg1. Okazaki fragment length can be used as a proxy for nucleosome deposition, since fragment length tends to be determined by the newly deposited nucleosome on the immediately preceding fragment [13,32]. To permit the visualization of Okazaki fragments, we used a strain background with an Auxin-Inducible Degron (AID)-tagged copy of the DNA ligase gene CDC9, which accumulates unligated Okazaki fragments during S phase in the presence of auxin. Cells were synchronized in G1 then released into S phase for 55 min, and then Okazaki fragments visualized by 3' end-labelling and gel electrophoresis as described [6,13] (Fig 1A & 1B). In normal cells, Okazaki fragment lengths tend to cluster around 180 bases and 360 bases corresponding to mono- and di-nucleosomal sizes. As previously described, Okazaki fragments are somewhat extended in the mutant cac2Δ which lacks the CAF-1 chromatin assembly factor (Fig 1C) [13,32]. This lengthening is believed to reflect aberrant and delayed nucleosome repositioning, which causes continued nick translation and Okazaki fragment lengthening by DNA polymerase δ, since it does not encounter a nucleosome on the previously synthesized DNA that would stimulate its disengagement. In an elg1Δ mutant, we found that Okazaki fragment lengths also differed from wild-type, showing a generally broader distribution with a higher proportion of fragments extended in length when compared to wild-type (Fig 1C & S1 Fig). This Okazaki fragment lengthening suggests that the elg1Δ mutation may cause a nucleosome assembly defect. The lengthened Okazaki fragment phenotype was not shared by a ctf18Δ mutant, which lacks the Ctf18-RLC complex that is involved in establishment of cohesion [33,34]. The effect of Elg1 in limiting Okazaki fragment length therefore appears specific to Elg1-RLC. Since Cdc9 depletion is intrinsic to the Okazaki fragment detection assay, we cannot exclude the possibility that lack of Cdc9 contributes to this Okazaki fragment lengthening effect in the elg1Δ mutant.
To examine chromatin re-assembly in elg1Δ using a different approach, we next tested the sensitivity of chromatin to digestion by Micrococcal Nuclease (MNase), since defective chromatin re-assembly can result in increased accessibility to digestion by this nuclease. There was no evident abnormality in MNase sensitivity of bulk chromatin in an elg1Δ mutant. However, defects in replication-coupled chromatin re-assembly tend to be transient and quickly restored following replication by redundantly acting histone chaperones and/or replication-independent histone turnover [35]. To test nucleosome deposition onto newly replicated DNA, we used cultures synchronized by release from α factor into S phase and examined the MNase sensitivity of nascent DNA labelled with the thymidine analog 5-Bromo 2-deoxyuridine (BrdU) (Fig 2A & 2B). These experiments used strains genetically modified to incorporate BrdU. After Southern blot transfer of MNase-digested DNA to membrane, nascent DNA was specifically visualized by probing the DNA on the membrane with anti-BrdU antibody. Validating the assay, nascent DNA in a cac1Δ mutant (S2C Fig) was more sensitive than wild-type to MNase digestion, due to delayed chromatin re-assembly [35]. We found that nascent DNA in the elg1Δ mutant (Fig 2C) was also more sensitive to MNase than wild-type, as evidenced by an increased proportion of mononucleosomal compared to disomal digested fragments (Fig 2C lower panel, compare proportion of disome and monosome bands and signal traces of 45 min samples of nascent DNA in Fig 2D). This increased sensitivity to MNase digestion in elg1Δ was reproducible, as illustrated by the additional gels shown in S2A & S2B Fig. The magnitude of the effect did vary between experiments: the proportion of mono-nucleosomal to total nascent DNA was increased 1.7-fold in elg1Δ relative to wild-type in Fig 2C, 1.2-fold in S2A Fig, and 2.6-fold in S2B Fig. Such variation is to be expected given the semi-quantitative nature of such experiments, but overall the elevated accessibility of nascent DNA to MNase digestion is indicative of defective or delayed nucleosome assembly. The differences in sensitivity to MNase are not caused by different rates of progression through S phase of WT and elg1Δ cells (S3 Fig). To summarize, our observation of extended Okazaki fragments and increased sensitivity to micrococcal nuclease in the elg1Δ mutant suggest a role for Elg1 in replication-coupled nucleosome re-organization.
The results presented above prompted us to investigate effects of the elg1Δ mutation on nucleosome assembly genome-wide. We used thymidine analog 5-ethynyl-2’-deoxy-uridine (EdU) to label newly replicated DNA in G1-arrested cells released into S phase. Following MNase digestion, EdU-labelled nascent DNA was isolated by affinity purification (Fig 3A). After deep sequencing [35], nucleosomal reads were then aligned with respect to origins of replication (Fig 3B & 3C) or transcription start sites (TSS) of all genes (S4 Fig). While no difference in the organization of nucleosomes either upstream or downstream of origins was observed in G1 control samples, a clear defect in organization of nucleosomes is observed in elg1Δ (Fig 3B) at early time points after release (27 min, 30 min, 33 min) when compared to WT. As cells reach the end of S phase (60 min) the nucleosomal pattern in the elg1Δ mutant becomes more organized and similar to WT, consistent with recovery of normal nucleosome distribution as previously described [35]. Defective nucleosome organization in elg1Δ mutant is somewhat similar to that seen in a cac1Δ mutant (Fig 3C & S4B Fig) although the cac1Δ mutant shows an increased spacing of nucleosomes on nascent DNA that is not obviously shared by elg1Δ.
To identify interaction partners of Elg1 potentially connected to nucleosome assembly, we used SILAC-based mass spectrometry to identify co-precipitating proteins. Strains expressing untagged or FLAG-tagged versions of Elg1 were differentially labelled with isotopically light or heavy lysine and arginine, and immunoprecipitated proteins (Fig 4A) were analyzed by mass spectrometry. As expected, the Elg1-FLAG samples showed strong enrichment of Elg1 and Rfc2-5 (the other Elg1-RLC subunits) and also of PCNA. Strikingly, the histone chaperone Rtt106 was also enriched at levels similar to the Rfc2-5 subunits (Fig 4B & 4C). Also enriched were Spt16 and Pob3, two subunits of the FACT complex. Both Rtt106 and FACT complex are implicated in replication-coupled nucleosome assembly: while FACT appears to mediate recycling of parental histones, Rtt106 is involved in depositing newly synthesized histones [18,19,36]. The interactions suggest that these histone chaperones, particularly Rtt106, could potentially mediate the nucleosome assembly role of Elg1.
We carried out further co-immunoprecipitation experiments with Rtt106 to confirm and investigate the Elg1-Rtt106 interaction. Immunoprecipitation of Elg1-FLAG pulled down HA-tagged Rtt106 (Fig 5A). Pulldown of Elg1 truncation mutants showed that both the Elg1 N-terminal and C-terminal regions are important for the interaction with Rtt106 (S6 Fig). These regions are unique to Elg1, having only very limited sequence similarity with Rfc1 or Ctf18. Consistently, neither Rfc1 nor Ctf18 showed interaction with Rtt106 in co-immunoprecipitation experiments (Fig 5B), suggesting interaction with Rtt106 is a property specific for Elg1 amongst the major subunits of RFC and its related complexes.
Immunoprecipitation of Elg1-FLAG pulled down not only Rtt106 but also PCNA, reflecting the function of Elg1-RLC as the major PCNA unloader. Co-immunoprecipitation experiments in the presence of increasing salt concentrations showed that interaction with PCNA was lost at a concentration where Rtt106-Elg1 interaction was retained (Fig 5C, 250mM potassium acetate & S7 Fig), indicating that the Elg1-Rtt106 interaction is not mediated through PCNA. Note that a band appearing in Western analysis slightly below full-length Elg1 (Fig 5A, 5B & 5C) appears to represent a degradation product whose appearance is stimulated by increased salt concentration. To summarize, our results indicate that robust interaction occurs between Elg1 and Rtt106, specific to Elg1 amongst the RFC-related complexes.
Since Elg1 is important for nucleosome deposition and interacts with Rtt106, we reasoned that, during DNA replication on the lagging strand, Elg1 might concomitantly recruit Rtt106 as it unloads PCNA, thereby coupling PCNA unloading and chromatin re-assembly. Alternatively, Rtt106 might participate in the PCNA unloading function of Elg1. Examining the accumulation of PCNA on chromatin in the absence of Rtt106 (S8 Fig) did not show clear evidence for a role for Rtt106 in PCNA unloading. We therefore followed up the possibility that Elg1 interaction is important to recruit Rtt106 for chromatin re-assembly, by investigating whether recruitment of Rtt106 to replicating regions is dependent on Elg1. We carried out ChIP-seq analysis of HA-tagged Rtt106 on cells released into hydroxyurea from a G1 arrest. However contrary to our expectation, we did not consistently observe association of Rtt106 with newly replicated regions at early origins (e.g. ARS306, ARS510, ARS310, S9B Fig). Nor did we observe convincing Rtt106 recruitment to replicating chromatin in a similar experiment carried out under unperturbed conditions (i.e. in WT cells with no HU treatment). Our ChIP experiments did effectively identify Rtt106 binding as we did observe Rtt106 localization at the promoter HTA1-HTB1 promoter (S9A Fig), as previously described [37]. Rtt106 recruitment to the HTA1-HTB1 promoter was not affected in the absence of Elg1 (S9A Fig). We did notice Rtt106 association with the promoters of some genes encoding putative drug exporters, that in some cases appeared Elg1-dependent. This promoter association does not appear replication-linked, since it was observed at some late-replicating regions that forks will not reach under the HU block conditions of the experiment. The importance of Rtt106 promoter binding will be described elsewhere.
Given the effect of Elg1 on chromatin re-assembly and its interaction with Rtt106, we tested whether the two proteins act in chromatin re-assembly in the same pathway. Specifically, we examined whether the elg1Δ and rtt106Δ mutations have similar effects on the length of Okazaki fragments. We found that rtt106Δ causes only mild lengthening of Okazaki fragments, the degree of lengthening much less than observed for elg1Δ. Moreover, the effect of elg1Δ rtt106Δ double mutation on Okazaki fragments appeared to be additive rather than epistatic when compared to the single mutations (Fig 6A). These effects suggest that Elg1 acts in a distinct pathway from Rtt106. Hence, we considered other mechanisms through which elg1Δ might affect chromatin re-assembly. The absence of Elg1 results in prolonged accumulation of PCNA on chromatin [11], which could potentially interfere with nucleosome deposition causing defective chromatin re-organization. To investigate this possibility, we made use of trimer instability mutations in PCNA. These mutations cause the PCNA ring to be disassembly-prone, falling off DNA spontaneously even in the absence of Elg1 and thereby suppressing the PCNA accumulation phenotype of the elg1Δ mutant [11]. Okazaki fragment length assays were performed in double mutants where elg1Δ was combined with two different trimer instability PCNA mutants, pol30-R14E (Paul Solomon Devakumar et al. in revision) and pol30-D150E [11]. We observed that in these double mutants, Okazaki fragments were restored to normal length, when compared to the elongated Okazaki fragments of the elg1Δ single mutant (Fig 6B & S10 Fig). Based on this observation, we propose that when normal PCNA unloading fails due to absence of Elg1, aberrant PCNA accumulation on the newly replicated DNA leads to defective nucleosome deposition.
In this investigation, we show that Elg1 contributes to proper nucleosome assembly across the genome after DNA replication, as evidenced by Okazaki fragment lengthening (Fig 1) and elevated sensitivity of nascent DNA to micrococcal nuclease digestion (Figs 2 & 3) in an elg1Δ mutant. Okazaki fragment length has previously been examined in several studies as a proxy for nucleosome deposition [32]. This assay could raise the concern that the DNA ligase-deficient background required to visualize Okazaki fragments might itself impact on fragment length or nucleosome re-assembly, but a different study [38] obtained consistent results, also finding that nucleosome position determines S. cerevisiae Okazaki fragment positioning, using a completely different approach that analyzed mutations inserted by an error-prone polymerase α prone to ribonucleotide insertion. Moreover, in assays that measure the micrococcal nuclease sensitivity of nascent DNA (in cells where DNA ligase activity is intact) we confirmed that nucleosome deposition is affected by the elg1Δ mutation. Therefore, the Okazaki fragment lengthening phenotype indeed reflects a nascent strand chromatin re-assembly defect.
To understand interactions that may contribute to the chromatin re-assembly effect of Elg1, we examined the proteins that co-precipitate with Elg1 in pull-down experiments, and identified novel interactions of Elg1 with histone chaperones, in particular Rtt106 and the FACT complex. Interestingly, Rtt106 appears to bind the Elg1-RLC in almost stoichiometric amounts, in an interaction that does not depend on PCNA. Rtt106 does not interact with either Rfc1 or Ctf18. Consistently, we found that both the N-terminal and C-terminal regions that are unique to Elg1 are needed for Rtt106 interaction (S6 Fig).
To examine the extent to which Rtt106-Elg1 interaction versus the Elg1 PCNA unloading function are important for chromatin re-assembly, we made use of disassembly-prone mutants of PCNA which do not accumulate on chromatin even in the absence of Elg1. Using these mutations to relieve PCNA accumulation on chromatin in an elg1Δ background restored Okazaki fragments to normal length, indicating that prompt and effective PCNA unloading is absolutely essential for normal nucleosome deposition in the wake of replication forks.
How might PCNA accumulation result in defective nucleosome assembly and associated Okazaki fragment lengthening? Okazaki fragment length is proposed to be regulated by nucleosome deposition on the previously synthesized section of DNA [13,38] as illustrated in Fig 7. The newly deposited nucleosome on the last piece of DNA synthesized is believed to form an obstacle to progression of polymerase δ as it carries out strand displacement synthesis, prior to completing synthesis of each Okazaki fragment. Encounter of pol δ with the nucleosome is suggested to favour pol δ disengagement and dissociation, allowing PCNA to recruit DNA ligase [39] with ligation of the completed Okazaki fragment to the nascent lagging strand determining the final Okazaki fragment length (Fig 7, Model i). We propose that in the absence of Elg1, accumulated PCNA in the wake of the replisome obstructs normal placement and spacing of nucleosome deposition, so that the nucleosomal barrier to pol δ synthesis is not present, resulting in longer Okazaki fragments being synthesized prior to their eventual completion and ligation (Fig 7, Model ii). Combining the elg1Δ mutation with a PCNA trimer-unstable mutant prevents the accumulation of PCNA, relieving the block to nucleosome deposition and restoring the normal mechanism of Okazaki fragment length determination (Fig 7, Model iii).
Our findings support the suggestion that nucleosome deposition is a very early event that precedes and stimulates pol δ dissociation, the polymerase in turn allowing DNA ligase recruitment by PCNA [39] and subsequent Okazaki fragment ligation, which is necessary for PCNA unloading by the Elg1-RLC. Our results are therefore consistent with the previously identified dependence of PCNA unloading on Okazaki fragment ligation [6].
A very recent study by [40] provides an interesting illustration of the consequences of disrupting PCNA removal by Elg1-RLC and nucleosome deposition. Janke et al used an assay that measures heterochromatin disruption, by testing for failure to silence expression of a Cre recombinase gene. Their finding that silencing is disrupted by an elg1Δ mutation (or by histone chaperone mutations) implies that normal replication-coupled chromatin assembly is needed to preserve silencing at a specific heterochromatic locus. Our study generalizes the conclusion that Elg1 activity is needed for normal chromatin inheritance, with the discovery that nucleosome deposition problems caused by failure to unload PCNA extend genome-wide.
Since delayed PCNA removal appears to be the main cause of the chromatin re-assembly defect observed in elg1Δ, what is the significance of Elg1 interaction with histone chaperones, in particular Rtt106 and FACT complex? Identification of these interactions raises the suggestion that Elg1 might recruit histone chaperones to assist in chromatin reassembly, with Elg1 thereby contributing to chromatin re-configuration or maturation. However, our ChIP analysis failed to identify a clear role for Elg1 in localizing Rtt106 to newly synthesized DNA. We did find that Elg1 has effects on Rtt106 chromatin association at the promoters of a number of genes, particularly genes involved in cellular transport and drug resistance. However, this effect is unlikely to be coupled to DNA replication since we observed Rtt106 association with several such sites in G1 phase samples. Slight sensitivity of an elg1Δ mutant to HU [8] would be consistent with a need for Elg1 in controlling the expression of genes required for drug response and export. The possibility of a non-replication-associated role for Elg1 in regulating gene expression through histone chaperone recruitment is the subject of ongoing study.
While PCNA accumulation appears to be the immediate cause of delayed nucleosome deposition in the elg1Δ mutant (Fig 7), our results do not exclude the possibility of a role for Rtt106-Elg1 in replication-coupled chromatin re-establishment, especially since presence of multiple, redundant histone chaperones activities in yeast complicates analysis of chromatin re-assembly. However, we could not obtain unambiguous, reproducible evidence of a role for Elg1-Rtt106 interaction following replication. One possibility is that Elg1 does contribute to coordination of chromatin re-assembly, operating through Rtt106 and/or other histone chaperones, in a pathway acting at a later stage of chromatin maturation operating after histone deposition and Okazaki fragment ligation.
The role of Elg1 appears to be conserved, since its mammalian homolog, called ATAD5, also appears to mediate PCNA unloading [41]. Mammalian cells lacking ATAD5 show PCNA accumulation on chromatin similar to that observed in yeast, and it seems likely that such PCNA retention may impact chromatin re-assembly. The major phenotype of mice lacking ATAD5 is cancer predisposition, and indeed ATAD5 mutations are also proposed to contribute to human ovarian cancers [42,43]. Defects in genomic function caused by derailed chromatin re-assembly following replication might therefore contribute significantly to human cancer development or progression.
All yeast strains used in this study are listed in S1 Table. Gene disruptions and epitope tags were introduced by standard PCR based methods [44,45].
Okazaki fragment purification and detection was performed as described previously in [13].
Yeast cells were grown to OD600 of 0.2 at 30°C in 60ml YPD media and then alpha factor was added to arrest cells in G1 phase. 400μg/ml BrdU was added to the culture and incubated for 30 minutes for cells to take up BrdU. Cells were then released into S phase by resuspending in fresh YPD containing 400μg/ml BrdU. Then 20 ml samples were collected at desired time points into formaldehyde (1% final concentration) and incubated with rotation for 15 minutes at room temperature. 125mM glycine was then added to neutralise formaldehyde. Cells were washed twice in 10 ml of ice cold 1X PBS, then with 2 ml of spheroplasting buffer (1M sorbitol, 1mM beta-mercaptoethanol) before resuspending in 1ml of spheroplasting buffer with 300μg/ml 100-T Zymolase then incubated at 30°C for 20 minutes. Spheroplasts were washed in 1ml of spheroplasting buffer and resuspended in 600μl of Digestion buffer (1M sorbitol, 50mM Nacl, 10mM Tris-HCl pH7.4, 5mM MgCl2, 5mM CaCl2, 0.075% Nonidet P-40, 1mM beta-mercaptoethanol, 0.5mM spermidine). 200μl aliquots were subjected to micrococcal nuclease (NEB, M0247S) digestion (200 or 600 gel units) for 5 minutes at 37°C. Digestions were stopped by adding 1/10 volume of stop solution (250mM EDTA, 5% SDS). 5μl of 20 mg/ml Proteinase K was added and incubated overnight at 55°C. Following phenol-chloroform extraction, DNA was precipitated using 1/10 volume of 3M sodium acetate and 2 volumes of 100% ethanol. The air-dried DNA pellet was resuspended in 20μl of TE buffer with RNase A (1mg/ml) and incubated for 2 hours at 37°C. DNA samples were electrophoresed on a 1.4% agarose gel, which was incubated in denaturing buffer (0.5M NaoH, 1M NaCl) twice for 15 minutes followed by incubation in neutralization buffer (0.5M Tris-Hcl, 3M Nacl) for 30 minutes. The DNA was then transferred to Amersham Hybond N+ membrane by Southern blotting. DNA was cross-linked to the membrane with UV light (1200J). The membrane was then incubated in 5% milk in TBS-tween for 60 minutes and probed with anti-BrdU antibody (ab12219, abcam).
Whole cell extract preparation, western blotting and co-immunoprecipitation experiments were performed as described previously [5,6]. Antibodies used were: anti-BrdU (ab12219, abcam), anti-FLAG (F1804, Sigma), anti-HA (HA.11 clone 16B12, Covance), anti-PCNA (ab70472, abcam).
SILAC Quantitative proteomic analysis was performed as described previously [46].
Yeast strains were grown to an OD600 of 0.25 in YPD. Alpha factor was added to arrest cells in G1 and released into YPD containing 0.2M hydroxyurea at 23°C for 60 minutes. Formaldehyde (1% final concentration) was added and incubated with rotation first at room temperature for 20 minutes and then at 4°C overnight. Cells were washed 3 times with ice-cold 1X Phosphate buffered saline. Cells were pelleted and frozen at -80°C. Rtt106 ChIP using anti-HA (HA.11 clone 16B12, Covance) and data analysis were performed as described previously [6].
ChIP-Seq data are uploaded to Array Express under accession number: E-MTAB-6985
|
10.1371/journal.pgen.1000605 | CHD3 Proteins and Polycomb Group Proteins Antagonistically Determine Cell Identity in Arabidopsis | Dynamic regulation of chromatin structure is of fundamental importance for modulating genomic activities in higher eukaryotes. The opposing activities of Polycomb group (PcG) and trithorax group (trxG) proteins are part of a chromatin-based cellular memory system ensuring the correct expression of specific transcriptional programs at defined developmental stages. The default silencing activity of PcG proteins is counteracted by trxG proteins that activate PcG target genes and prevent PcG mediated silencing activities. Therefore, the timely expression and regulation of PcG proteins and counteracting trxG proteins is likely to be of fundamental importance for establishing cell identity. Here, we report that the chromodomain/helicase/DNA–binding domain CHD3 proteins PICKLE (PKL) and PICKLE RELATED2 (PKR2) have trxG-like functions in plants and are required for the expression of many genes that are repressed by PcG proteins. The pkl mutant could partly suppress the leaf and flower phenotype of the PcG mutant curly leaf, supporting the idea that CHD3 proteins and PcG proteins antagonistically determine cell identity in plants. The direct targets of PKL in roots include the PcG genes SWINGER and EMBRYONIC FLOWER2 that encode subunits of Polycomb repressive complexes responsible for trimethylating histone H3 at lysine 27 (H3K27me3). Similar to mutants lacking PcG proteins, lack of PKL and PKR2 caused reduced H3K27me3 levels and, therefore, increased expression of a set of PcG protein target genes in roots. Thus, PKL and PKR2 are directly required for activation of PcG protein target genes and in roots are also indirectly required for repression of PcG protein target genes. Reduced PcG protein activity can lead to cell de-differentiation and callus-like tissue formation in pkl pkr2 mutants. Thus, in contrast to mammals, where PcG proteins are required to maintain pluripotency and to prevent cell differentiation, in plants PcG proteins are required to promote cell differentiation by suppressing embryonic development.
| In higher eukaryotes only a small proportion of genomic information is required in any specific cell type at a given developmental stage. The intricate decision whether a gene should be active or repressed is made by the counteractive activities of trithorax group (trxG) and Polycomb group (PcG) proteins that form part of a chromatin-based cellular memory system. Here we show that the CHD3 proteins PICKLE and PICKLE RELATED2 (PKR2) have trxG-like functions in plants and activate PcG protein target genes. Lack of PKL function can partially suppress PcG mutant leaf and flower phenotypes, supporting the idea that CHD3 proteins and PcG proteins act antagonistically during plant development. We identified PcG genes among the direct PKL/PKR2 targets in roots and demonstrated that lack of pkl pkr2 results in reduced PcG protein activities, leading to similar root phenotypes in pkl pkr2 and PcG protein mutants. Previous studies have implicated PKL as a transcriptional repressor, but we provide evidence that CHD3 proteins such as PKL and PKR2 act as transcriptional activators in plants and assume trxG-like function to counteract PcG protein–mediated gene repression.
| Dynamic regulation of chromatin structure is the underlying scheme for modulating genome activities in higher eukaryotes. There are two major classes of proteins with enzymatic activities directed at chromatin - histone modifying enzymes and ATP dependent chromatin remodelers. Histone modifying enzymes add or remove posttranslational modifications such as acetylation, methylation, phosphorylation and ubiquitinylation on histones. These modifications are recognized and bound by factors that cause changes in chromatin structure by not well understood mechanisms [1]. ATP dependent chromatin remodelers modify chromatin structure by altering interactions between DNA and histone octamers, resulting in changes of nucleosome position or composition associated with changes in nucleosomal DNA accessibility [2].
CURLY LEAF (CLF) and PICKLE (PKL) are examples of these two enzyme classes in plants. CLF is a Polycomb group (PcG) protein with histone methyltransferase activity [3],[4], and PKL is a predicted ATP-dependent chromatin remodeling factor of the chromodomain/helicase/DNA-binding domain (CHD3) subfamily [5],[6].
Members of the CHD3 subfamily are characterized by the presence of two tandemly arranged chromodomains as well as the presence of one or two PhD (plant-homeo-domain) zinc fingers preceding the chromodomains [7]. CHD3 family members of flies and mammals are part of the NuRD (nucleosome remodeling and deacetylase) multiprotein complex that is implicated to couple ATP-dependent chromatin remodeling and deacetylation resulting in transcriptional repression [7]. However, several studies also implicate a function of CHD3 family members in transcriptional activation [8]–[10].
CLF is a homolog of the metazoan SET domain protein Enhancer of zeste, and similar to animal PcG proteins CLF is part of a multiprotein Polycomb repressive complex 2 (PRC2)-like complex that trimethylates histone H3 on lysine 27 (H3K27me3) [4],[11],[12]. This modification is recognized by the chromodomain containing protein LIKE HETEROCHROMATIN PROTEIN 1 (LHP1) that together with the RING finger domain proteins AtRING1a and AtRING1b causes gene repression by not yet understood mechanisms [13]–[15]. Lack of CLF function causes reduced H3K27me3 levels associated with pleiotropic developmental aberrations like formation of curled leaves, homeotic transformations of flowers and early flowering [4],[16],[17]. CLF acts partially redundant with its homolog SWINGER (SWN), and lack of both proteins causes cells to de-differentiate and to form callus-like tissues that give rise to somatic embryos [18]. Similarly, lack of PKL function causes derepression of embryogenic traits in seedling roots, accumulation of seed storage reserves and formation of somatic embryos; albeit this pickle root phenotype only occurs with very low penetrance [19]. The embryonic master regulator gene LEAFY COTYLEDON1 (LEC1) is activated in both, pkl and clf swn double mutants [6],[20]. Overexpression of LEC1 causes somatic embryogenesis [21],[22], suggesting that LEC1 is critically responsible for somatic embryogenesis in pkl and clf swn mutants. Thus, lack of PcG proteins CLF and SWN as well as lack of PKL causes cell dedifferentiation and somatic embryogenesis, however, the underlying molecular mechanisms for this common phenotype remain unclear. Recent studies observed an overlap of genes being up-regulated in pkl mutants and genes enriched for H3K27me3, suggesting a functional connection of PKL and PcG pathways [23]. This idea was supported by the finding that lack of PKL caused reduced H3K27me3 levels at selected loci, whereas histone acetylation levels remained largely unaffected in pkl mutants [23]. Thus it seemed unlikely that PKL is part of a NuRD-like complex in plants but rather assumes an as yet unidentified role in gene regulation.
We set out to identify the functional connection of PKL and PcG proteins and found that in contrast to its anticipated role as a repressor, PKL has trithorax group (trxG)-like functions and is required for the activation of PcG target genes. Among its direct targets we identified the genes for the PcG proteins SWN and EMF2. Lack of PKL as well as its close homolog PKR2 caused reduced expression of genes for PcG proteins in primary roots, concomitantly with reduced H3K27me3 levels. This in turn is likely responsible for increased LEC1 expression and derepression of embryonic traits in pkl pkr2 primary roots. Expression of PcG genes is independent of PKL in aerial plant tissues and pkl can partly suppress the clf leaf and flower phenotype, supporting the idea that PKL and PcG proteins antagonistically determine cell identity in plants.
Investigations of the underlying molecular mechanism of the pickle root phenotype have been hampered by the very low penetrance of this phenotype. Therefore, we tested whether double mutants of pkl with mutants in close PICKLE homologs PICKLE RELATED1 (PKR1) and PKR2 [24] had an increased penetrance of this phenotype. We isolated mutant alleles for both genes, located in exon 9 in pkr1-1 and in exons 9 and 5 in pkr2-1 and pkr2-2, respectively (Figure 1A). Based on the expression of PKR1 and PKR2 in isolated homozygous mutant alleles, we concluded that all three alleles are likely to be null alleles (Figure 1B). Neither pkr1 nor pkr2 homozygous mutants exhibited significant phenotypic differences to wild-type plants under standard growth conditions (data not shown). However, whereas pkl pkr2 had a strongly increased penetrance of the pkl root phenotype, no increase was observed in the pkl pkr1 double mutant (Figure 1C). This suggests that PKR2 acts redundantly with PKL in suppressing cell dedifferentiation in the seedling root. This idea is supported by the finding that PKR2 expression was induced in pkl roots (Figure 1D). The pkr2 mutant did not enhance other aspects of the pkl phenotype (data not shown); consistent with lack of PKR2 expression in other vegetative plant organs (Figure 1D). PKR2 was strongly expressed in flowers and siliques; however, reproductive development was not disturbed in pkr2 single and pkl pkr2 double mutants (data not shown).
To investigate the molecular basis of the pickle root phenotype, we profiled transcriptomes of pkl and pkl pkr2 roots at five days after germination. Consistent with the strongly increased penetrance of the pickle root phenotype in the pkl pkr2 double mutant, we observed a synergistic increase in the number of up- and down-regulated genes in the double mutant (Figure 2A, Table S1). Next we used principal components analysis (PCA) to visualize the relation of pkl and pkl pkr2 mutant roots to wild-type roots, leaves and seeds. PCA was performed on the 4 samples from this study and 11 samples from the AtGenExpress developmental reference data set [25] using expression data of the 611 genes with altered expression in pkl or pkl pkr2 (Figure S1). The primary principle component accounted for 45% of the variation in the data and differentiated between seeds containing embryos and non-embryonic tissue such as roots or leaves. The second principle components accounted for 29% of the variation in the data and differentiated between photosynthetic active (leaves) and inactive (root) samples. Leaf, root and seed samples all clustered tightly in the PCA plot (Figure S1). The pkl and pkl pkr2 samples did not cluster tightly with wild-type or pkr2 roots but were located between the root and seed clusters indicating a partial change in cell identity from non-embryonic to embryonic fate. The positions of pkl and pkl pkr2 in the PCA plot were consistent with the hypothesis that the pkr2 mutation enhances the pickle root phenotype. Previous studies revealed that expression of LEC1 is critically important for cell dedifferentiation and embryonic fate [21],[22],[26]; consistent with this idea we found that LEC1 as well as embryonic regulators FUS3 and ABI3 and other seed-specific genes were synergistically up-regulated in the pkl pkr2 double mutant (Figure 2B and Figure S2A, S2B).
To explore the connection between PcG proteins and PKL, we tested whether genes that had altered expression in the pkl and pkl pkr2 double mutant were enriched for H3K27me3 [27]. We found a significant overlap with both, up- as well as down-regulated genes in pkl and in pkl pkr2 mutants (Figure 2C, Table S1). It has been reported previously that up-regulated genes in the pkl mutant are enriched for H3K27me3 [23], however, the strong enrichment for H3K27me3 among down-regulated genes was unexpected.
The strong enrichment for H3K27me3 among down-regulated genes prompted us to ask whether PKL was directly required for gene repression, gene activation or whether it had dual function. To address this question we performed chromatin immunoprecipitation (ChIP) using PKL-specific antibodies (Figure S3) and tested binding of PKL to the promoter region of genes with altered expression in pkl and pkl pkr2 mutants. We detected significant PKL binding to three genes that we picked from the top seven down-regulated genes (Figure 3); however, we did not detect significant PKL binding to the up-regulated genes LEC1, FUS3 and ABI3 (Figure 3), suggesting that PKL is directly required for the activation, but not repression of defined genes.
Consistent with results showing reduced H3K27me3 levels at several up-regulated genes in the pkl mutant [23], we detected significantly reduced H3K27me3 amounts at LEC1 and ABI3 promoter regions in pkl pkr2 mutants. No reduction in H3K27me3 levels was observed at the FUS3 locus (Figure 3), suggesting increased FUS3 expression is mediated by LEC1 that was previously shown to activate expression of FUS3 and ABI3 [26]. However, we also detected significantly reduced H3K27me3 levels in the promoter region of one of the genes with reduced expression in pkl and pkl pkr2 mutants (Figure 3), suggesting that loss of H3K27me3 is not sufficient for gene activation in pkl and pkl pkr2 mutants.
To test this hypothesis, we analyzed expression of confirmed PKL target genes and other genes with reduced expression in the pkl mutant in clf and pkl clf double mutants. It was known that lack of CLF causes strong reductions in H3K27me3 [4]. We found that lack of CLF in a PKL+/+ background led to increased expression for three of five tested genes with decreased expression in pkl or pkl pkr2 (Figure 4A). In contrast, lack of CLF in a pkl background did not affect expression of the five tested genes. Thus, increased expression of the test genes upon loss of the repressor CLF requires the presence of PKL. These results support our hypothesis that PKL activity is indeed required for gene activation.
In contrast, LEC1 and FUS3 were synergistically up-regulated in pkl clf double mutants (Figure 4B), supporting the idea that PKL and CLF are required for repression of both genes. Given that LEC1 and FUS3 are not direct target genes of PKL (Figure 3) suggests that PKL indirectly represses target genes by activating a repressor. Consistent with increased expression of LEC1 and FUS3 in pkl clf mutants, we observed a significantly increased penetrance of the pickle root phenotype in the double mutant (Figure 4C). To summarize, a set of PcG target genes were directly bound by PKL, had reduced expression in pkl and pkl pkr2 mutants, and additional loss of CLF did not affect expression. Other PcG target genes were not directly bound by PKL, had increased expression in pkl and pkl pkr2 mutants, and additional loss of CLF led to a further increase in expression.
A subset of PcG target genes was de-repressed in pkl and pkl pkr2 mutants, and we wondered whether this could be caused by reduced expression of genes for PcG proteins. Therefore, we tested expression of FIE, EMF2, VRN2, CLF, SWN, MEA, and MSI1 in roots of pkl and pkl pkr2 mutants. Indeed, we detected strongly reduced expression of EMF2, CLF, and SWN (Figure 5A), suggesting that PKL is directly or indirectly required for the activation of PcG protein encoding genes. To distinguish between both possibilities, we performed ChIP analysis and tested binding of PKL to the promoter regions of EMF2, CLF and SWN. We clearly detected binding of PKL to EMF2 and SWN promoter regions, but no binding was detected to the promoter region of CLF (Figure 5B) and neither to regions within the gene body (data not shown). Thus, we conclude that PKL is directly required for the activation of EMF2 and SWN, whereas PKL-mediated activation of CLF is possibly an indirect effect. We found SWN and EMF2 promoter regions marked by H3K27me3, but no significant enrichment for this mark was detected at the CLF promoter, suggesting that PKL is targeted preferentially to PcG target genes. This conclusion was supported by the observation that genes with reduced expression in pkl and pkl pkr2 mutants were also significantly enriched for H3K27me3 (Figure 2C). Previous whole genome analysis of H3K27me3 distribution did not reveal enrichment for H3K27me3 at EMF2 and SWN loci [27], which might be attributed to the use of whole seedlings by Zhang and colleagues (2007) in contrast to the root tissues used here.
Loss of CLF function leads to reduced H3K27me3 levels [4],[28]; therefore, we tested whether reduced expression of genes for PcG proteins EMF2, CLF and SWN in pkl pkr2 was reflected in reduced H3K27me3 levels. We assayed global H3K27me3 levels and indeed found less H3K27me3 in pkl pkr2 than in wild-type primary roots (Figure 5C). Previously, we observed induced expression of embryo-specific genes such as LEC1 and FUS3 in clf swn seedlings [20]. Therefore, we tested whether clf swn seedlings developed similar embryonic characteristics like pkl [19] and pkl pkr2 mutants. Indeed, clf swn seedlings were clearly stained with the neutral lipid staining dye Fat Red [19], indicating the accumulation of seed storage specific triacylglycerols (Figure 6A). Triacylglycerol accumulation in clf swn seedlings was only detected in structures developing from above-ground organs, suggesting that the third E(Z) homolog MEA, which is expressed in wild-type and clf swn roots ([29] and data not shown), can compensate the lack of CLF and SWN functions in roots. Triacylglycerols accumulated in pkl pkr2 seedlings only in primary roots, suggesting that PKL-mediated repression of EMF2, CLF and SWN was restricted to primary root tissues. When testing this hypothesis we found normal expression of EMF2, CLF and SWN in aerial parts of pkl and pkl pkr2 seedlings (Figure 6B). However, lack of PKL function strongly enhanced the clf swn phenotype, and pkl clf swn triple mutants only formed callus-like tissues that accumulated triacylglycerols (Figure 6A). It is possible that PKL is required for PcG gene activation in primary roots of wild-type plants but also in aerial parts of clf swn mutants; reduced expression of other PcG genes would then enhance the clf swn phenotype. To summarize, we propose that development of embryonic traits in pkl pkr2 is a secondary consequence of reduced expression of genes for PcG proteins, resulting in reduced levels of H3K27me3 and faulty expression of embryonic regulators such as LEC1, FUS3 and ABI3. This hypothesis predicts a significant overlap of genes up-regulated in pkl pkr2 and genes up-regulated in LEC1 overexpressing lines [22]. In agreement with this prediction the overlap of genes up-regulated in pkl pkr2 and in LEC1 overexpressing lines was significant (p<1E-15). In contrast, no significant overlap was detected between down-regulated genes of both datasets (Figure S4 and Table S1).
Our transcriptional profiling experiments revealed a significant overlap of genes with reduced expression levels in pkl and pkl pkr2 mutants and genes marked by H3K27me3 (Figure 2C), suggesting that PKL acts as transcriptional activator for PcG protein target genes. To test this hypothesis we analyzed adult phenotypes of pkl clf double mutants. Expression of PcG genes CLF, SWN and EMF2 was not affected by loss of PKL/PKR2 function in adult leaves (Figure 7A); therefore, lack of PKL function in a clf mutant background is expected to suppress at least partially the clf mutant phenotype.
There are two prominent phenotypic changes caused by lack of CLF function, (i) clf mutants have narrowed and upward curled leaf blades and (ii) clf flowers have partial homeotic transformations of sepals and petals towards carpels and stamens, respectively [16]. Both phenotypes were clearly suppressed in the pkl clf double mutant. The leaf blade of pkl clf plants was flat like the blade of wild-type leaves (Figure 7B) and we did not observe flowers with homeotic transformations in pkl clf plants. In contrast, about 30% of clf flowers developed homeotic transformations (Figure 7C and 7D). Thus, consistent with our hypothesis that PKL and CLF have antagonistic roles, lack of PKL function largely suppressed the clf mutant phenotype. To test whether we could find further molecular support for this hypothesis, we tested expression of the known CLF target genes AP3, AG and FLC [4],[28] in clf, pkl and clf pkl double mutants. All three genes had increased expression levels in clf mutants; but expression levels were greatly reduced in the clf pkl double mutant (Figure 7E). Finally, we tested whether PKL directly binds to AP3, AG and FLC and performed ChIP analysis of wild-type, clf and pkl seedlings. We clearly detected binding of PKL to the promoter regions of all three genes in wild-type as well as in clf seedlings (Figure 7E). Because PKL binding to AP3, AG and FLC occurred in wild-type as well as in clf seedlings while PKL-dependent activation of these genes was only observed in clf mutants, it is possible that PKL can activate transcription only in the absence of H3K27me3. In line with this hypothesis we detected significantly reduced levels of H3K27m3 at the three tested loci in clf seedlings. Thus, developmental and molecular phenotypes of clf pkl double mutants and direct binding of PKL to PcG target genes support the conclusion that PKL is required for the activation of PcG target genes.
The chromatin remodeling factor PKL has been implicated in maintenance of cell identity in plant seedlings by suppressing seed-associated developmental programs [19], [30]–[32], and we found that PKL acts redundantly with PKR2. PKL and PKR2 are homologs of metazoan CHD3/CHD4 proteins [24] that are part of multisubunit complexes with histone deacetylase activity such as the NuRD complex [7],[33]. Therefore, it was suggested that PKL acts as transcriptional repressor and suppresses embryonic regulators like LEC1 and FUS3 [6], [23], [30]–[32]. However, previous studies did not detect any effect of PKL activity on acetylation levels, casting doubt on the idea that PKL might be part of a plant NuRD-like complex [23]. Instead, Zhang and colleagues (2008) proposed that PKL is involved in H3K27me3-mediated transcriptional repression, because they found in pkl mutants reduced H3K27me3 levels and increased expression for LEC1, FUS3 and several other loci. Nevertheless, the connection between PKL and PcG protein-mediated H3K27me3 remained unclear.
We used ChIP to test binding of PKL to genes that have altered expression in pkl mutants and therefore are potential direct PKL target genes. We failed to detect direct binding of PKL to any of the up-regulated genes. In contrast, we detected direct binding of PKL to several of the down-regulated genes. Therefore, we conclude that PKL does not act as transcriptional repressor but as transcriptional activator. Interestingly, Drosophila dCHD3 acts as a monomer and is not part of a NuRD-like complex [10]. Furthermore, Drosophila CHD3/CHD4 proteins dMi-2 and dCHD3 colocalize with active RNA polymerase II on polytene chromosomes [9],[10]. Thus, it is possible that at least some members of the CHD3/CHD4 protein family could have a role in transcriptional activation in animals as well.
PKL can act as transcriptional activator, and genes that are down-regulated in pkl mutants are of particular interest. We observed a significant overlap of this set of down-regulated genes with genes reported to carry H3K27me3. All identified direct PKL target genes carry H3K27me3. Thus, one major group of genes activated by PKL consists of PcG protein target genes. Because PKL acts as transcriptional activator of PcG protein target genes, PKL can be considered as a plant trxG protein. A trxG function of a CHD protein is not without precedence, as the Drosophila CHD protein Kismet-L counteracts PcG protein-mediated repression by promoting transcription elongation through recruiting ASH1 and TRX histone methyltransferases [34]. For PKL, this idea is supported by the partial suppression of the clf mutant phenotype by pkl. In summary, direct activation of PcG genes by PKL explains the down-regulation of many genes with H3K27me3 in pkl mutants.
In addition to down-regulation of H3K27me3-covered genes in pkl, we observed up-regulation of many H3K27me3-covered genes as well. This is consistent with earlier observations by others [23]. Because we failed to detect direct binding of PKL to any of the up-regulated genes, we conclude that increased expression of these genes in pkl is an indirect effect. We show that this indirect effect is caused by reduced expression of PcG protein encoding genes in pkl pkr2 roots. We find that in roots EMF2 and SWN loci contain H3K27me3, the hallmark of PcG protein-mediated regulation. Thus, EMF2 and SWN, which code for PcG proteins, are themselves PcG protein targets. Autoregulation of genes for PcG proteins has been observed before in Drosophila [35]. In Arabidopsis, the MEDEA (MEA) gene, a homolog of E(Z), is repressed by PcG proteins in post-embyronic tissues [36],[37]. Similar to many other PcG protein target genes, EMF2 and SWN require PKL for efficient expression, because in pkl pkr2 roots expression of both genes is strongly reduced. Expression of CLF in pkl pkr2 roots is reduced as well, but this could be an indirect effect because we detected neither H3K27me3 at the CLF locus nor PKL binding to CLF. Together, reduced expression of EMF2, SWN and CLF explains the reduced H3K27me3 levels in pkl pkr2 roots and the de-repression of a number of PcG target genes. As lack of PKL did not prevent increased expression of LEC1, FUS3, ABI3 as well as many other PcG target genes, we propose that PKL is required for the activation of only a subset of PcG target genes.
PKL and PKR2 are expressed mostly in the seedling root ([32] and Figure 1C), and loss of cell identity in pkl pkr2 is restricted to primary root tissues. Thus, PKL and PKR2 function mainly in the seedling root; other proteins might activate PcG protein target genes in aerial organs, possibly other PKL homologs.
PcG proteins in plants and animals are master regulators of genomic programs [38]. However, whereas in mammals PcG proteins are required to maintain pluripotency and to prevent cell differentiation, in plants PcG proteins are required to promote cell differentiation by suppressing embryonic development. The PcG proteins CLF and SWN act redundantly and lack of both, CLF and SWN causes cells to de-differentiate and to form callus-like tissues that give rise to somatic embryos [18]. EMF2 is likely part of PRC2-like complexes together with CLF and SWN [18]; EMF2 interacts with both, CLF and SWN in yeast and a weak emf2 mutant allele resembles clf [18]. Mutant studies support the idea that PcG protein function is impaired in pkl and pkl pkr2 roots: First, the pkl pkr2 and clf swn double mutants have similar phenotypes. Both activate the embryonic master regulator LEC1 (this study and [20]) and both express embryonic traits in seedlings. Second, the clf swn mutant phenotype is strongly enhanced in the pkl clf swn triple mutant, causing complete transformation of germinating seedlings into callus-like tissues. For several reasons we believe that reduced expression of PcG genes is rather the cause than the consequence of the pkl root phenotype: (i) PcG genes EMF2 and SWN are direct target genes of PKL, (ii) about 35% of pkl pkr2 mutants undergo transformation to pkl roots, whereas expression levels of PcG genes CLF and SWN are reduced to 20% of wild-type expression levels, indicating reduced expression levels of PcG genes in roots that do not adopt a pkl phenotype, (iii) in line with the last argument, expression of PcG genes was indeed reduced in pkl pkr2 roots that did not undergo discernable transformations (data not shown).
LEC1 is sufficient to induce somatic embryogenesis [21],[22] and development of embryogenic characteristics in pkl pkr2 roots is likely a consequence of LEC1 de-repression due to reduced PcG protein activity. LEC1 activates FUS3 [26],[39], suggesting that increased FUS3 expression in pkl pkr2 is a consequence of increased LEC1 expression. FUS3 expression increases in pkl pkr2 despite no detectable decrease in H3K27me3 levels at FUS3; this is consistent with previous observations that H3K27me3 is not sufficient for gene silencing [4]. Finally, we conclude that PKL restricts embryogenic potential by regulating expression of genes for PcG proteins that are needed to repress activators of embryonic cell fate such as LEC1.
Taken together, our study revealed that the plant CHD3 proteins PKL/PKR2 directly activate PcG protein target genes; thus, PKL/PKR2 have trxG-like functions and counteract PcG protein repressive activities during development. In the future, it will be important to find out how CHD3 proteins and PcG proteins target the same genes and why at certain loci repression dominates and at other loci activation dominates.
All Arabidopsis thaliana mutants used in this study are in the Columbia accession. The pkl mutant used in this study was the pkl-1 allele described by Ogas et al. (1997). Mutant alleles clf-29 and swn-3 were described previously [17],[18]. pkr1-1, pkr2-1 and pkr2-2 correspond to WiscDsLox407C12, SALK 109423 and SALK 115303 respectively [40]. Single, double and triple homozygous mutant plants were characterized by PCR (for primers see Table S2). Seeds were surface sterilized (5% sodium hypochlorite, 0.1% Tween-20) and plated on MS medium (MS salts, 1% sucrose, pH 5.6, 0.8% bactoagar). After stratification for one day at 4°C, plants were grown in a growth cabinett under a long day photoperiod (16 h light and 8 h dark) at 23°C. For monitoring pickle root development, plates were incubated in vertical position and the phenotype was scored after 7 days. Experiments were performed in triplicates (three plates per experiment) and each experiment was performed at least three times. 10 day old seedlings were transferred to soil and plants were grown in a growth room at 60% humidity and daily cycles of 16 h light at 23°C and 8 h darkness at 18°C.
Root tips of five-day-old seedlings were harvested and total RNA was extracted using the RNeasy kit (Qiagen). For quantitative RT-PCR, RNA was treated with DNaseI and reverse transcribed using the First strand cDNA synthesis kit (Fermentas). For transcript analysis of aerial tissues, RNA was extracted using Trizol reagent (Invitrogen) and cDNA was synthesized as described above. Gene-specific primers and SYBR green JumpStart TaqReadyMix (Sigma-Aldrich) were used on a 7500 Fast Real-Time PCR system (Applied Biosystems). PP2A was used as reference gene. For sequences of primers see Table S2. Quantitative RT-PCR was performed using three replicates and results were analyzed as described [41].
Anti-PKL antibodies were generated against the C-terminus of the PKL protein (amino acids 1111–1384) by immunizing rabbits with the purified protein. For analysis of PKL protein in wild-type and pkl-1 mutant plants, rosette leaves were ground in liquid nitrogen and incubated in 2×urea sample buffer (65 mM Tris, pH 6.8, 5% β-mercaptoethanol, 2% SDS, 10% glycerin, 0.25% bromphenol blue, 8 M urea) for 5 min at 70°C. After centrifugation, the protein samples were loaded on a SDS-polyacrylamide gel and analyzed on immunoblots using antibodies against PKL. Equal loading and transfer of proteins was verified by staining the membrane in Ponceau Red solution (0.1% Ponceau S, 5% acetic acid). For analysis of H3 and H3K27me3, nuclear proteins from five-day-old seedling roots were extracted as described previously [42]. Protein blots were first probed with anti-H3K27me3 (Millipore, cat. 07-449) followed by anti-H3 antibodies (Millipore, cat. 07-690).
Root tips of five-day-old seedlings or aerial parts of ten-day-old seedlings were harvested and proteins were crosslinked in 10 mM dimethyladipimate for 20 min. After washing with distilled water proteins were crosslinked to DNA with 1% formaldehyde for 15 min. ChIP was performed as previously described [20] using antibodies against histone H3 (Millipore, cat. 07-690), H3K27me3 (Millipore, cat. 07-449) and rabbit IgG (Santa Cruz Biotechnology, cat. sc-2027). All tested regions were within 400 bp upstream of the start ATG. For sequences of primers see Table S2. PCR products were analyzed by agarose gel electrophoresis and quantified using ImageJ (http://rsbweb.nih.gov/ij/). Three PCR reactions were used for quantification and results presented as percent of input. Alternatively, gene-specific primers and SYBR green JumpStart TaqReadyMix (Sigma-Aldrich) were used on a 7500 Fast Real-Time PCR system (Applied Biosystems). Quantitative ChIP PCR was performed with four replicates and results were analyzed as described and presented as percent of input [41]. All ChIP experiments were performed at least twice.
Whole seedlings were incubated for 1 h in filtered Fat red solution (0.5% Fat Red Bluish in 60% isopropanol), washed three times with water and inspected.
|
10.1371/journal.ppat.1006418 | Structural and functional dissection reveals distinct roles of Ca2+-binding sites in the giant adhesin SiiE of Salmonella enterica | The giant non-fimbrial adhesin SiiE of Salmonella enterica mediates the first contact to the apical site of epithelial cells and enables subsequent invasion. SiiE is a 595 kDa protein composed of 53 repetitive bacterial immunoglobulin (BIg) domains and the only known substrate of the SPI4-encoded type 1 secretion system (T1SS). The crystal structure of BIg50-52 of SiiE revealed two distinct Ca2+-binding sites per BIg domain formed by conserved aspartate or glutamate residues. In a mutational analysis Ca2+-binding sites were disrupted by aspartate to serine exchange at various positions in the BIg domains of SiiE. Amounts of secreted SiiE diminish with a decreasing number of intact Ca2+-binding sites. BIg domains of SiiE contain distinct Ca2+-binding sites, with type I sites being similar to other T1SS-secreted proteins and type II sites newly identified in SiiE. We functionally and structurally dissected the roles of type I and type II Ca2+-binding sites in SiiE, as well as the importance of Ca2+-binding sites in various positions of SiiE. Type I Ca2+-binding sites were critical for efficient secretion of SiiE and a decreasing number of type I sites correlated with reduced secretion. Type II sites were less important for secretion, stability and surface expression of SiiE, however integrity of type II sites in the C-terminal portion was required for the function of SiiE in mediating adhesion and invasion.
| The interaction of Salmonella enterica with polarized epithelial cells depends on the function of SiiE, a 595 kDa adhesin containing 53 repeats of a bacterial immunoglobulin (BIg) domain. SiiE is secreted and surface-expressed by a cognate type I secretion system (T1SS). We found that BIg domains contain amino acid (aa) residues forming binding sites for Ca2+ ions. Two types of Ca2+-binding sites can be distinguished, termed type I and type II sites. We performed a structural and functional dissection of Ca2+-binding sites of SiiE. After mutation of aa residues forming type I and/or type II Ca2+-binding sites, we investigated the secretion, surface expression and function as adhesin for interaction with polarized epithelial cells of the SiiE variants. We found that Ca2+-binding sites are critical for supporting the secretion of SiiE. Integrity of type I sites in any position of SiiE is essential for efficient secretion and surface expression. In contrast integrity of type II sites is less important for secretion. However, loss of type II in the C-terminal, most distal portion of SiiE ablated SiiE-mediated adhesion, while loss of the type II sites in middle or N-terminal portions of SiiE had less or no effect on SiiE function. We propose a novel mechanism of Ca2+-dependent secretion and conformational fine tuning of SiiE as a large T1SS substrate with a central role in the interaction of S. enterica with host cells.
| Salmonella enterica is a food-borne Gram-negative pathogen which causes self-limiting gastroenteritis. To survive inside the host, Salmonella possesses sophisticated virulence factors and protein secretion systems [1]. A Salmonella pathogenicity island (SPI) 1-encoded type 3 secretion system (T3SS) is necessary for invasion [2]. This protein secretion system is capable to secrete a distinct cocktail of effector proteins, which manipulate the host cell. In order to establish the initial contact to the apical side of polarized epithelial cells and to enable translocation by the SPI1-T3SS, Salmonella deploys the SPI4-encoded T1SS and the giant non-fimbrial adhesin SiiE [3].
The SPI4 locus encodes SiiE and the T1SS for secretion of SiiE, with SiiF being the inner membrane transport ATPase, SiiD acting as periplasmic adaptor protein (PAP), and outer membrane secretin SiiC [4]. SiiE, a 595 kDa non-fimbrial adhesin, is the only known substrate for the SPI4-T1SS [4]. SiiE mediates the first intimate contact to the host cell through binding to glycostructures containing N-acetyl-glucosamine and/or α2,3-linked sialic acid [5]. This contact positions the SPI1-T3SS to efficiently translocate effector proteins which lead to actin remodeling and macropinocytosis of the bacteria. As a T1SS substrate protein, SiiE possesses a C-terminal secretion signal [4]. The adhesin is transiently retained within the secretion system and at later time points present in the supernatant [6]. The two accessory proteins SiiA and SiiB are located in the inner membrane presumably forming a proton-conductive channel. This channel may use the proton motive force (PMF) at the cytoplasmic membrane to regulate the retention of SiiE, either through sensing the physiological state of the cell or by inducing conformational changes to binding partners [7].
The adhesin SiiE is composed of an N-terminal domain containing β-sheet and coiled-coil repeats, followed by 53 repeats of bacterial immunoglobulin (BIg) domains [6]. BIg52 and BIg53 are separated by a putatively unfolded element termed insertion. Sequence alignments of all 53 BIg domains revealed that a prototypical BIg domain possesses 6 conserved aspartate (D) or glutamate (E) residues, of which 5 form two binding sites for Ca2+ ions. Recently we solved the crystal structure of SiiE BIg50-52. Despite not having explicitly added any Ca2+ ions during protein production, the crystal structure revealed that up to two Ca2+ ions are bound per BIg domain in SiiE [8]. The SiiE-wide conservation of D and E residues that are involved in Ca2+ binding suggests that SiiE binds about 100 Ca2+ ions per molecule [8]. Ultrastructural analysis showed that chelation of Ca2+ ions of purified secreted SiiE molecules distorts the linear rod-like structure of SiiE, indicating a stabilizing effect of Ca2+ ions. Ca2+ binding has been demonstrated for other T1SS substrate proteins, such as adhesins, the antifreeze protein of Marinomonas primoryensis (MpAFP) [9] or SpaA of Corynebacterium diphtheria [10]. Repeat in Toxin (RTX) proteins are a family of T1SS-secreted toxins and E. coli HlyA und Bordetella pertussis adenylate cyclase CyaA are well studied RTX toxins. The RTX motif is a glycine-rich nonapeptide involved in Ca2+ binding. For both HlyA and CyaA, binding of Ca2+ ions was shown to support the secretion by the T1SS [11–14].
The BIg domains of SiiE possess two distinct types of Ca2+-binding sites that are distinct from the RTX motif. The conserved D residues are numbered according to the multi-sequence alignment of SiiE BIg domains shown by Griessl et al. [8]. Type I Ca2+-binding sites are positioned at the interface of two BIg domains and contain three D residues, namely BIgn-1117D and BIgn43D and 97D. Type I Ca2+-binding sites are frequently found in BIg domain proteins. In contrast, type II Ca2+-binding sites are specific to SiiE and built by two D residues within one BIg domain (BIgn16D and 24D). Please note that each Ca2+ ion is coordinated by 6 ligands and therefore other residues and water molecules are involved as well in Ca2+ ion binding, in addition to the conserved D residues [8]. We also observed conserved tryptophan residues, i.e. 74W, in most of the BIg domains. This residue is distal to the Ca2+-binding sites, but may be involved in interaction of SiiE with glycostructures as observed for transport proteins [15] or fimbrial adhesins [16].
The roles of the distinct type I and type II Ca2+-binding sites for secretion of SiiE and function as adhesin are not known. Here, we report the functional dissection of SiiE Ca2+-binding sites. We found that with an increasing number of conserved D residues exchanged to the non-charged amino acid serine (S), the amounts of secreted SiiE were dramatically decreased. Exchanges of single D residues or of either single type I or type II Ca2+-binding sites showed no effect, while exchange of multiple type I or type II Ca2+-binding sites showed a more dramatic effect when type I Ca2+-binding sites are missing. Our data demonstrate a critical role of type I sites to support transport of SiiE through the T1SS, while type II sites are important to structure secreted SiiE and to maintain a BIg domain conformation that enables interaction with cognate ligands on the host cell surface.
The giant adhesin SiiE possesses 53 BIg domains, most of which contain five conserved D or E residues that coordinate binding of two Ca2+ ions. We investigated the role of Ca2 +-binding sites in BIg domains for secretion of SiiE. The Gaussia luciferase (GLuc) [17] was used as reporter for quantification of amounts of secreted SiiE (Fig 1, S1 Fig). Compared to Firefly luciferase, GLuc is ATP independent, and more robust and progressive [18]. To quantify the secretion, the reporter GLuc BIg50-53 was constructed by fusing GLuc to the C-terminal moiety of SiiE, i.e. BIg domains 50–53, the insertion and the C-terminal secretion signal (Fig 1A). Another construct was generated in which all five conserved D residues forming the type I and type II Ca2+-binding sites in BIg51 and BIg52 were exchanged to S (D/S exchange), termed GLuc BIg50-53Δ2. The number of deleted Ca2+-binding sites is indicated by Δn. A further reporter fusion consisting of GLuc and BIg47-53 was generated and D/S exchanges of various extent were introduced (S1C Fig).
The synthesis and secretion of GLuc fusions with WT or mutant SiiE portions was compared using Salmonella WT and the siiF-deficient strain unable to form a functional T1SS. GLuc activities in the lysate and supernatant obtained after 6 h of subculture represent the cytosolic and surface-bound, or secreted portion of GLuc fusions, respectively (S1 Fig, Fig 1B). GLuc activities in lysates were similar for GLuc BIg50-53 and GLuc BIg50-53Δ2 reporters, indicating similar rates of synthesis and stability (Fig 1B). Secreted GLuc activity for the SiiE WT reporter was 46.7-fold lower in the ΔsiiF background, demonstrating SPI4-T1SS-dependent secretion of the reporter. In the SPI4-T1SS-proficient background, secreted GLuc activity for the GLuc BIg50-53Δ2 reporter was 122.4-fold lower than for the GLuc BIg50-53 reporter.
We also analyzed secretion of a GLuc fusion protein containing BIg47-53 (S1C and S1D Fig). Here, five D/S exchanges in GLuc BIg47-53Δ2 did not reduce secretion. Also, the exchanges resulting in GLuc BIg47-53Δ4 were without effect on the secretion of the reporter, while additional D/S exchanges for deletion of another two Ca2+-binding sites caused a five-fold reduced secretion of GLuc BIg47-53Δ6. The complete removal of 10 Ca2+-binding sites in GLuc BIg47-53Δ10 resulted in 73.7-fold reduced secretion, similar to levels of the WT reporter fusion in ΔsiiF background.
We conclude that Ca2+-binding sites in BIg SiiE are required for the secretion of SiiE. The removal of two Ca2+-binding sites in a secretion reporter with four BIg was sufficient to ablate T1SS-dependent secretion, while removal of at least 6 Ca2+-binding sites was required to affect secretion of a reporter fusion with 7 BIg.
To analyze the role of Ca2+-binding sites in SiiE for SiiE-dependent virulence functions of Salmonella, we transferred mutant alleles with D/S exchanges of various extent into chromosomal siiE using λ Red recombineering (Fig 2A). The resulting strains synthesized mutant forms of SiiE with D/S exchanges resulting in removal of 2, 5, 6 or 10 Ca2+-binding sites (Fig 2B). No SiiE was detected for the ΔsiiE strain and compared to WT SiiE, variable amounts of mutant SiiE were observed. Compared to the WT, all mutant strains investigated contain lower amount of cell-associated SiiE. Since whole bacterial lysates were analyzed, one cannot distinguish between SiiE present in cytosol, or SiiE retained on the bacterial surface. Mutant forms of SiiE with reduced surface retention will lead to lower amounts of cell-associated SiiE, although levels of SiiE synthesis are comparable to WT SiiE.
SiiE retention on the bacterial surface and secretion into culture supernatant was analyzed by dot blots of whole cells, and protein precipitated from culture supernatants, respectively (Fig 2C and 2D). Our previous work demonstrated that SiiE is mainly retained on the bacterial surface at 3.5 h of subculture, and predominantly released into the supernatant at 6 h and later of subculture [6]. Of the various mutant strains analyzed, only the strain producing SiiE BIg52Δ2 showed SiiE retention after 3.5 h of subculture similar to WT. After 6 h subculture levels of SiiE retention of WT and all mutant strains were as low as the negative control.
The number of mutated Ca2+-binding sites correlated with reduction of secreted SiiE at 3.5 and 6 h of subculture. With increasing numbers of D/S exchanges, lesser amounts of secreted SiiE were detected. Amounts of SiiE BIg47-52Δ10 were as low as the negative control, indicating a complete loss of secretion for this SiiE mutant (Fig 2D).
We compared secretion of WT and mutant SiiE quantified by dot blot analyses to GLuc activities of the GLuc-SiiE reporter (Fig 2E, S1C Fig). GLuc reporters for SiiE BIg52Δ2 and SiiE BIg47Δ1 BIg51-52Δ4 resulted in GLuc activities similar to GLuc-SiiEWT. If introduced in chromosomal siiE, the mutations resulted in reduced amounts of secreted SiiE. Secretion of SiiE BIg47Δ2 BIg51-52Δ4 was highly reduced in both assays, while no secretion of SiiE BIg47-52Δ10 was detected in GLuc and dot blot assays. The data demonstrate that Ca2+-binding sites in BIg are important for the secretion of SiiE, and that amounts of secreted SiiE decreases with an increasing number of D/S exchanges in BIg domains.
We next determined the effect of deletion of Ca2+-binding sites on SiiE-dependent virulence functions, i.e. adhesion to polarized epithelial cell followed by SPI1-T3SS-mediated invasion (Fig 2F). Only SiiE BIg52Δ2 conferred invasion of MDCK cells at a level comparable to WT SiiE. All other mutant SiiE we investigated resulted in highly reduced invasion, comparable to the siiF-deficient strain that is unable to secrete SiiE.
We conclude that Ca2+-binding sites in the C-terminal part of SiiE are essential for secretion and function of the adhesin. Removal of Ca2+-binding sites in more than one BIg in this moiety results in loss of function.
The C-terminal moiety of SiiE contains the signal for T1SS secretion, is secreted first and is likely to be exposed most distal to the cell envelope. We next tested the functional relevance of Ca2+-binding sites in the middle or N-terminal portions of SiiE. Strains were generated with mutations in chromosomal siiE resulting in deletion of Ca2+-binding sites in BIg2, BIg40, or BIg1-5 (Fig 3A). Since 117D of a previous BIg domain (BIg(n-1)) forms a type I Ca2+-binding site with 43D and 97D of a subsequent BIg domain (BIg(n)), 117D of BIg1 and BIg39 were exchanged instead of 117D of BIg2 and BIg40, resulting in SiiE BIg2Δ2 and SiiE BIg40Δ2, respectively. To control the precision of the Red recombineering method applied here and the absence of unwanted attenuating mutations, we used a siiE mutant strain and restored the WT sequence. This strain, termed WTrestored, showed SiiE-dependent phenotypes as the WT strain.
Strains expressing mutant chromosomal siiE were tested for protein synthesis. No expression could be detected for the negative control ΔsiiE. All mutant strains synthesized SiiE of correct size (Fig 3B). The deletion of Ca2+-binding sites only in BIg2 or BIg40 had no or only small effects on surface retention of SiiE (Fig 3C), secretion (Fig 3D), or SiiE-dependent invasion (Fig 3E). Secretion of SiiE BIg1-5Δ10 was slightly reduced, but there was still more secretion after 6 h of subculture than after 3.5 h of subculture. Interestingly, the level of SiiE retention was also reduced to approximately 50% of WT and maintained at the same level at 6 h of subculture. Destruction of all Ca2+-binding sites in BIg1-5 (BIg1-5Δ10) led to a 74.6-fold decreased invasion of polarized cells (Fig 3E), while the same extent of deletions in BIg47-52 (BIg47-52Δ10) resulted in 392.2-fold reduced invasion (Fig 2F). Compared to SiiE BIg47-52Δ10, reduction of retention and secretion is less pronounced for SiiE BIg1-5Δ10. If extracellular Ca2+ ions facilitate secretion of SiiE, the secretion might come to a halt earlier for BIg47-52Δ10 than for SiiE BIg1-5Δ10.
In addition to the conserved D or E residues involved in Ca2+ binding, 47 of 53 BIg domains possess a conserved tryptophan residue at position 74. To test a potential role of these conserved tryptophan residues in SiiE function, we performed W to F (W/F) exchanges in 1, 2, or 3 BIg in the C-terminal moiety of SiiE (S2A Fig). These mutations only resulted in minor changes of the amounts of SiiE retained and secreted at 3.5 h or 6 h of subculture (S2B, S2C and S2D Fig). Functionally, none of the mutant forms of SiiE with various degrees of W/F exchanges resulted in reduced invasion of polarized epithelial cells (S2E Fig), indicating that conserved 74W residues in the C-terminal moiety of SiiE are neither important for secretion and retention of SiiE, nor for the SiiE-dependent adhesion and invasion.
T1SS substrate proteins are secreted in an unfolded state [19]. For example, secretion of E. coli HlyA was highly reduced if the protein was modified in a way that allowed folding in the cytosol [20]. To further investigate parameters known to affect secretion of T1SS substrate proteins, we tested if folding rate influences secretion as for HlyA. We fused the C-terminal portion of SiiE harboring the secretion signal to MalE. Distinct point mutations in the MalE portion of the fusion protein led to different folding rates as previous established by Bakkes et al. [20]. Secretion of various MalE-SiiE fusion proteins was analyzed at 3.5 h and 6 h of subculture. Similar amounts of fusion proteins were detected in the culture supernatant at both time points (S3 Fig). This indicates that the rate of intracellular folding did not affect SiiE secretion, supporting that binding of extracellular Ca2+ ions by the secreted portion of BIg domains is more important.
In order to assess the role of Ca2+ ions for the conformational stability, molecular dynamics (MD) simulations of WT and mutant SiiE were performed. We focused on BIg domains 50–52 for the following reasons: (i) a high-resolution crystal structure is available for this portion of SiiE, (ii) the fragment is sufficiently small to allow for extensive MD simulations, and (iii) the role of Ca2+-binding sites in BIg50-52 constructs was experimentally investigated (Fig 1). The role of Ca2+ ions was assessed from inspection of the tilt and twist angles defining the relative orientation of BIg 51 and 52 (Fig 4A–4D). The percentage of tilted or twisted structures detected over the simulation time is summarized in Fig 4E. A comparison of WT and mutant forms revealed that mutation of both type I and type II sites caused an enhanced tilting of the structure and thus a less extended domain arrangement compared to the WT structure. The stronger effect was observed for the type I site, which is consistent with the direct location in the domain interface. Notably, the concomitant mutation of both sites had an additive effect resulting in the highest portion of tilted structures among all systems investigated. Mutation of type I and type II site did not only affect tilting, but also resulted in an enhanced twisting of the domain pair (Fig 4E). However, in contrast to tilting, type I and type II site had a similar effect on domain twisting and there was no additive effect upon mutation of both sites. These data support the role of Ca2+-binding sites in increasing the rigidity of SiiE.
To assess the relative Ca2+ binding affinities of the type I and type II sites, steered molecular dynamics (SMD) simulations were performed. The setup is schematically depicted in Fig 5A and 5B. The Ca2+ ions were independently removed from both sites and 10 simulations were performed for each site. For the type II site, all 10 work plots display an overall similar shape (Fig 5C). Up to a distance of ~ 15 Å, the work linearly increases reflecting the disruption of the interactions between the Ca2+ ion and its protein ligands. For larger distances there is only a marginal further increase of the work, indicating that dissociation is almost complete at a distance of ~ 15 Å. The work required in the 10 SMD runs of the type II site ranges from 246–329 kcal x mol-1. For the type I sites, the qualitative appearance of the curves is similar (Fig 5D); however, less work is required for the removal of the ion from this site. The resulting work ranges from 134–212 kcal x mol-1, which is significantly lower than for the type II site.
SiiE variants with altered type I and type II Ca2+-binding sites in the C-terminal moiety of SiiE (SiiECterm) were characterized in detail in in vitro experiments. These variants encompass BIg domains 48–53, the insertion and the C-terminal segment that includes the secretion signal (Fig 6A). We have previously shown that recombinant protein production of a fragment that covers BIg domains 50 to 52 in E. coli and subsequent purification of the fragment without the addition of any Ca2+ ions leads to a protein sample in which all Ca2+-binding sites are fully occupied [8]. This observation was corroborated not only by the final electron density map of the solved crystal structure, but also by analyzing in detail the anomalous scattering signal and via X-ray fluorescence measurements [8]. Here, we now used inductively coupled plasma—atom emission spectroscopy (ICP-AES) to verify the presence and/or absence of Ca2+ ions in protein variants that were produced following a similar purification protocol as previously described for the BIg domains 50 to 52 protein fragment [8]. In case of WT SiiECterm the occurrence of 8 to 10 Ca2+-binding sites was expected. In the ICP-AES experiment the protein was analysed at a concentration of 6.5 mg x ml-1 (88.1 μM), which results in a theoretical maximum calcium content of 28–35 μg x ml-1 (705.2–881.5 μM). The experimentally determined calcium concentration was 23 μg x ml-1 which amounts to 81.8% to 65.44% of the expected theoretically maximum content (Table 1, S4 Fig). Variant SiiECterm with type I and type II Ca2+-binding sites mutated (SiiECterm BIg48-52Δ8type I + II) was measured at a concentration of 4.2 mg x ml-1 (57.63 μM). For this variant, a calcium signal below the 0.5 μg x ml-1 calibration standard and outside the calibration range was detected (S4 Fig). Thus, no calcium binding is observed for variant SiiECterm BIg48-52Δ8type I + II (Table 1). These measurements show that in case of the WT protein the Ca2+-binding sites are almost fully occupied in the recombinantly produced protein sample whereas the substitution of defined aspartic residues against serines in the SiiECterm BIg48-52Δ8type I + II results in a protein that is devoid of any calcium binding. Conversely, variants with some of the Ca2+-binding sites disrupted should display reduced Ca2+ binding if one assumes the absence of any cooperativity between the binding sites.
To experimentally address individual roles of type I and type II Ca2+-binding sites for the conformational stability of SiiE, we performed circular dichroism (CD) measurements. Highly similar FarUV spectra were recorded for all variants, namely the SiiECterm construct with WT sequence, mutations of type I, type II and of both type I and type II Ca2+-binding sites (Fig 6B). Estimation of the secondary structure content using the BeStSel server suggests that the SiiECterm wild-type variant consist of around 50% β-sheets, 10% turns and 40% others (e.g. random coil). These values are close to those derived from the available SiiE BIg-domain crystal structure (BIg domains 50–52, PDB entry: 2YN5). This suggests that the so far structurally uncharacterized domains (BIg domains 48–49 and BIg domain 53) display similar secondary structures as domains 50 to 52. Most importantly, however, the highly similar spectra and concomitant results from the secondary structure analysis of the variants studied here demonstrate that the secondary structure composition of the proteins is not altered, thus excluding pronounced mis- or unfolding, when mutating the type I and/or type II Ca2+-binding sites (S4 Table). The thermal scanning CD measurements revealed distinct effects of type I and type II site mutations on the conformational stability of SiiECterm (Fig 6C). All protein variants exhibit a decrease in ellipticity upon heating, which suggests that instead of a thermally induced unfolding an increased secondary structure and/or β-sheet formation via aggregation occurs. However, the magnitude of this transition is lower for SiiECterm WT and SiiECterm BIg48-52Δ4type II than for SiiECterm BIg48-52Δ4type I and SiiECterm BIg48-52Δ8type I + II. Also, Tonset of these structural changes is higher for WT and Δ4type II variants at 72°C and 69°C, respectively, than for Δ4type I and Δ8type I + II variants at 50°C and 45°C, respectively. Thus, while the thermal scanning CD curve of SiiECterm BIg48-52Δ4type II resembles that of SiiECterm WT, the spectra of proteins with mutations of type I and both type I and II sites indicated that the proteins are more prone to aggregation.
To further investigate the conformational stability, the SiiECterm variants were subjected to native PAGE (Fig 6D and 6E). All variants covering the C-terminal moiety of SiiE migrated as a single band under mild conditions (Fig 6D). To further investigate the aggregation behavior, individual samples of each variant were incubated at increasing temperatures and subsequently analyzed by native PAGE. SiiECterm WT was resistant to aggregation up to 80°C and SiiECterm BIg48-52Δ4type II behaved similar to WT protein, although pronounced aggregation was detected at a slightly lower temperature (Fig 6E). A ladder-like pattern indicates that SiiECterm BIg48-52Δ8type I + II started to form oligomers and aggregated already at 50°C. Aggregation of SiiECterm BIg48-52Δ4type I resembled SiiECterm BIg48-52Δ8type I + II, although slightly delayed (Fig 6D). The analysis of the aggregation behavior therefore reflects the results of the thermal scanning CD measurements. Faster migrating protein species are visible in the native PAGE of the Δ4type I and Δ8type I + II mutants. To control whether unwanted proteolysis might have caused the occurrence of these so-called lower bands, SDS-PAGE analysis was performed for SiiECterm samples after incubation at various temperatures (S5 Fig). Neither WT SiiECterm nor any of the mutant forms indicate a temperature-dependent occurrence of proteolytic fragments. The increased migration behavior thus results from a partial collapse of the expected linear overall structure of SiiE into a more globular domain arrangement as the result of the removal of the type I Ca2+-binding sites that are located in the interface between BIg domains.
Next, the conformation and compactness of the SiiECterm variants was probed by limited proteolysis using α-Chymotrypsin and Proteinase K (S6 Fig). Multi-domain proteins with flexible domain surface loops and/or interdomain linkers are expected to be more prone to proteolytic cleavage than proteins with very rigid domain architecture. Resistance against proteolytic cleavage by α-Chymotrypsin was clearly reduced for SiiECterm BIg48-52Δ8type I + II in comparison to WT protein or protein with only type I or type II binding site mutations. This suggests that both types of Ca2+-binding sites help to stabilize the fold against proteolytic degradation. The effect is possibly enhanced by the close spatial proximity of the two binding sites [8]. Resistance against proteolytic cleavage by Proteinase K was also reduced most for SiiECterm BIg48-52Δ8type I + II similarly to the proteolysis using α-Chymotrypsin. However, the stability of the individual type I or type II Ca2+-binding site mutants was also decreased, although to a lesser extent than for the variant with both binding sites mutated.
Finally, we set out to functionally dissect the roles of type I and type II Ca2+-binding sites in SiiE. We generated mutant alleles by site-directed mutagenesis for single aa exchanges in BIg51 and BIg52 (S7 Fig), or exchanges of all residues of either the type I or the type II Ca2+-binding sites within BIg52, within BIg47-52, or BIg1-5 (Fig 7).
Mutation of single aspartate residues or D/S exchanges in single Ca2+-binding sites in chromosomal siiE did not affect SiiE synthesis, surface retention, secretion, or SiiE-dependent invasion (S7B, S7C, S7D and S7E Fig). In contrast, if either type I or type II Ca2+-binding sites are missing within the five C-terminal BIg domains 47–52, invasion is reduced to the level of the negative control (Fig 7E). For both mutants, retention (Fig 7C) and secretion (Fig 7D) was reduced. This reduction was more pronounced for SiiE BIg47-52Δ5type I than for SiiE BIg47-52Δ5type II. Retention was fully abolished for SiiE BIg47-52Δ5type I, as well as for BIg47-52Δ10 for all time points tested, while secretion was reduced to 50% of WT SiiE. For SiiE BIg47-52Δ5type II retention after 3.5 h of subculture was reduced to 40% of WT SiiE and after 6 h and later time points no surface retention was detected, similar to WT SiiE.
Removal of type I or type II Ca2+-binding sites in BIg1-5 had only minor effects on retention and secretion of SiiE. In contrast to SiiE BIg1-5Δ5type II, SiiE BIg1-5Δ5type I was retained at late time points at a level similar to BIg1-5Δ10 (8 and 24 h). SiiE BIg1-5Δ5type II was also retained at late time points, but to a lesser extent than SiiE BIg1-5Δ5type I or SiiE BIg1-5Δ10. The mutation of five type I sites in BIg1-5 resulted in highly (43.5-fold) reduced invasion, while removal of five type II Ca2+-binding sites in the same moiety (BIg1-5Δ5type II) did not reduce invasion of polarized epithelial cells (Fig 7E). These results suggest distinct roles of type II Ca2+-binding sites in N- and C-terminal portions of SiiE.
To further analyze the role of type II Ca2+-binding sites for function of SiiE, we removed type II Ca2+-binding sites by D/S exchanges in BIg31-35, BIg 36–40, BIg 41–45, or BIg 46–50. The invasion of MDCK cells of strains expressing these mutant forms of siiE was compared to invasion by strains with WT SiiE, SiiE BIg1-5Δ5type II and SiiE BIg47-52Δ5type II (Fig 8). We observed that strains producing SiiE with type II sites removed in BIg31-35, BIg 36–40, BIg 41–45, BIg 46–50 or BIg47-52 all exhibited reduced invasion compared to strains with SiiE WT or SiiE BIg1-5Δ5type II. The reduction of invasion was pronounced if BIg domains in the C-terminal region were affected, and smallest reduction of invasion was observed for the strain with SiiE BIg30-35Δ5type II.
The fact that SiiE BIg1-5Δ5type II, but not SiiE BIg47-52Δ5type II still mediates binding to, and invasion of MDCK cells indicates that type II Ca2+-binding sites are more important for the correct local conformation of the protein, which might be necessary for proper binding. We conclude that in SiiE BIg1-5Δ5type II the C-terminal part is correctly folded and can mediate binding to the host cell, while this is not the case for SiiE BIg47-52Δ5type II.
Our comprehensive mutational and functional analyses revealed a role of conserved aspartate residues in BIg domains of SiiE in secretion and adhesin function of this giant adhesin. An increasing number of exchanges of conserved aspartate residues resulted in decreased secretion of SiiE. This observation was made for a C-terminal plasmid-encoded portion of SiiE, as well as for chromosomally encoded variants of SiiE with the same amino acid exchanges. For chromosomally encoded SiiE BIg52Δ2, the secretion was reduced. This reduced amount of secreted SiiE was not associated with higher levels of SiiE retention, or reduced SiiE-dependent invasion. Also, 5 D/S exchanges in BIg2Δ2 or BIg40Δ2 in chromosomally encoded SiiE did not lead to reduced invasion. Based on these results we conclude that the number of functional SiiE molecules on the bacterial surface is still high enough to mediate apical adhesion and subsequent invasion.
We found that deletion of two Ca2+-binding sites by 5 D/S exchanges in the N- or C-terminal parts of SiiE showed no or only mild phenotypic difference, indicating that the remaining Ca2+-binding sites in adjacent domains can compensate for a certain degree of loss of Ca2+-binding properties of SiiE. Upon deletion of 5 or more Ca2+-binding sites, we observed loss of SiiE retention, dramatically decreased amounts of secreted SiiE and attenuated invasion. Thus, the lack of 5 Ca2+-binding sites in the C-terminal portion of SiiE could not be compensated by the function of residual domains. We propose a model in which binding of extracellular Ca2+ ions promotes directionality in the secretion of SiiE (Fig 9A). If many consecutive D residues are missing, Ca2+ binding is ablated and the secretion is reduced. Lack of a few Ca2+-binding sites is not critical since Ca2+ ions will bind to the next available Ca2+-binding site of BIg domains that are already outside of the T1SS. If too many Ca2+-binding sites are missing, the next available Ca2+-binding sites are still within the channel of the T1SS and not accessible for the extracellular Ca2+ ions (Fig 9A). Dependent on the position of the missing Ca2+-binding sites within SiiE, secretion of SiiE is arrested at a certain stage. If Ca2+-binding sites were removed in BIg1-5, SiiE was also surface-retained at 6 h of subculture and later, indicating that the secretion process stopped. Exchanges in the C-terminal moiety, namely BIg47-52Δ10 led to secretion stalling early in the process of secretion, so that surface expressed or secreted SiiE was highly reduced.
If fusion proteins covering a short portion of SiiE are investigated (Fig 1), lack of already a few Ca2+-binding sites leads to a dramatic decrease in secreted SiiE. Possibly, not only the number of lacking Ca2+-binding sites is responsible for this, but also the overall number of available Ca2+-binding sites.
Thomas et al. [13] recently described a Ca2+ driven folding that may facilitate secretion in E. coli pro-HlyA. HlyA contains RTX motifs that bind Ca2+. If no Ca2+ is bound, or if Ca2+ is chelated by e.g. EDTA, the protein remains unfolded [13, 21, 22]. A similar mechanism could be considered for secretion of SiiE, with binding of extracellular Ca2+ ions initially facilitating secretion and later supporting the proper conformation and interaction with ligands. Such an interaction could, for example, occur with the carboxyl group of SiiE ligand α2,3-linked sialic acid.
From other T1SS substrates it is known that they are secreted in an unfolded state. For SiiE, the intracellular folding rate is not the most critical parameter for the initialization of the secretion process, as seen for the E. coli HlyA system [13] (S3 Fig). Since the SPI4-encoded T1SS possesses two unique accessory proteins SiiA and SiiB, these proteins could promote the start of SiiE release until the protein becomes accessible to extracellular Ca2+ ions. SiiE has to be transiently retained on the bacterial surface in order to function as an adhesin. The proton channel formed by SiiA and SiiB or possibly other interactions between subunits of the T1SS could act as the retention signal, which has to be stronger than the extracellular Ca2+ ions. More experiments are needed to fully understand the retention process of this exceptional adhesin.
Why does SiiE possess two distinct types of Ca2+-binding sites? Conserved D residues forming type I Ca2+-binding sites can be found in other bacterial adhesins or secreted enzymes, like BapA from Salmonella enterica, LapF from Pseudomonas putida, or the PKD domain from the plant cell wall-targeting endoglucanase of Clostridium thermocellum [8, 23–25]. In contrast, type II Ca2+-binding sites are specific for SiiE. Since type I and type II Ca2+-binding sites are structurally different, also distinct functions have to be considered. We speculate that type II Ca2+-binding sites are required for proper fine-tuning of the conformation of SiiE, while type I Ca2+-binding sites promote an overall rigidification of the tertiary structure of SiiE and thereby secretion through the T1SS. To elucidate the impact of type I and type II Ca2+-binding sites, several binding sites within the C-terminal portion were exchanged. All recombinantly produced C-terminal variants showed similar CD spectra showing that the secondary structure content was not altered in these variants. However, compared to type II sites, removal of type I sites had more dramatic effects on secretion and retention, independent of the region in SiiE in which the mutations were introduced. This observation would be in line with a role of type I sites in supporting secretion (Fig 9B). Interestingly, mutations BIg47-52Δ5type I, and BIg1-5Δ5type I led to loss of function of SiiE in supporting invasion of polarized cells. Loss of SiiE function was observed for mutation BIg47-52Δ5type II, while mutation BIg1-5Δ5type II only caused minute alteration in invasion of MDCK cells. Thus, type II Ca2+-binding sites in the N-terminal region of SiiE are dispensable for secretion and surface expression of SiiE.
BIg domains of LapF and BapA are more similar to SiiE BIg50, harboring only a type I Ca2+-binding site [8]. Until now, no distinct role for BIg50 could be identified. Since it is much shorter than the other SiiE BIg domains, one Ca2+-binding site could be sufficient to stabilize this domain, whereas longer BIg domains need two Ca2+-binding sites to be stabilized.
We previously reported that Ca2+ ions stabilize a rigid, linear conformation of SiiE, suggesting a role of Ca2+-binding sites for the overall protein stability [8]. The MD simulations of the present study showed that mutations of type I as well as type II sites destabilize SiiE and cause more frequent deviations from the geometry of the crystal structure. In particular, changes of the tilt angle result in less extended conformations (Fig 4B) that are expected to exhibit a reduced SiiE functionality. It is also interesting to note that the majority of the tilt angles observed during simulation are in the range of 15° to 30°. These rather small tilt angles suggest that larger deviations from the extended geometry require kinking at multiple sites at the same time. This is in line with the experimental data that mutations of multiple type I sites are required in order to disturb SiiE function. The observation that tilting is predominantly enhanced by mutation of the type I site suggests that these sites might also be more critical for the structural rigidification of SiiE and is in line with the observation that disruption of the type I sites leads to an earlier onset of temperature-induced aggregation of SiiE. Although the work calculated by SMD simulations cannot readily be converted into binding affinities, it suggests that type II sites exhibit higher Ca2+ affinity compared to type I sites (Fig 5). This is in line with a sequential mechanism of Ca2+ binding during secretion, in which initially the type II sites get occupied (probably during the folding of the individual domains), and subsequently Ca2+ binding of the type I sites stabilizes the extended domain arrangement of SiiE.
We propose that local conformational distortions in the C-terminal BIg domains due to the BIg47-52Δ5type II mutations may result in loss of binding to host cell ligands, despite sufficient amounts of SiiE being surface expressed. For SiiE BIg1-5Δ5type II, the C-terminally located BIg domains are correctly folded and proficient in ligand binding. This model also implies that BIg domains located in the N-terminal and perhaps also in the central portion of SiiE are not directly involved in binding to ligands on apical membranes of host cells (Fig 9C, 9D and 9E). Although Ca2+ ions promote SiiE secretion and folding, we do not assume that they are also directly involved into binding to the target structure. SiiE binds to GlcNAc- and sialic acid-containing structures [5]. Griessl et al. [8] also showed that the majority of SiiE BIg domains possess a conserved tryptophan residue, however, exchanges of this aromatic residue in BIg50-52 did not influence SiiE specific phenotypes (S2 Fig). Future mutational and functional analyses may identify the residues in C-terminal BIg domains of SiiE that directly contribute to the interaction with ligands.
The 'C-terminus first secretion' of T1SS substrate proteins has been formally demonstrated for HlyA [26]. The C-terminal RTX domain of B. pertussis CyaA has 5 repeat blocks containing GGxGxDxxx motifs that form β-rolls coordinating Ca2+ binding. A total of 40 Ca2+ ions bound per CyaA molecule have been estimated [27]. Recent analyses of T1SS secretion of CyaA and related RTX toxins demonstrated a vectorial push-ratchet mechanism [12, 14]. In the bacterial cytosol, the RTX domain is intrinsically disordered. After exit of the T1SS duct, the β-rolls sequentially bind Ca2+ at the outside of the cell envelope, initiate folding of the C-terminal domain and thereby provide directionality and secretion support for the rest of the protein [28, 29]. The proposed model is in line with earlier findings, namely that ATP-hydrolysis and membrane potential are only required for the initiation of secretion, while further secretion is driven by the folding of portions of the substrate protein that have left the T1SS [11, 14]. Interestingly, a recent analysis revealed that HlyA secretion efficiency is independent from Ca2+ concentration ranging from 0 to 5 mM Ca2+ in the external medium [30]. This observation would argue against a role of Ca2+ in supporting T1SS secretion of HlyA, however, the local Ca2+ pool in the outer membrane of secreting bacteria also has to be considered.
Although Ca2+ binding by type I and type II Ca2+-binding sites in SiiE is mediated by structurally distinct motifs, we propose a function similar to RTX repeats regarding disordered-to-folded transition and directionality of secretion. While Ca2+-binding sites in CyaA are restricted to the C-terminal RTX domain, Ca2+-binding sites are present in all BIg domains of SiiE. This would be in line with a dual function of Ca2+ binding in promoting secretion, as well as stabilizing the ligand-binding competent overall conformation of SiiE.
Salmonella enterica serovar Typhimurium (S. Typhimurium) NCTC 12023 was used as wild-type strain in this study and all mutant strains are isogenic to this strain. The characteristics of strains used in this study are listed in Table 2. Bacterial strains were routinely grown in LB broth or on LB agar containing antibiotics if required for selection of specific markers. The Ca2+ concentration in LB media is not defined. Carbenicillin or kanamycin were used at 50 μg x ml-1, and tetracycline or chloramphenicol were added to a final concentration of 20 or 10 μg x ml-1 respectively, if required for the selection of phenotypes or maintenance of plasmids.
Madin-Darby Canine Kidney Epithelial (MDCK) cells are an immortalized cell line initially derived from renal tube of a cocker spaniel. MDCK Pf subclone used for the generation of polarized epithelial cell monolayers was kindly provided by Department of Nephroplogy, FAU Erlangen-Nürnberg. MDCK cells were used for the generation of polarized epithelial cell layer. Cell culture conditions were previously described by Wagner et al. [6]. Briefly, cells were cultures in MEM with Earle’s salts, 4 mM Glutamax, non-essential amino acids and 10% heat-inactivated fetal calf serum. The Ca2+ concentration in this medium is 1.8 mM.
The synthetic DNA fragments (GeneArt or IDT) were subcloned in blunt end restriction sites of the pJET1.2 vector backbone. The synthetic DNA fragment SiiE-BIg49half-52D-S was first cloned into pWRG454 (pWSK29::PsiiAGlucM43LM110L::siiE-BIg50-53) via HindIII, NheI digest and ligation. To obtain the whole SiiE-BIg49-52D-S, the second synthetic DNA fragment SiiE-BIg47-49half D-S was cloned into pWSK29::PsiiAGlucM43LM110L::siiE-BIg49half-52D-S via HindIII digestion and ligation. Orientation was checked by PstI, ClaI diagnostic digest and subsequent sequencing confirmed the construct.
The Q5 site-directed mutagenesis kit (NEB) was used to create plasmid with exchanges in codons for single conserved aspartate residues or exchanges for type I or type II Ca2+-binding sites in SiiE. Primers included new recognition sites for restriction enzymes through silent mutations. After performing colony PCR, the PCR fragment was digested with an appropriate restriction enzyme to confirm the silent mutations. Clones have also been confirmed by sequencing.
The construction of scar-less in-frame deletions in siiE using the I-SceI site was previously described by Blank et al. [32] and the protocol modified by Hoffmann et al. [33] was applied here. Briefly, the I-SceI aph resistance cassette was amplified from pWRG717 using primers with 20 bp homology and 40 bp 5’ overlap. The resistance cassette was inserted by Red-mediated recombination within the desired region. The plasmid pWRG730 was used instead of pKD46 and features a heat-inducible promoter for red genes. For preparing competent cells of WT [pWRG730], cells were grown in LB Cm10 to OD600 of 0.4–0.5 in a baffled flask in a shaking water bath at 150 rpm at 30°C. Subsequently, cells were immediately transferred to another water bath pre-heated to 42°C and incubated for 12.5 min at 100 rpm. After incubation of heat-induced cells on ice for 15 min, competent cells were prepared as described before [34]. After DpnI-digestion and purification, the PCR product was electroporated into Salmonella and transformants were selected on LB Km25 agar plates. Transformants were checked by colony PCR and confirmed clones were streaked on LB Cm10 plates with or without 100 ng x ml-1 anhydrotetracycline (AHT, Sigma-Aldrich). Clones with the highest inhibition on AHT containing plates at 30°C were selected for further procedures.
PCR fragments were amplified from plasmids containing exchanges of conserved aspartate codons in siiE or synthetic DNA were used. These fragments contain 5’ overlaps which are homolog to the insertion site of the chromosomally integrated I-SceI aph cassette. The strain containing the I-SceI aph cassette and harboring pWRG730 was then transformed by electroporation with either purified PCR product or synthetic DNA fragments. Serial dilutions from 10−1 to 10−4 were plated on LB Cm10 plates containing 100 ng x ml-1 AHT and incubated overnight at 30°C. The next day, large colonies were re-streaked on LB Cm10 AHT100 plates. Clones were verified by testing sensitivity to kanamycin and by colony PCR.
The Gaussia luciferase assay was performed as previously described by Wille et al. [17].
Bacterial strains were diluted 1:31 in LB from O/N cultures and grown at 37°C for 3.5 h. Aliquots of 1 ml of bacterial culture were collected, cells pelleted and resuspended in 1 ml of sterile LB. After an additional washing step with sterile LB, optical density was measured and adjusted to OD600 of 1 in 500 μl of 3% PFA in PBS. After fixation of bacterial cells for 15 min at RT, cells were pelleted (10,000 x g; 5 min) and resuspended in 500 μl PBS. Five microliters of bacterial suspensions were spotted on a nitrocellulose membrane which has been pre-wetted with PBS and dried again before adding bacteria. After drying of the spots, membranes were blocked with 5% dry milk powder in TBS/T (TBS; 0.1% Tween20) for at least 30 min. For detection of SiiE on the bacterial surface, antiserum against the C-terminal moiety of SiiE [35] was diluted 1:10,000 in blocking solution and applied to the membrane. LPS was detected using antiserum against Salmonella O-antigen (Becton-Dickinson) at the same dilution. After incubation O/N at 4°C, membranes were washed thrice with TBS/T and bound primary antibodies were detected with anti-rabbit IRDye 800CW (LI-COR) at a dilution of 1:20,000 in PBS/T (PBS; 0.1% Tween20). Subsequently, membranes were incubated for 1 h at RT in the dark and washed thrice with PBS/T. Membranes were rinsed in PBS and signals were quantified using the Odyssey Imaging System (LI-COR Biotechnology).
Bacterial strains were diluted 1:31 in LB from an O/N culture and grown at 37°C for 6 h. The OD600 was measured. Aliquots of 2 ml of bacterial culture were taken and pelleted. Supernatants were filter sterilized (0.2 μm Millex filter units, Millipore). 200 μl of 100% TCA was added to 1.8 ml of filtered sterilized supernatant and incubated O/N at 4°C. The precipitated proteins were pelleted by centrifugation at 4°C for 45 min. After two washing steps with 1 ml ice-cold acetone (14,000 x g, 30 min, 4°C) the precipitate was air dried and afterwards resuspended in 25 μl PBS per OD600 of 1. Dot blot analysis was carried out as described before without detection of LPS.
For Western blot analysis, O/N cultures were diluted 1:31 in LB and grown for 3.5 h at 37°C. OD600 was measured, and 150 μl were pelleted in 2 min at 4°C at 16.000 x g. The pellet was resuspended in OD600 x 50 μl in 1 x SDS sample buffer and incubated at 100°C for 5 min. 15 μl of each sample were loaded onto 3–8% gradient gels (NuPage) and electrophoretically separated for 1 h at 150 V. Semi-dry blotting was performed using a 0.2 μm nitrocellulose membrane with 64 mA/blot for 4 h. After blocking with 5% milk/TBS/T, the membrane was incubated first with a primary antibody against SiiE and subsequently with a secondary HRP-conjugated antibody against rabbit IgG. The signals were determines using ECL reagent (ThermoScientific) and the ChemiDoc system (BioRad).
Invasion assay was performed as previous described by Wagner et al. [6]. Briefly, O/N cultures of Salmonella strains were diluted 1:31 in LB and grown for 3.5 h in test tubes with aeration in a roller drum. The cultures were diluted in MEM medium to obtain a multiplicity of infection (MOI) of 5 and this inoculum as added to the MDCK cells. After infection for 25 min cells were washed three times with PBS to remove non-internalized bacteria, and medium was replaced by medium containing 100 μg x ml-1 gentamicin to kill remaining extracellular bacteria. After incubation for 1 h, cells were washed again with PBS, lysed by addition of 0.5% sodium desoxychlate in PBS, and colony forming units were determined by plating serial dilutions of the lysates onto agar plates.
All SiiECterm variants (p4033, p4034, p4462, p4463, S1 Table) were cloned into the pGEX-6P-1 vector (GE Healthcare) and expressed as GST fusion constructs in E. coli BL21 (DE3) (Novagen). The bacteria were chemically transformed with the expression plasmid and transformants selected on LB agar plates containing 100 μg x ml-1 ampicillin. The expression was done in terrific broth containing 100 μg x ml-1 ampicillin. A starter culture was inoculated with a single colony and grown over night at 37°C and 180 rpm. On the next day, the main expression cultures were inoculated to an OD600 of 0.08 and incubated at 37°C and 180 rpm. At OD600 0.6 to 0.7 the temperature was reduced to 20°C and at OD600 1.0 to 1.2 the protein expression was induced by adding 0.5 mM IPTG. After induction, the proteins were expressed for 20 h at 20°C and 180 rpm, the bacteria harvested by centrifugation and the pellets stored at -80°C until used for purification. Briefly, the GST-tagged proteins were captured by Glutathione Sepharose affinity column (GE Healthcare) using standard buffers given in the manual. The GST tag was cleaved by adding GST tagged HRV 3C protease at a mass ratio of 1 to 250 (protease-to-fusion protein) and the tag, undigested fusion proteins and the protease were extracted by a second Glutathione Sepharose purification. As a final purification step, the proteins were separated by size exclusion chromatography using a HiLoad 26/60 Superdex 200 pg column (GE Healthcare) and a buffer with 25 mM Tris-HCl, 150 mM NaCl, pH 8.0. No calcium was added to the buffers during the chromatographic purification. The protein was concentrated to 8 mg x ml-1, frozen in liquid nitrogen and stored at -80°C until use.
The calcium content of the SiiECterm variants was analysed by inductively coupled plasma—atom emission spectroscopy. All proteins were purified as described above, but without the second Glutathione Sepharose step after tag cleavage. The purified SiiECterm proteins were concentrated and dialysed against the same batch of buffer (5 mM Tris-HCl, pH 8.0, ratio of sample-to-buffer volume 1:100). Calcium standard for ICP (Sigma-Aldrich) was diluted to 0.5, 5 and 50 μg x ml-1 with the same buffer. The ICP-AES analyses were performed using a Ciros CCD (Spectro Analytical Instruments GmbH). Three individual measurements of the same sample were conducted for each variant and the mean calculated.
Limited proteolysis was performed in order to investigate the influence of the mutations on the conformation and compactness of the proteins [36]. The proteolysis experiments were conducted at 20°C and 550 rpm in a benchtop shaker. The assay was done in a buffer containing 25 mM Tris-HCl, 150 mM NaCl, pH 8.0 and the SiiE protein concentration was adjusted to 1 mg x ml-1. 10 μg α-Chymotrypsin or 0.5 μg Proteinase K (Proti-Ace & Proti-Ace 2 Kit, Hampton Research, Aliso Viejo, USA) were added per mg of SiiE protein. Aliquots of 18 μl were taken prior (0 min), and at various time points after protease addition (1 min, 5 min, 15 min, 30 min, 1 h, 2 h, 4 h, 7 h, O/N). The samples were mixed with 6 μl 4 x SDS-PAGE loading buffer and boiled at 95°C for 5 min to stop the cleavage reaction. The heat-treated samples were briefly spun down and stored at -20°C. The 10 μl of each sample were analyzed by SDS-PAGE using 15% polyacrylamide gels. Gels were stained with Coomassie Blue.
The conformation and stability of SiiE variants were probed using circular dichroism (CD) spectroscopy. All measurements were done with a Jasco J-815 spectropolarimeter (Jasco, Tokyo, Japan) using a cuvette with a 0.1 cm path length. CD spectra were recorded at 20°C in the far UV region between 185 and 260 nm in 10 mM potassium phosphate buffer, pH 8.0. Protein concentrations of 15 μM were used for SiiECterm variants. The band width was set to 1.0 nm, the scan speed to 20 nm x min−1, data integration time to 1 sec, data pitch to 0.1 nm and sensitivity to standard. Each measurement was averaged across ten accumulations and the protein spectra corrected for the sample buffer.
The stability of the proteins was compared by thermal scanning analysis. Changes in the secondary structure composition were investigated between 20°C and 96°C by monitoring the CD signal at wavelengths of 222 nm for SiiECterm proteins. A band width of 1.0 nm, data integration time of 8 sec, heat rate of 1°C x min-1, sampling rate of one data point per 0.2°C and standard sensitivity was used for all thermal scanning experiments. Conversion of the data to concentration- and length-independent mean residue weight (MRW) ellipticities [θ]MRW was done as described previously [37]. The secondary structure analysis of the CD-spectra and of the Ig-domain structure of SiiE BIg domains 50 to 52 (PDB entry: 2YN5) was done with single spectrum analysis and the secondary structure and beta-sheet decomposition for PDB-structures tools of the BeStSel server, respectively (http://bestsel.elte.hu/ssfrompdb.php) [38]. The wavelength range between 190 nm and 250 nm of the CD spectra was used for estimation of the secondary structure content.
Native polyacrylamide gel electrophoresis (PAGE) was used to analyze the aggregation tendency of the SiiE variants. All protein samples were adjusted to 0.3125 mg x ml-1 and 20 μl samples of each protein were incubated at various temperatures (20, 30, 41.7, 50, 60, 70, 80 and 90°C) for 5 min, briefly spun down and chilled on ice for 1 min. 5 μl of 5 x native PAGE-buffer (0.25% (w/v) Bromophenol blue, 4.5% (w/v) sucrose) were added to each sample to achieve a final protein concentration of 0.25 mg x ml-1 and 15 μl (3.75 μg protein) were loaded per sample. Native PAGE was done using 7.5% native polyacrylamide gels and native PAGE running buffer (50 mM Tris, 384 mM glycine). The gel runs were done at 5 mA per gel for 5 h and 6–8°C in the cold room and gels were stained with Coomassie Blue. For the SDS-PAGE analysis, 4 μl of 4 x SDS sample buffer were added to 16 μl of each sample, the mixture boiled at 95°C for 5 min and briefly centrifuged. 10 μl of each of the samples were analyzed on 15% acrylamide gels. The gel runs were done at 200 V for 60 min.
The crystal structure of SiiE wild-type BIg domains 50–52 (PDB code 2YN5, chain A) was used for all computational studies. Based on the wild-type system, three mutants were modelled that lack either the type I, the type II, or both types of Ca2+-binding sites. Mutants were generated by replacing the respective coordinating aspartate and glutamate residues by serine.
All systems were neutralized by adding an appropriate amount of sodium counter ions. Each system was placed in a periodic TIP3P water box [39] extending at least 12 Å in all directions from the solute. All simulations were done with Amber 14 [40] using the ff99SB force field [41]. Long-range electrostatics were calculated with the particle mesh Ewald (PME) approximation [42]. Shake was used to constrain hydrogen atoms during equilibration and simulation [43].
Minimization, equilibration and MD calculations were carried out with the pmemd module of AMBER. Minimizations were run for 10,000 steps and switched from steepest descent to conjugate gradient after 500 cycles. During equilibration the system was gradually heated from 30 K to 310 K in 60 ps with backbone restraints of 2.0 kcal/mol Å2 and then relaxed at 310 K with backbone restraints of 0.2 kcal/mol Å2 for another 20 ps. Two independent production runs of 300 ns were generated for each system using the weak-coupling algorithm [44] and a Berendsen barostat in an NPT ensemble.
For the SMD simulations ten restart files containing atomic coordinates and velocities were taken from wild-type MD simulation (at intervals of 10 ns). Ca2+ ions were pulled from type I and type II sites of SiiE individually by a harmonic potential with the spring energy constant of 50 kcal/mol Å2. The center of this potential was moved away from the center of mass (CoM) of the backbone atoms of one BIg domain with a constant velocity of 0.2 Å/ps. In this way the force was evenly distributed on the protein and no positional restraints had to be used.
|
10.1371/journal.pmed.1002565 | Access to antiretroviral therapy in HIV-infected children aged 0–19 years in the International Epidemiology Databases to Evaluate AIDS (IeDEA) Global Cohort Consortium, 2004–2015: A prospective cohort study | Access to antiretroviral therapy (ART) is a global priority. However, the attrition across the continuum of care for HIV-infected children between their HIV diagnosis and ART initiation is not well known. We analyzed the time from enrollment into HIV care to ART initiation in HIV-infected children within the International Epidemiology Databases to Evaluate AIDS (IeDEA) Global Cohort Consortium.
We included 135,479 HIV-1-infected children, aged 0–19 years and ART-naïve at enrollment, between 1 January 2004 and 31 December 2015, in IeDEA cohorts from Central Africa (3 countries; n = 4,948), East Africa (3 countries; n = 22,827), West Africa (7 countries; n = 7,372), Southern Africa (6 countries; n = 93,799), Asia-Pacific (6 countries; n = 4,045), and Latin America (7 countries; n = 2,488). Follow-up in these cohorts is typically every 3–6 months. We described time to ART initiation and missed opportunities (death or loss to follow-up [LTFU]: last clinical visit >6 months) since baseline (the date of HIV diagnosis or, if unavailable, date of enrollment). Cumulative incidence functions (CIFs) for and determinants of ART initiation were computed, with death and LTFU as competing risks. Among the 135,479 children included, 99,404 (73.4%) initiated ART, 1.9% died, 1.4% were transferred out, and 20.4% were lost to follow-up before ART initiation. The 24-month CIF for ART initiation was 68.2% (95% CI: 67.9%–68.4%); it was lower in sub-Saharan Africa—ranging from 49.8% (95% CI: 48.4%–51.2%) in Central Africa to 72.5% (95% CI: 71.5%–73.5%) in West Africa—compared to Latin America (71.0%, 95% CI: 69.1%–72.7%) and the Asia-Pacific (78.3%, 95% CI: 76.9%–79.6%). Adolescents aged 15–19 years and infants <1 year had the lowest cumulative incidence of ART initiation compared to other ages: 62.2% (95% CI: 61.6%–62.8%) and 66.4% (95% CI: 65.7%–67.0%), respectively. Overall, 49.1% were ART-eligible per local guidelines at baseline, of whom 80.6% initiated ART. The following children had lower cumulative incidence of ART initiation: female children (p < 0.01); those aged <1 year, 2–4 years, 5–9 years, and 15–19 years (versus those aged 10–14 years, p < 0.01); those who became eligible during follow-up (versus eligible at enrollment, p < 0.01); and those receiving care in low-income or lower-middle-income countries (p < 0.01). The main limitations of our study include left truncation and survivor bias, caused by deaths of children prior to enrollment, and use of enrollment date as a proxy for missing data on date of HIV diagnosis, which could have led to underestimation of the time between HIV diagnosis and ART initiation.
In this study, 68% of HIV-infected children initiated ART by 24 months. However, there was a substantial risk of LTFU before ART initiation, which may also represent undocumented mortality. In 2015, many obstacles to ART initiation remained, with substantial inequities. More effective and targeted interventions to improve access are needed to reach the target of treating 90% of HIV-infected children with ART.
| Access to antiretroviral therapy (ART) has been highlighted as an urgent global priority area by a diverse range of stakeholders, including the World Health Organization, the Joint United Nations Programme on HIV/AIDS (UNAIDS), and, more specifically for children, the United Nations Children’s Fund.
In 2014, UNAIDS set the ambitious 90-90-90 targets that, by 2020, 90% of people living with HIV should know their HIV status, 90% of HIV-infected people who know their HIV status should receive antiretroviral treatment, and 90% of people on treatment should be virologically suppressed.
In low/middle-income countries, the attrition across the continuum of care for HIV-infected children between their HIV diagnosis, linkage to care, and ART initiation is not well characterized.
We used the International Epidemiological Databases to Evaluate AIDS (IeDEA), bringing together data from 6 different regions worldwide (in Asia-Pacific, sub-Saharan Africa, and Latin America), including the regions most affected by the HIV epidemic, to describe access to pediatric ART since the scaling up of ART.
We described the cumulative incidence of and time to ART initiation in HIV-infected children in care between 2004 and 2015 (n = 135,479) according to region, sex, age, and eligibility criteria at baseline, since enrollment in care or HIV diagnosis, when this was available.
The global probability for initiating ART within 2 years in care was 68%; 20% of the children were lost to follow-up, and 2% died before initiating treatment.
ART initiation rates were lower in sub-Saharan Africa compared to the other regions, and children aged <1 year and those aged 15–19 years were the least likely to initiate treatment.
Overall, 49% of the children were eligible for ART at baseline, and 81% of these children initiated treatment.
In many low- and middle-income countries, obstacles to ART initiation for children remained, with substantial inequities, in 2015.
More effective and targeted interventions to improve access are needed to reach the target of treating 90% of children with ART.
With adoption of universal treatment recommendations since 2015, it will be crucial to further monitor progress and identify gaps in ART coverage to achieve the 90-90-90 targets for children and adolescents.
| By the end of 2016, the Joint United Nations Programme on HIV/AIDS (UNAIDS) estimated that 2.1 million children aged <15 years were living with HIV worldwide [1]. Despite effective interventions for the prevention of mother-to-child transmission, the pediatric epidemic persists, and an estimated 160,000 children were newly infected with HIV in 2016 [1]. Furthermore, the incidence of HIV remains alarmingly high in adolescents and young people aged 15–24 years. According to UNAIDS, 37% of new HIV infections occurring in sub-Saharan African adults in 2016 were among this population [1].
In the absence of timely antiretroviral therapy (ART), mortality in HIV-infected children reaches 52% by the age of 2 years [2]. Systematic early ART initiation in infants <3 months of age has proven an effective intervention in reducing early infant mortality, and the World Health Organization (WHO) has recommended ART in all HIV-infected children aged <2 years since 2010; in 2013, the recommendation to initiate ART regardless of clinical stage or CD4 count was extended to children <5 years [3–5]. In 2015, guidelines were revised to recommend ART initiation regardless of age, clinical, or immunological criteria [6]. However, in 2015, 49% of children aged <15 years who were eligible for ART based on previous guidelines were still not receiving treatment worldwide [7]. Many of those children who did not initiate ART either had no access to treatment or had unknown HIV status, mainly through lack of access to early HIV diagnosis.
In 2014, UNAIDS set the ambitious targets that, by 2020, 90% of people living with HIV should know their HIV status, 90% of people who know their HIV status should receive treatment, and 90% of people on treatment should be virologically suppressed [8]. However, in low-income and middle-income countries, attrition across the continuum of care for HIV-infected children between their HIV diagnosis and ART initiation is not well known. The period between diagnosis of HIV infection and ART initiation, which includes periods from testing to linkage to care and then from inclusion in care to ART initiation, is known as the pre-ART cascade. Throughout the pre-ART period, children are at risk for attrition due to death and loss to follow-up (LTFU). A better understanding of the attrition of children across the pre-ART cascade is crucial to reach the 90-90-90 targets, particularly for the second target, to initiate ART in 90% of HIV-infected children identified, assuming that 90% have been identified among those HIV infected.
In 2006, the US National Institutes of Health launched the International Epidemiology Databases to Evaluate AIDS (IeDEA) to describe trends in HIV epidemiology in the context of ART access across regions of the world (https://www.iedea.org/). Clinics from 7 international regional data centers contribute both retrospective cohort data, prior to 2006, and prospective data since, on care and treatment of HIV to evaluate the outcomes of people living with HIV/AIDS. To better understand the continuum of care from HIV diagnosis to ART initiation in HIV-infected children and in collaboration with WHO, we performed a multiregional analysis of the pre-ART retention cascade of HIV-infected children from HIV diagnosis to ART initiation within IeDEA from 2004 to 2015.
This multiregional analysis was prespecified, except for the sensitivity analysis, in an approved concept plan available in the Supporting Information (S1 Concept Plan). We pooled individual patient data from 6 pediatric cohorts of IeDEA and included clinical care sites from Asia-Pacific, West Africa, East Africa, Central Africa, Southern Africa, and Latin America. We included all HIV-infected children aged 0–19 years at enrollment into any IeDEA-affiliated pediatric care program who were ART-naïve at enrollment (except for exposure to perinatal prevention of mother-to-child transmission prophylaxis) between 1 January 2004 (corresponding to the beginning of the era of ART access in low-income countries) and 31 December 2015.
Although each clinic within IeDEA has its own protocol for routine follow-up, HIV-infected children who have not yet been initiated on ART are typically seen at least every 6 months.
The data abstracted for this analysis were from routine care and included region, country, site, patient demographics (sex, date of birth, date of HIV diagnosis if available, and date of enrollment in care), clinical WHO/CDC staging at enrollment and ART initiation, laboratory values and dates (CD4 cell count, CD4 percent), cotrimoxazole start date, date of ART initiation, ART regimen, and date of death, LTFU, or transfer out.
The outcomes of interest were (1) time to ART initiation and (2) missed opportunities for ART initiation, defined as either death or LTFU (last clinical visit >6 months before database closure and whereabouts unknown). Baseline was defined as the date of diagnosis, whether this occurred before or after enrollment in the HIV care program; when this date was unavailable, which was frequent in these contexts, we defined baseline as the date of enrollment in the HIV care program. The follow-up period was defined as the time between baseline and the date of ART initiation, death, LTFU, transfer out, or database closure (31 December 2015), whichever came first.
We assessed the proportion of children initiating ART and the proportion with missed opportunities for ART initiation (death or LTFU) within the first 24 months after enrollment. This conservative time-point was prespecified to provide a reference analysis in scaling up HIV diagnosis and ART initiation before the treat-all era from 2015 onwards. Baseline categorical data are presented as frequency (percent), and continuous data are presented as median (interquartile range [IQR]). Continuous variables were compared using the Kruskal–Wallis test, and categorical variables using the chi-squared or Fisher’s test. Time to ART initiation was estimated using cumulative incidence functions (CIFs): mortality and LTFU were considered competing events to ART initiation [9], while those transferred out were right-censored, with the assumption that they remained in HIV care and had similar outcomes as patients still in observation.
To better understand the different patterns of ART initiation, taking into account the evolving WHO eligibility criteria for ART initiation between 2004 and 2015 [4,5,10,11], we described the percentage of children eligible for ART initiation at baseline and during follow-up according to WHO guidelines, combining clinical criteria (WHO stage 3 or 4 or AIDS, though these data were unavailable in the Southern Africa database) and severe immunodeficiency for age (CD4 ≤ 25% if age < 5 years or CD4 ≤ 350 cells/μl if age ≥ 5 years) if enrollment in care occurred prior to 1 April 2008; additional age criteria were added if enrolled later (age <1 year between 1 April 2008 and 30 June 2010, age < 2 years between 1 July 2010 and 31 May 2013, and age < 5 years on or after 1 June 2013).
Correlates of ART initiation were described in a multivariate competing risks analysis using the Fine and Gray proportional sub-distribution hazards regression model, where time of origin was baseline [12]. Explanatory variables included sex, age at baseline, region, country income as defined by the World Bank (http://data.worldbank.org/about/country-and-lending-groups), period of enrollment based on the changing WHO treatment guidelines (before April 2008, April 2008–June 2010, July 2010–May 2013, and after May 2013), and clinical/immunological criteria for ART eligibility at baseline and during follow-up. Model fit was checked graphically, plotting the Schoenfeld-type residuals against time for each of the covariates included in the model; we used the %pshreg SAS macro for this [13,14].
There was a high degree of variability of data collection practices across the HIV programs included in IeDEA, and some centers followed up only children who were eligible at baseline to be initiated on ART. It is likely that those programs only recorded data if children initiated ART or were intended to initiate ART before being followed up in a decentralized center. Consequently, incidence of ART initiation among all children in HIV care may be overestimated by these programs. For this reason, we performed an un-prespecified sensitivity analysis excluding the clinics where the ART coverage was >95% of all the patients followed up.
Analyses were conducted using the package cmprsk in R statistical software version 2.11.1 (R Foundation for Statistical Computing, Vienna, Austria). The adjusted sub-distribution hazard ratios (asHRs) were reported with their 95% confidence intervals (CIs). A p-value less than 0.05 was considered statistically significant.
Each participating IeDEA region formally agreed to contribute pediatric data, with local institutional review board and US National Institutes of Health approvals to contribute to multiregional analyses.
Overall, 180,419 children and adolescents were included in the IeDEA, of whom 120,413 were <15 years; this represents about 6% of all children living with HIV worldwide according to UNAIDS estimates. Of these, 3,262 (1.9%) were excluded due to incoherent data, 35,439 (19.6%) because they were >19 years of age, 2,194 (1.2%) because they were not ART-naïve at baseline—these children most likely entered in IeDEA active files transferring from other clinics while they were already on ART but without a date of ART initiation—and 4,045 (2.2%) because the date of enrollment in the site was before 2004 or after 2015, leaving 135,479 (75.1%) who met the inclusion criteria and were included in the study; 69.2% were from Southern Africa. Among all children included, 27,831 (20.5%) had a documented date of HIV diagnosis at or after enrollment (for those HIV-exposed) and 18,035 (13.3%) had a documented date of HIV diagnosis prior to enrollment; for the remaining 89,613 (66.1%), the date of enrollment was used as baseline (Fig 1).
Table 1 describes the patients’ characteristics at baseline. Children were enrolled at a late age, 6 years in median (IQR: 2–12), but the median age varied by region, ranging from 4 years in West Africa (IQR: 2–9) and Asia-Pacific (IQR: 2–7) to 8 years in Latin America (IQR: 2–16). This variation was mainly driven by the proportion of adolescents enrolled in the HIV care programs in each region. Overall, 32.1% of children were adolescents (≥10 years) at baseline. This varied across regions, with Asia-Pacific having the smallest proportion (9.9%), and Latin America the largest (43.2%). Children aged 5–9 years were the most commonly represented group overall (23.5%). Median CD4 percentage at baseline was 17% (IQR: 10%–25%) overall: 12% in Asia-Pacific, 15% in West Africa, 16% in Central Africa and Latin America, 17% in Southern Africa, and 19% in East Africa (p < 0.01). At the time of their enrollment, 49.1% of children were known to be eligible for ART initiation according to WHO recommendations. This percentage ranged from 43.0% in Central Africa to 74.5% in the Asia-Pacific region. Forty-eight percent of children could not be classified for ART eligibility at baseline due to missing clinical and/or immunological data.
Access to care varied greatly over regions. Overall, 31.5% of children were enrolled prior to April 2008; this represented 52.9% of the cohort in the Asia-Pacific region but only 28.1% in Southern Africa. Of all children in the database, 66.1% did not have a date of confirmed HIV diagnosis in the database: in Latin America and Asia-Pacific, a smaller proportion of children lacked HIV diagnosis date (0.0% and 2.8%, respectively), but in Southern Africa this proportion was 78.4%. Among those with a date of confirmed HIV diagnosis, 39.3% of children were diagnosed prior to enrollment in IeDEA, ranging from 14.8% in East Africa to 60.3% in West Africa (p < 0.001; Table 1); the overall median time to enrollment in care since HIV diagnosis among these children was 1 month (IQR: 0–7).
Among the 135,479 children included in the study, 99,404 (73.4%) initiated ART. The median time to ART initiation was 1 month (IQR: 0–6 months); 1.9% died, 1.4% were transferred out, and 20.4% were lost to follow-up before any ART initiation (Table 2). Time to ART initiation varied according to the timing of diagnosis. Among those diagnosed prior to enrollment in care, the median time to ART since diagnosis was 4 months (IQR: 1–19); among those diagnosed after enrollment in care, median time to ART was 3 months (IQR: 1–10). Among those with no documented date of HIV diagnosis, median time from enrollment to ART was 1 month (IQR: 0–5). Finally, among those enrolled at time of diagnosis, the median time from baseline to ART was 2 months (IQR: 0–8).
After 1 month of pre-ART follow-up, the CIF for ART initiation in all HIV-infected children was estimated to be 35.7% (95% CI: 35.4%–35.9%) while the CIF for missed opportunities (death or LTFU) was 10.7% (95% CI: 10.5%–10.9%). By 24 months of pre-ART follow-up, the CIF for ART initiation had reached 68.2% (95% CI: 67.9%–68.4%), and the 24-month cumulative incidence for missed opportunities was 19.3% (95% CI: 19.1%–19.5%). Fig 2A presents the 24-month CIF for time to ART initiation since baseline by region. It was 71.0% (95% CI: 69.1%–72.7%) and 78.3% (95% CI: 76.9%–79.6%) for Latin America and the Asia-Pacific, respectively. In sub-Saharan Africa, the 24-month CIF for ART initiation was significantly lower, ranging from 49.8% (95% CI: 48.4%–51.2%) in Central Africa to 60.9% (95% CI: 60.3%–61.6%) in East Africa, 70.1% (95% CI: 69.8%–70.4%) in Southern Africa, and 72.5% (95% CI: 71.5%–73.5%) in West Africa. We also noted that results differed by age at baseline: children and adolescents aged 15–19 years and those aged <1 year at baseline had the lowest ART initiation rates compared to other ages, with a 24-month CIF of 62.2% (95% CI: 61.6%–62.8%) and 66.4% (95% CI: 65.7%–67.0%), respectively (Fig 2B).
Fig 2C presents CIFs for ART initiation according to ART eligibility at baseline. Among children eligible at baseline, the 24-month CIF was 78.6% (95% CI: 78.3%–78.9%) compared to 56.4% (95% CI: 55.1%–57.7%) among those known not to be eligible at baseline who became eligible during follow-up. We also note that the CIF for ART initiation reached 39.4% (95% CI: 37.6%–41.1%) at 24 months among children who did not meet ART eligibility criteria. Among those with unknown eligibility criteria at baseline, the 24-month CIF for ART initiation was 58.7% (95% CI: 58.3%–59.1%) (S3 Fig).
To better understand the patterns of ART initiation, we present the ART eligibility for the overall population (Fig 3). Of the 135,479 children included in the analysis, 66,482 (49.1%) were known to be eligible for ART initiation according to WHO criteria at baseline, of whom 53,616 (80.6%) initiated ART. Among the 3,674 (2.7%) not eligible for ART at baseline, 5 (0.1%) became eligible during follow-up, of whom all initiated ART. Among the 65,323 (48.2%) who were not classified for ART eligibility at baseline due to missing clinical and/or immunological data, 5,345 (8.2%) became eligible, of whom 4,275 (80.0%) initiated ART, and 692 (1%) did not meet eligibility criteria during follow-up, of whom 398 (57%) initiated ART. Overall, 13,936 (19.4%) children known to be eligible for ART initiation, either at baseline or during follow-up, never accessed treatment over the first 24 months of follow-up. These missed opportunities for ART initiation highlight a substantial proportion of unmet needs.
In 2015, no region had yet reached the UNAIDS target of 90% of those diagnosed with HIV infection on ART, though the Asia-Pacific, Latin America, and Southern and West Africa IeDEA regions had percentages of ART initiation close to 80% (Fig 4). Central and East Africa had the lowest coverage (49% and 59%, respectively). A substantial proportion of eligible children did not initiate ART during the study period, ranging from 9% in Latin America to 15% in West Africa and 16% in Central Africa.
In the multivariate Fine and Gray analysis, sex, region, age at baseline, period of enrollment, country income, and clinical or immunological eligibility were all associated with ART initiation (Table 3). Adjusted for the above variables, females were less likely to initiate ART compared to their male counterparts (asHR: 0.94, 95% CI: 0.92–0.95). Compared to Latin America, 3 sub-Saharan African regions were less likely to initiate ART, with asHRs ranging from 0.69 (95% CI: 0.66–0.72) in Central Africa and 0.80 (95% CI: 0.77–0.83) in East Africa to 0.93 (95% CI: 0.90–0.97) in Southern Africa. Children from the Asia-Pacific and West Africa regions had similar hazard of treatment initiation compared to those from Latin America (asHR: 1.01, 95% CI: 0.97–1.05, and asHR: 1.02, 95% CI: 0.97–1.06, respectively). Adolescents aged 10–14 years at baseline were the most likely to initiate ART compared to all other age groups. Infants aged <1 year and adolescents aged ≥15 years were less likely to initiate ART compared to those 10–14 years old (asHR: 0.83, 95% CI: 0.81–0.85, and asHR: 0.73, 95% CI: 0.71–0.75, respectively). Children enrolled prior to June 2013 were also less likely to initiate ART than those enrolled more recently, adjusted for other variables, and we noted a tendency towards lower likelihood of initiating ART in earlier enrollment periods (asHR: 0.78, 95% CI: 0.76–0.80, for those with enrollment in July 2010–May 2013; asHR: 0.63,95% CI: 0.61–0.64, for those enrolled in the period April 2008–June 2010; and asHR: 0.57, 95% CI: 0.55–0.58, for those enrolled prior to April 2008) compared to those enrolled in or beyond June 2013. ART initiation was also associated with country income: children from countries of low or lower middle income were less likely to initiate ART compared to those from upper-middle- and high-income settings (asHR: 0.76, 95% CI: 0.75–0.78). Finally, we found that children who became eligible for ART initiation (per clinical and/or immunological criteria, but not on age criteria—that was a separate variable) during follow-up were less likely to initiate ART compared to those who were eligible at baseline (asHR: 0.61, 95% CI: 0.60–0.63).
Overall, 112,134 children were followed up in clinics where ART coverage was ≤95%. Results are presented in the Supporting Information (S2 and S3 Figs; S1 Table). Among these children, the 24-month CIF for ART initiation was estimated to be much lower than in the whole population, at 62.9% (95% CI: 62.6%–63.2%), and the 24-month probability for missed opportunities for ART was higher, at 23.3% (95% CI: 23.1%–23.6%), than in the whole population (S2 Fig). The children from Asia-Pacific had the highest ART initiation rates, reaching 73.1% (95% CI: 71.4%–74.8%), compared to all others, including those from sub-Saharan Africa, with 49.8% (95% CI: 48.4%–51.2%) in Central Africa, 60.2% (95% CI: 58.8%–61.7%) in West Africa, 60.9% (95% CI: 58.8%–61.7%) in East Africa, and 60.2% (95% CI: 58.8%–61.7%) in Southern Africa (S3 Fig). In the multivariate Fine and Gray competing risk analysis, we did not observe significant changes in correlates of ART initiation in this population, except for the effect of region, with the lowest asHR for ART initiation in sub-Saharan Africa compared to Latin America, ranging from 0.60 in Central Africa (95% CI: 0.57–0.63) and West Africa (95% CI: 0.57–0.63) and 0.70 in East Africa (95% CI: 0.67–0.73) to 0.80 in Southern Africa (95% CI: 0.77–0.83) (S1 Table).
This pooled analysis from the IeDEA Global Cohort Consortium documents time to ART initiation since enrollment in HIV programs treating HIV-infected children and adolescents between the ages of 0 and 19 years within multiple geographic regions, between 2004 and 2015. We report 3 major findings. First, in HIV-infected children and adolescents, the cumulative incidence of initiating ART within 24 months of enrollment into a program or HIV diagnosis was 68%, with a substantial risk for mortality or LTFU before ART initiation (19%), representing multiple missed opportunities for ART initiation. Second, among children eligible for ART initiation and followed up, 19% did not initiate treatment within the first 24 months of follow-up. Third, we report a number of inequities in ART access: female sex, children <10 years at baseline (and those <1 year in particular), adolescents aged 15–19 years at baseline (compared to those aged 10–14 years), those becoming eligible during follow-up (compared to those eligible at baseline), and those living in sub-Saharan Africa compared to other regions were all less likely to initiate treatment.
The 24-month cumulative incidence for ART initiation was 68%. According to ART eligibility, this was 78.6% among those eligible at baseline and 56.4% among those who became eligible during the study. While these rates are low compared to the UNAIDS target of 90% of those diagnosed with HIV, they are encouraging compared to previous available studies. For example, in South Africa, only 34.8% of ART-eligible HIV-infected children initiated ART in 2003 [15]. More recently, in Lesotho, 41.2% of eligible children initiated ART in 2008 [16]. In Côte d’Ivoire, 55% of eligible children initiated ART in 2009 [17]. Still, the rate of ART initiation among children often lags compared to adults and occurs late, despite progressive guidelines stressing immediate treatment initiation for the youngest children regardless of immunological/clinical status since 2008 [18].
At inclusion in HIV programs, 66.1% of children did not have a date of confirmed HIV diagnosis, median age was 6 years, and 49.1% were already eligible for ART, highlighting the late access to ART. Delayed ART initiation is most likely the result of late access to HIV diagnosis for HIV-exposed children [19,20]. Difficulties in identifying HIV-exposed infants, limited capacity to perform routine virological testing in HIV-exposed infants, and long result turnaround time remain important barriers to timely initiation of treatment [20]. In addition, the lack of integration of prevention of mother-to-child transmission and pediatric HIV care programs hampers the delivery of early infant diagnosis [21]. This is mainly related to infrastructure limitations and time constraints as well as staff shortages [20,22,23]. In older children, similar structural barriers have been identified [21], along with additional key barriers such as stigma, lack of knowledge in the adolescent population, and socio-cultural beliefs [24–27].
There are many points at which children may drop out the HIV care cascade. First and foremost, linkage to care after HIV diagnosis remains a complex issue [28]. There are many reasons for failing to link a child to care including fear of stigma [29–31], community and economic factors such as lack of support and finances for transport, missed days of work, and healthcare worker and infrastructure constraints (e.g., drug stock outs, lack of knowledge among providers on when to prescribe ART, patients missing appointments, eligible patients not identified appropriately) [32,33]. Challenges continue even after linkage to care. Although our results demonstrate the feasibility of large-scale ART rollout for children, 1.9% died and 20.4% were lost to follow-up before ART initiation. Some of the children who were lost to HIV care programs may represent undocumented mortality, but they may also be out of care or may have transferred to other facilities without documentation. Once children were linked to HIV care, we also observed suboptimal rates of ART initiation among children who became eligible during follow-up compared to those eligible at enrollment, further undermining the HIV care cascade.
Direct comparisons with other studies are made difficult by differences in methodology and definition of outcomes, but our observations underline the many missed opportunities for ART initiation and the difficulties in keeping children in the pre-ART care cascade. Low retention among HIV-diagnosed patients waiting to initiate ART, as observed in our study, has also been previously described in both adult and pediatric populations. Most losses happen between HIV diagnosis and CD4 staging [34]. Current guidelines no longer require CD4 staging to start ART, which may contribute to improved retention in care and access to ART.
Weaknesses in the continuity of care services must urgently be addressed in order to improve ART coverage and survival among HIV-infected children. For perinatally infected children, other interventions such as family-centered models have also been proposed to improve linkage to care after diagnosis [35,36]. In older children, use of youth-friendly models of care may be an important intervention [37–39]. Regardless of the mode and age at infection, multiple efforts are necessary to reach high uptake of services [40,41]. HIV testing and care need to be decentralized and brought to communities [42]. There is a need to support families and healthcare workers to provide HIV services for children [6,43,44]. Finally, the “test and treat” strategy recommended by WHO in 2015 that advocates starting HIV-infected individuals on ART immediately regardless of any eligibility criteria could further prevent these missed opportunities for ART initiation among HIV-diagnosed children.
We observed disparities in ART initiation between regions, with 24-month cumulative incidence of ART initiation ranging from <75% in sub-Saharan Africa to 78.3% in Asia-Pacific. This variability could be partly explained by the smaller number of HIV-infected children treated in Asia-Pacific compared to sub-Saharan Africa, with a larger sample size [6].
We also highlight other inequities in the rollout of ART. Females, children aged <10 years, and in particular those aged <1 year, along with adolescents aged ≥15 years, were less likely to initiate treatment. Previous studies have reported on missed opportunities for ART initiation in both very young children, mostly explained by early mortality before accessing HIV diagnosis and subsequent ART initiation, and adolescents, where fear of stigma in the family and community as well as parental consent requirements are major barriers [25,40,45]. Missed opportunities for ART initiation in adolescents aged 15–19 years could also be a reflection of noncompliance with visits. In addition, we observed that children who became eligible during follow-up were less likely to initiate treatment (asHR: 0.69, 95% CI: 0.68–0.70) compared to those eligible at enrollment. From a programmatic point of view, this observation strongly supports the universal “test and treat” strategy [6].
Our results indicate insufficient levels of ART initiation among children who were treatment eligible over the whole study period, but we also observed a gradual improvement in these rates during more recent time periods. As WHO guidelines began recommending universal ART in all children and adolescents irrespective of clinical stage or CD4 count in 2015, we expect the number needed to be treated increased in 2016. While there are still many obstacles that will impede the target of 90% ART coverage, there is an ethical priority to trace all HIV-exposed children in order to determine HIV status and link them into care if HIV infected and to treat all children who have already linked to care.
This study has major strengths but also several limitations. First, time to ART initiation since HIV diagnosis may be incorrectly estimated as data regarding confirmed dates of HIV diagnosis were scarce. We used enrollment as a proxy for HIV diagnosis, and therefore time between HIV diagnosis and ART initiation is likely underestimated in programs that do not have well-documented HIV diagnosis dates. Furthermore, a left-truncation phenomenon would mask deaths among HIV-infected children between their HIV diagnosis and inclusion into HIV programs. This survivor bias undoubtedly leads to further underestimation of the incidence of missed opportunities for ART initiation among HIV-infected children, and, consequently, the true cumulative incidence of ART initiation among all HIV-infected children is likely lower than that estimated by these data. Our results illustrate this well: programs where higher proportions were eligible at diagnosis seemed to do better in terms of ART initiation, whereas in reality these programs were doing worse as many children were diagnosed too late, with advanced HIV disease at entry. Imputation of time of diagnosis for those with missing date of diagnosis could have addressed this limitation in theory, but missing values were too numerous for this to be done. Because diagnosis is often performed at the same time as inclusion in care, our results reflect as best as possible the situation in routine care.
Second, we observed limitations inherent to data quality: 45% of children had missing data on variables used to assess ART eligibility. We thus advise caution in the interpretation of our results, particularly in the context of current universal treatment recommendations. Third, 24-month incidence of ART initiation and missed opportunities for care were derived from data collected over a 10-year period, during which both national and international guidelines varied over time. Although we adjusted for evolving ART eligibility criteria during follow-up in our final model to address this, further analyses would be necessary to better describe the progress made and treatment gaps on a national level.
Fourth, the outcomes of children lost to follow-up were not well known, and the high proportion of children lost to follow-up (20.4%) includes undocumented mortality, those out of care, and silent transfers. Both death and being out of care represent poor outcomes from missed opportunities, but silent transfers may or may not represent ultimate ART initiation. In the absence of outreach data, it is unclear how these results should be interpreted. To overcome these limitations, we combined mortality and LTFU as a single outcome (of programmatic failure) in estimating the cumulative rate of missed opportunities for ART initiation after enrollment in an HIV care program.
Fifth, HIV programs across and within regions vary. Some programs were specific to children once they started therapy, making it likely that those programs only recorded data if children initiated ART or were intended to initiate ART but not during the pre-ART period. As a result, our analysis may further overestimate the overall proportion of HIV-infected children starting ART as the denominator of all children may have shrunk in these instances. In sensitivity analyses where we excluded such programs from consideration, we observed no significant differences in our results, except in overall and regional ART initiation probabilities, which, as expected, were lower when restricted to sites with ART coverage ≤95%.
The mitigating factor of all these limitations is that the bias in estimating overall cumulative ART initiation rates goes in one direction, leading to overestimation of the cumulative rate of the start of treatment. This actually strengthens rather than undermines our conclusions regarding the continued challenge of universal care and treatment of children and adolescents living with HIV around the world. In addition, our study is the largest study reported to our knowledge to document the global pre-ART cascade in pediatrics in 2015, including data from a large number of diverse programs with significant geographic coverage; this study is a generalizable, authoritative investigation of the state of the worldwide response to the HIV epidemic in pediatric and adolescent populations living with HIV.
In conclusion, this large global cohort study of children with HIV reported a 24-month cumulative incidence of ART initiation of 68.2% between 2004 and 2015, and a high 19.3% cumulative incidence of program attrition prior to treatment start driven by mortality and loss to programs. Given the limitations to our study data, actual coverage of ART initiation in children during the study period is likely to have been lower than the estimates reported. As of 2015, there remain many obstacles to ART initiation, with substantial risks of loss to programs and death before ART initiation in the context of incomplete early infant diagnosis, linkage to care, and treatment initiation even after enrollment in care. In particular, infants <1 year of age and older adolescents urgently need more effective and targeted interventions to improve their HIV testing uptake and access to ART in order to facilitate their survival. With expanding adoption of universal treatment recommendations since 2015, it will be crucial to further monitor progress and identify gaps in ART coverage to achieve the 90-90-90 targets for children and adolescents.
|
10.1371/journal.pcbi.1003312 | Improving the Modeling of Disease Data from the Government Surveillance System: A Case Study on Malaria in the Brazilian Amazon | The study of the effect of large-scale drivers (e.g., climate) of human diseases typically relies on aggregate disease data collected by the government surveillance network. The usual approach to analyze these data, however, often ignores a) changes in the total number of individuals examined, b) the bias towards symptomatic individuals in routine government surveillance, and; c) the influence that observations can have on disease dynamics. Here, we highlight the consequences of ignoring the problems listed above and develop a novel modeling framework to circumvent them, which is illustrated using simulations and real malaria data. Our simulations reveal that trends in the number of disease cases do not necessarily imply similar trends in infection prevalence or incidence, due to the strong influence of concurrent changes in sampling effort. We also show that ignoring decreases in the pool of infected individuals due to the treatment of part of these individuals can hamper reliable inference on infection incidence. We propose a model that avoids these problems, being a compromise between phenomenological statistical models and mechanistic disease dynamics models; in particular, a cross-validation exercise reveals that it has better out-of-sample predictive performance than both of these alternative models. Our case study in the Brazilian Amazon reveals that infection prevalence was high in 2004–2008 (prevalence of 4% with 95% CI of 3–5%), with outbreaks (prevalence up to 18%) occurring during the dry season of the year. After this period, infection prevalence decreased substantially (0.9% with 95% CI of 0.8–1.1%), which is due to a large reduction in infection incidence (i.e., incidence in 2008–2010 was approximately one fifth of the incidence in 2004–2008).We believe that our approach to modeling government surveillance disease data will be useful to advance current understanding of large-scale drivers of several diseases.
| Disease data collected by the government surveillance system are frequently used to understand the influence of large-scale phenomena (e.g., climate) on human health because these data often have a large temporal and/or geographical span. The down side is that a) these data are often biased towards individuals that come to the health facilities (i.e., symptomatic individuals); and b) the number of individuals examined can vary substantially regardless of concurrent changes in prevalence or incidence (e.g., due to shortage of personnel or supplies in health facilities), directly impacting the number of disease cases detected. Current modeling approaches typically ignore these peculiarities of the government data. Furthermore, current approaches do not take into account that observations directly influence disease dynamics since individuals with a positive diagnosis are often subsequently treated for the disease. In this article, we develop a novel model to circumvent these shortcomings and apply it to simulated data, highlighting how inference on infection incidence and prevalence might be misleading when some of the issues mentioned above are ignored. Finally, we illustrate this model using malaria data from the Brazilian Amazon, revealing the strong role of precipitation on infection prevalence seasonality and striking patterns in infection incidence.
| Current best practices regarding the collection of disease data consist in the unbiased sampling of individuals (e.g., through aggressive active case detection; [1], [2]) using the most sensitive pathogen detection method available (e.g., polymerase chain reaction (PCR) for malaria). This type of individual-level data has provided important information regarding infection and disease (symptoms+infection) prevalence and risk factors; however, these data are costly and thus tend to be spatially and temporally restricted, curtailing their ability to detect important disease drivers that vary over long temporal and large spatial scales. Studies that focus on large geographical and/or long temporal-scale disease drivers typically rely on government-based surveillance data (e.g., malaria [3]–[6], cholera [7], [8], measles [9], [10], american cutaneous leishmaniasis [11], pertussis [12], meningitis [13], and dengue [14]). While government-based surveillance data provide a wealth of information on disease, these data are often collected opportunistically, which may severely bias inference drawn from these data [e.g.], [ 15,16]. For instance, individuals routinely sampled by the government health facilities are often symptomatic [17], [18]. As a result, if part of the population is infected but asymptomatic, infection prevalence for the overall population cannot be estimated as if these data came from a random sample (i.e., the number detected to be infected divided by number of tested individuals) nor as if all infected individuals had been detected (i.e., the number detected to be infected divided by total population size). Similarly, the number of individuals that seek help at a particular health facility may fluctuate considerably with time regardless of concurrent changes in infection prevalence or incidence (e.g., due to increases in catchment area, or a shortage of personnel or supplies), directly affecting the number of observed disease cases. Unfortunately, past analyses have typically considered only the number of disease cases per unit time (e.g., weekly or monthly), ignoring the total number of individuals examined per unit time (but see [19]).
The standard approach to analyze time-series data from the government surveillance system is to search for trends [e.g.], [ regression analysis]; [ 3], [4], [11], [20]–[23] or scales of variability [e.g.], [ wavelet analysis]; [ 10], [24–26] that match those of the explanatory variables. Recent work, however, has increasingly employed sophisticated statistical models, typically within the state-space modeling framework, to fit mechanistic disease dynamics models [e.g.], [ 7], [9], [27]–[32]. An important assumption within these state-space models is that observations provide information about the states but do not affect the underlying process. In the particular context of disease dynamics, the assumption is that the number of individuals diagnosed with a particular disease provides information on infection incidence or prevalence but does not influence disease dynamics (the underlying temporal process). This is a valid assumption if tested individuals are not informed about test results nor treated for the disease (e.g., data consist on the number of deaths due to a particular disease). However, this assumption is violated if individuals that have a positive diagnosis are subsequently treated for the disease because treatment decreases the pool of infected individuals and thus affects disease dynamics.
Here we refine the state-space framework to overcome the shortcomings we have described. Our approach scales-up the results from a detailed individual-level study to allow unbiased inference on infection prevalence from government-based syndromic surveillance data over larger geographical and longer temporal scales than would be possible using solely the individual-level data. Our approach also properly accounts for changes in sampling effort and the number of individuals diagnosed/treated for the disease and makes use of several short time-series (rather than one long time-series) to infer changes in infection prevalence and the drivers of these changes. While some of our assumptions are tailored to malaria, the general approach we put forth should be adaptable to other human diseases.
We start our article by describing our data and the model we are proposing. We then use a ten-fold cross-validation exercise to show that the proposed model has a better out-of-sample predictive performance than a more phenomenological statistical model and a more mechanistic disease dynamics model. Next, we employ simulated data to show how inference on disease incidence can be severely distorted if one does not take into account concurrent changes in sampling effort and that observations affect disease dynamics. Finally, we illustrate our model by applying it to real malaria data from the western Brazilian Amazon.
Malaria health posts are the only source of antimalarial medication in the Brazilian Amazon and this medication can only be obtained with a positive malaria exam result. As a result, data from these health posts provide considerable information regarding changes in malaria prevalence and incidence, being the basis of the malaria surveillance system in Brazil [33]. The malaria data we use arise from the Brazilian surveillance network in three counties (Acrelandia – AC, Placido de Castro – PC, and Senador Guiomard – SG) in Acre state, western Brazilian Amazon. These data are aggregated by week t and county l. Over the entire 2004–2010 period, there were approximately 160,000 malaria tests, from which ∼20,000 were positive (Figure 1). In this dataset, individuals are sampled and tested for malaria (through microscopy) either because they believed they had malaria and sought help at the local government health facility (passive case detection) or because they were symptomatic when health agents visited their houses (active case detection). In either case, individuals tend to be predominantly symptomatic.
Let and be parameter sets containing the process and observation parameters, respectively. To draw samples from the posterior distribution of our latent states and parameter sets and , we need to determine up to a proportionality constant. Our approach adopts a slightly different factorization than the one used in the standard state-space models because the disease dynamics process depends on the observations from the previous time step. Here is our factorization:
The posterior distribution of the states and parameters is obtained by Gibbs sampling. We use Metropolis-within-Gibbs sampling steps for all states and parameters due to the lack of a closed form expression for the full conditional distributions. Convergence of our Monte Carlo Markov Chain (MCMC) algorithm was evaluated using trace-plots. All analyses and figures were created using R version 2.13.2 [55].
We compare the out-of-sample predictive ability of the proposed model (eqns. 8 and 11) with that of two alternative models. The first model is a phenomenological state-space model, where the latent states follow an AR-1 temporal process, while the second model is a mechanistic Susceptible-Infectious-Susceptible (SIS) model. The goal here is to compare the proposed model to models that would typically be proposed by a statistician (AR-1 process on latent states) or by a mathematical biologist (SIS disease dynamics model). Details regarding the AR-1 and the SIS models are given in Text S1.
To determine the out-of-sample predictive performance of these three models, we conduct a 10-fold cross-validation exercise. First, we randomly partition our dataset into 10 sets. Then, we exclude one of these sets and use our algorithms to predict it based on information from the nine remaining sets. We compare the performance of these models by determining their mean squared error (MSE, a standard model comparison measure that takes into account both bias and variance of estimators), where lower MSE values are preferred.
Our ten-fold cross-validation exercise (i.e., prediction of 10% of the real malaria dataset using the other 90% of the data to train the model) revealed that the proposed model had a consistently better out-of-sample predictive performance when compared to the phenomenological AR-1 state-space model and the mechanistic SIS disease model (Table 3). In particular, the SIS disease model had a substantially worse MSE when compared to the other two models, revealing the negative impact of not allowing for process uncertainty. Based on these cross-validation results, we just report on the results from the proposed model from here onwards.
Using the out-of-sample results, we indeed find that the proposed model fitted well the weekly number of malaria cases (Figure 4). The 95% credible intervals tended to include most of the out-of-sample observations, both in terms of the total number of positive malaria exams (left panels in Figure 5) and the proportion of positive exams (right panels in Figure 5), indicating that uncertainty was adequately represented.
Simulated data using eqns. 8 and 11 show that trends in the number of malaria cases do not necessarily correspond to equivalent trends in infection prevalence or incidence. For instance, increasing number of malaria cases does not necessarily imply increases in infection prevalence (left panels in Figure 6). Similarly, decreasing number of malaria cases might just reflect decreases in the number of individuals examined, rather than decreases in infection prevalence (middle panels in Figure 6). Finally, trends in the number of malaria cases do not imply similar trend neither in infection prevalence nor in infection incidence (right panels in Figure 6). These simulation results are intuitive if we recognize that the expected number of disease cases depends both on infection prevalence and on the total number of sampled individuals (i.e., in eqn. 8). As a consequence, inference on infection prevalence or incidence based solely on the number of positive exams (i.e., ignoring the number of individuals examined) might lead to spurious conclusions.
The importance of allowing observations to directly affect disease dynamics is also illustrated using simulated data. We created a mock dataset where the number of malaria cases, the number of individuals examined, and infection incidence all exhibit the same temporal pattern (Panels A, B and D in Figure 7, respectively). As a result of the cancelling effect of greater number of individuals being treated precisely when infection incidence is higher, infection prevalence remains relatively constant (Panels C in Figure 7).
We then estimated infection prevalence and incidence using our original model (eqns. 8 and 10) and compared the resulting inference to that of a similar model that ignores that the observations (i.e., number of treated individuals) decreases infection prevalence. To implement this assumption, we modify equation 10 as(10a)
Assuming that the observation parameters are known, both the original model and this alternative model inferred well the underlying infection prevalence (top six panels in Figure 8) but led to substantially different inference on infection incidence (bottom two panels in Figure 8). In particular, the original model correctly inferred infection incidence (bottom right panel in Figure 8) while the alternative model inferred an infection incidence of approximately zero (bottom left panel in Figure 8). The intuition for these results is simple. If the number of individuals being treated is changing but the inferred infection prevalence remains constant, this has to imply that the number of individuals being treated is precisely off-setting infection incidence. On the other hand, since the alternative model does not take into account the fact that treated individuals decrease prevalence, an estimated constant infection prevalence implies zero incidence. These results highlight the problem of ignoring that individuals treated for the disease directly influence disease dynamics.
The depiction of the real data in Figure 1 already illustrates that sampling effort exerts considerable influence on the number of positive test results. For instance, the correlation between the number of exams and the number of disease cases was equal to 0.71 in our malaria dataset. Furthermore, there is considerable variation through time in the number of individuals that are examined. Thus, the common assumption that sampling effort is constant is likely to be unrealistic, particularly given the length of many of the disease time-series typically employed, such as those used to detect the effect of climate change on disease. As a result, analyses that rely solely on trends in the number of positive exams may generate misleading conclusions regarding disease dynamics.
Our estimates of infection prevalence reveal a relatively high initial infection prevalence (mean infection prevalence from 2004 to 2008 was 4%, with 95% credible interval (CI) of 3%–5%) with large seasonal outbreaks, which was then followed by a substantial decline in prevalence (mean infection prevalence for 2008–2010 was equal to 0.9% with 95% CI of 0.8–1.1%) (red line and polygon in Figure 9). A large increase in infection incidence seems to occur immediately after the rainy season, leading to subsequent peaks in infection prevalence (which can be as high as 18%) during the dry season, although there is considerable variability both geographically (from county to county) and temporally (year to year). A quantitative measure of association between prevalence and rainfall can be obtained using a permutation test, akin to the ones described in [23]. In this test, we compare precipitation when infection prevalence was at its highest versus at its lowest, for each year and location, yielding 21 (7 years×3 locations) observations for each level of infection prevalence. Our permutation test strongly suggests that the observed difference in mean precipitation is highly unlikely under the null hypothesis of no association (p-value<0.01), consistent with the results from a large-scale analysis of malaria data spanning 7 states of the Brazilian Amazon, which found a negative correlation between precipitation and number of malaria cases [23].
The declining trend in infection prevalence may be attributed to a sharp decrease in incidence after week 210 (from 2007 to 2008, Figure 10); incidence in 2008 to 2010 was approximately 1/5 of the incidence in 2004 to 2007. This abrupt decrease in incidence does not seem to be associated neither with land use/land cover changes (e.g., fire, deforestation rate, and forest cover) nor with climate (e.g., Southern Oscillation index or Oceanic Niño Index) (data not shown). This decrease may be attributable to enhanced vector control activities but we lack data on these activities to test this hypothesis. Posterior distributions for the remaining model parameters are given in Text S1.
We have described a novel model that circumvents some of the shortcomings of earlier modeling approaches. For example, our model is able to estimate infection prevalence despite the biases associated with government surveillance data by up-scaling information from a detailed individual level study. This capability of our model is particularly important for public health, where estimates of infection prevalence (rather than disease prevalence) are vital for disease control and elimination strategies. The ability to build on individual-level data (unbiased but geographically limited and costly) to extract information from the government surveillance data (geographically extensive but often biased) is likely to be important for the modeling of data from several other diseases. In particular, it reveals the potential benefits of coordinating careful individual level data collection with the modeling of large-scale patterns using government data. However, for this strategy to work well, it is critical that the collection of individual level data is done so that the results are representative for the region and time-frame of interest.
Disease dynamics model are typically more complex than the model we have presented here, including age structure of the host population, vector dynamics, multiple parasites and strains, and an exposed state. Models containing these additional complexities, however, are rarely fitted to data, with parameters often simply assumed to be known [e.g., 32] or extracted from the literature [e.g.], [ 12], [44,46]. Attempts to fit these models directly to data often reveal that several parameters are unidentifiable [28], [32], [42]–[44], [46] or rely on equilibrium assumptions to estimate these parameters [e.g., 56]. Furthermore, these attempts typically assume either just observation error or just process stochasticity, but not both as our model [50]. Finally, these disease dynamic models have numerous simplifying assumptions of their own, which may lead to substantially different conclusions [47], [48]. For these reasons, we have chosen to employ a model that is not as phenomenological as a regression model or wavelet analysis (i.e., we employ a realistic observation model to infer the underlying infection prevalence and allow for prevalence to decrease with the treatment of individuals) nor mechanistic as disease dynamics models (e.g., we do not account for infection incidence being influenced by current infection prevalence). Cross-validation results suggest that our model may outperform more phenomenological methods (e.g., AR-1 state-space model) and more mechanistic disease models that do not account for process uncertainty (e.g., the deterministic SIS disease dynamics model) (Table 3).
The statistical literature has traditionally assumed that observations do not alter the phenomenon or object that is being measured or assessed. Yet, some types of time-series data can clearly violate this assumption. In our case, a high number of individuals diagnosed to have malaria has the dual-role of suggesting a high infection prevalence at a particular time and a substantial decrease in infection prevalence in the next time step, since these individuals are subsequently treated for the disease. A similar example refers to the use of the number of carcasses encountered or harvested animals as a proxy for animal abundance [57], [58]. The model we propose explicitly accounts for the fact that observations (i.e., the number of individuals diagnosed and then treated for the disease) influence the underlying temporal process (i.e., infection prevalence dynamics), thus modifying the usual state-space approach. Using simulated data, we show that this characteristic is critical when inferring infection incidence (bottom two panels in Figure 8). When applied to the real malaria data, this model characteristic has allowed the identification of pronounced seasonal and long-term trends on infection incidence and prevalence, which might be associated with rainfall. The importance of letting observations affect disease dynamics depends on the nature of the observations. For instance, we believe this is an important problem that has been overlooked in previous malaria models [28], [29]. On the other hand, this feedback of observations on the disease dynamics might not be necessary if the observations consist on the reported number of deaths attributed to a particular disease [e.g.], [ 7,27]. In this case, observations can be modeled simply as a fraction of the true number of individuals that died and left the infected pool.
The proposed model also accounts for sampling effort (i.e., number of individuals sampled), an important characteristic that is surprisingly absent from the disease modeling approaches we know of, mechanistic or not. For example, there has been considerable contention regarding the role of climate change on the increasing number of malaria cases in the African highlands [3], [22], [59]–[61]. Could an increasing trend in sampling effort be a simple explanation for the observed trend in number of malaria cases? Simulated and real data suggest that the effect of sampling effort might be substantial (e.g., Figure 1 and Figure 6), which may be particularly important given the long-term nature of most of the time-series used for disease dynamics modeling [30]. Similar examples highlighting how changes in detection probability and health treatment seeking behavior can distort inference on disease dynamics are also given by [46], [50]. Finally, the lack of more long time-series has been blamed for the considerable uncertainty regarding how climate and other environmental drivers affect disease [29], [30], [62], [63]. Instead of relying on long but rare disease time-series, our model utilized multiple short time-series to infer on the effect of climate on disease dynamics.
In summary, we have focused on three aspects that have typically been ignored by earlier modeling approaches, namely: a) changes in sampling effort (i.e., total number of individuals examined), b) the fact that government surveillance data are often biased towards symptomatic individuals, and; c) the fact that observations (i.e., individuals diagnosed and subsequently treated for the disease) often directly influence disease dynamics by decreasing infection prevalence. We note that the relevance of these aspects fundamentally depends on the particular disease and data that are being analyzed; yet, we highlight them because they (to the best of our knowledge) are overlooked in the literature, either individually or jointly. Furthermore, we emphasize that these shortcomings are not restricted to state-space models; they may occur in other modeling approaches as well. We believe that some of these problems are a legacy from the biomathematical origins of these disease dynamics models. Researchers employing these models have traditionally focused on studying the long-term behavior of this complex non-linear system, thus relying on parameters from the literature or on rough parameter estimates [64]. However, as the focus shifts to parameter estimation and quantitative disease prediction, greater attention will be needed regarding how disease data arise and how to properly estimate parameters from it.
Our modeling approach has five important limitations. First, the proposed model conditions on the total number of exams at each time and county. By doing so, we avoid having to worry about factors that influence the total number of individuals examined, such as the opening of new health facilities, temporary lack of personnel, or shortage of supplies. However, this feature of our model precludes future predictions of future infection prevalence. This limitation can potentially be avoided by creating an additional model to predict the total number of exams. Second, we rely on individual level data to correct for the biased nature of the government surveillance data but individual level data might not be available or might not be representative of the geographical or temporal scale of the aggregate data. In this case, data from the literature might be used in place of the individual level data to create informative priors on the observation model parameters. Third, our observation model assumes that a) symptom status is binary whereas, in reality, there is often a whole spectrum of symptoms [53], which may in turn influence the probability of sampling the individual and detecting the pathogen; and b) that the probability of symptoms given infection does not change with time. These assumptions may or may not be reasonable for other diseases and we believe that changing our observation model to accommodate for alternative assumptions, without compromising the ability to fit the model, is an important topic for future research.
Fourth, our process model does not take into account the nonlinearities in disease transmission that are the hallmark of disease dynamics models. As noted before, it remains an important challenge to estimate parameter for these biologically inspired disease dynamics models, particularly if one is willing to take into account process uncertainty and a more realistic observation model. Finally, our results suggest large and relatively abrupt changes in infection incidence (Figure 10), which may not be realistic. Future research could focus on developing methods to infer smooth changes in infection incidence.
In this article, we have conceptualized and implemented a model that takes into account how data arise and affect prevalence dynamics. While the exact model formulation (e.g., eqns. 8 and 11) was tailored to the available data and current understanding regarding malaria, the main contribution of this article is to shed light on the importance of a few shortcomings of current disease modeling approaches and to suggest some general strategies to overcome them. We believe that these features have the potential to considerably improve inference on the drivers of disease dynamics when using government surveillance data.
|
10.1371/journal.pcbi.1007258 | Reappraising the utility of Google Flu Trends | Estimation of influenza-like illness (ILI) using search trends activity was intended to supplement traditional surveillance systems, and was a motivation behind the development of Google Flu Trends (GFT). However, several studies have previously reported large errors in GFT estimates of ILI in the US. Following recent release of time-stamped surveillance data, which better reflects real-time operational scenarios, we reanalyzed GFT errors. Using three data sources—GFT: an archive of weekly ILI estimates from Google Flu Trends; ILIf: fully-observed ILI rates from ILINet; and, ILIp: ILI rates available in real-time based on partial reporting—five influenza seasons were analyzed and mean square errors (MSE) of GFT and ILIp as estimates of ILIf were computed. To correct GFT errors, a random forest regression model was built with ILI and GFT rates from the previous three weeks as predictors. An overall reduction in error of 44% was observed and the errors of the corrected GFT are lower than those of ILIp. An 80% reduction in error during 2012/13, when GFT had large errors, shows that extreme failures of GFT could have been avoided. Using autoregressive integrated moving average (ARIMA) models, one- to four-week ahead forecasts were generated with two separate data streams: ILIp alone, and with both ILIp and corrected GFT. At all forecast targets and seasons, and for all but two regions, inclusion of GFT lowered MSE. Results from two alternative error measures, mean absolute error and mean absolute proportional error, were largely consistent with results from MSE. Taken together these findings provide an error profile of GFT in the US, establish strong evidence for the adoption of search trends based 'nowcasts' in influenza forecast systems, and encourage reevaluation of the utility of this data source in diverse domains.
| Google Flu Trends (GFT) was proposed as a method to estimate influenza-like illness (ILI) in the general population and to be used in conjunction with traditional surveillance systems. Several previous studies have documented that GFT estimates were often overestimates of ILI. In this study, using a recently released archive of data of provisional incidence from a large surveillance system in the US (ILINet), we report errors in GFT alongside errors from ILINet’s initial estimates of ILI. This comparison using information available in real-time allows for a more nuanced assessment of GFT errors. Additionally, we describe a method to correct errors in GFT and show that the corrected GFT estimates are at least as accurate as initial estimates from ILINet. Finally, we show that inclusion of corrected GFT while forecasting ILI in the next four weeks considerably improves forecast accuracy. Taken together, our results indicate that the GFT model could have added value to traditional surveillance and forecasting systems, and a reevaluation of the utility of the underlying search trends data, which is now more openly accessible, in fields beyond influenza is warranted.
| Surveillance of seasonal influenza and other respiratory illnesses deservedly receives significant attention from public health agencies in the United States. To complement traditional surveillance systems, both internet- [1–7] and non-internet-based [8–11] proxy indicators of incidence have been developed. Among these, of note is Google Flu Trends (GFT) [1, 12], which estimated influenza-like illness (ILI) from online search activity. GFT estimates from an initial model and subsequent revisions to the model were publicly available until 2015, when the service was discontinued [13]. Although Google has not offered reasons for the termination, one contributing factor could well have been the widely reported propensity of GFT to over-estimate ILI, which effectively morphed it in the public perception from a poster child for the power and utility of big data to one of its hubris [14–20].
However, this perception is probably misplaced. The most comprehensive and commonly cited study of GFT errors for locations in the United States was published by Lazer et al [14], following an anomalous season during which the errors were much larger than previously observed. These findings were supported by several other studies that were smaller in scope but reported errors of approximately the same magnitude at different locations and geographical resolutions [21, 22]. In this paper, using newly available surveillance data, we revisit GFT estimates for locations in the US and show that its errors are less substantial than previously reported.
The severity of a respiratory viral infection in an individual depends on multiple factors, and in most cases the symptoms are mild and do not require medical attention. As a consequence, the more widely used surveillance systems in the US–the Centers for Disease Control and Prevention (CDC)’s ILINet and FluSurv-NET systems, for example–only capture infections that are severe enough to precipitate a visit to a physician's office or hospital. On the other hand, the relationship between the severity of a respiratory infection and the likelihood that an individual initiates an online search session for related information, is unknown; hence, the signals that drive GFT and the surveillance systems are intrinsically different. Nonetheless, as GFT used incidence data from ILINet as its response variable, it has been a common practice, and one that we follow in this study, to use these rates as reference or ground truth when reporting the accuracy of GFT estimates.
However, in reporting US national and regional errors, most previous studies, including Lazer et al [14], did not account for delayed reporting to ILINet. The fully observed ILINet rates (ILIf) are finalized no sooner than 2–3 weeks after the conclusion of a surveillance week, as some of the surveillance network data are submitted late, and in some instances, revisions can even occur several months later. The rates released in the interim are estimates based on partial observations (ILIp) and the magnitude of difference between ILIp and ILIf, as we report here, varies by location, influenza season and the phase of a season.
Given these reporting delays and revisions, it is ILIp rather than ILIf that informs real-time decisions. Thus, a more appropriate error analysis, one that better reflects an operational scenario, should compare errors of GFT (ILIf—GFT) to errors of ILIp (ILIf—ILIp). An archive of ILIp at US national and Health and Human Service (HHS)[23] regional levels for 6 seasons has been made available [24] recently, and in this study we used this archive to recompute GFT and ILIp errors[25, 26]. Additionally, we extended the analysis to finer geographical resolutions as ILIf is now also available for US states. Finally, we report for the first time, errors from the final GFT model, updated in fall 2014, before the service was discontinued.
Google's recent initiative to provide access to its search trends through an API[13] supports more open data sharing. This effectively decouples data from model and facilitates the development of alternative models to GFT. Through the analysis described here, we hope to establish an error profile of GFT that can serve as a baseline for comparing these alternative models.
More importantly, although GFT was proposed by its developers as a supplement to traditional surveillance systems and not a replacement, the focus to date has been disproportionately on evaluating GFT's ability to mimic surveillance systems rather than on evaluating its utility when deployed in conjunction with these systems in operational settings. Previous findings suggest that the errors in GFT can be reduced by combining GFT estimates with lagged surveillance rates [14, 27, 28]. Here we propose a similar remedial step with a parsimonious regression model and show that the corrected GFT is more accurate than ILIp.
A natural extension is to assess whether GFT, its errors thus corrected, could have improved longer term forecasts by providing more timely outbreak information than traditional surveillance systems. For this purpose, we generated forecasts of ILI one to four weeks in the future using ILIp alone, and using both ILIp and error corrected GFT. We demonstrate that the inclusion of GFT considerably improves the accuracy of near-term forecasts and thus adds value to traditional surveillance systems.
In this section we describe in detail the two data sources used—an outpatient surveillance system and GFT—access information for the two sources, and the measures used to calculate errors of these estimates. We then describe the autoregressive model framework used to generate near term forecasts, followed by details of the forecast generation and validation process.
The ILINet surveillance system [29], developed and supported by the CDC, collects data from nearly 3000 healthcare providers in the US on outpatient visits for ILI, which is defined as fever (temperature above 100°F) co-occurring with cough and/or sore throat. Weekly counts of patients seen for ILI and for any reason are submitted to the system. These count data are used to calculate the percentage of outpatient visits due to ILI. In this study, by ILI rate we refer to population-weighted aggregates of ILI.
A Morbidity and Mortality Weekly Report (MMWR)[30] surveillance week runs from Sunday thru Saturday and aggregated ILI rates at US state-, HHS regional- and national levels are publicly released through the FluView [31] website on Friday (6 days after a week concludes). The system allows for delayed reporting from providers and the delayed data are included in subsequent weekly releases. Hence, the ILI estimates for a week can change for multiple weeks following initial release. We refer to the ILI rates calculated from incomplete reporting as partially observed ILI rates, and in this paper denote the rates as per the first week of release as ILIp. An archive of revisions for the 2009/10 season onwards has been recently made available [24, 32] and for the 2013/14 season and later, these data are also accessible through the DELPHI group's epidata API [33].
Although ILIf is available for the US, HHS regions and states, ILIp is not currently available at the state level. ILI rates for the 2009/10 to 2014/15 seasons that were available on FluView at the end of surveillance week 20 of the 2017/18 season (May 13–19, 2018) were considered to be ILIf. This date is over two years after the end of the time period studied, and hence we assume that it is very unlikely that these rates would be further revised. Note that both ILIp and ILIf are rates, and ILIp can over or underestimate ILIf.
Originally developed in 2008, GFT estimated ILI rates in a population based on the frequency of a selected set of queries to the Google search engine [1]. The 2008 model used 45 queries, whose search frequencies were historically well correlated [34, 35] with ILI rates, as explanatory variables. To generate the estimates for the US, ILI rates were used as the response variable in the model. In response to observed deficiencies in the predictions, revisions to the model, including updates to the feature set, were made in 2009, August 2013 and August 2014. GFT estimates that were published in real-time from September 2008 through August 2015, along with estimates from revised models applied to past seasons continue to be hosted publicly [12].
Fig 1 shows the availability of GFT, ILIf and ILIp at different locations in the US. For US and HHS regions, GFT, ILIf and ILIp are available for six seasons—2009/10 to 2014/15—and for the states ILIf and GFT are available for the last 5 of these 6 seasons. The vertical lines indicate the time points of revisions to the GFT model; therefore estimates for seasons 2009/10 thru 2012/13 seasons, season 2013/14, and season 2014/15 are from different models.
Unlike ILINet, GFT estimates for a week are finalized at the end of the week. Furthermore, as the GFT estimates were completely automated, and computed in real-time, they did not have the 6-day lag between the end of a week and the release of data as is the case with ILINet. This translates to GFT providing weekly incidence estimates for at least one more week than ILINet, at any given point of time. The estimate for this one additional week is sometimes referred to as a nowcast.
For each week and location, error is defined as y−y^, where y is the reference, ILIf, and y^ the estimate from GFT or ILIp. Aggregate error measures, Mean Squared Error (MSE), Mean Absolute Error (MAE) and Mean Absolute Proportional Error (MAPE) are respectively the mean of the square of errors, of the absolute error and of absolute error as a proportion of the reference value, and are reported across all seasons and locations, or for each season (across all location) and each location (across all seasons). During the study period, the reference value was never zero, and hence MAPE was computable. Formally,
MSE=1n∑i=1n(yi−y^i)2
MAE=1n∑i=1n|yi−y^i|
MAPE=1n∑i=1n|yi−y^i|yi
As the errors in 2012/13 are reportedly much larger than during the other seasons included in the study, inclusion of this season could obscure overall results, and hence we report aggregate measures both with and without this season.
A non-seasonal ARIMA model is specified by three parameters—p, the order of the autoregressive component; q, the order of the moving average component; and d, the degree of differencing required to make the given time series stationary. For a time series, Y, let y denote the time series obtained by d degree differencing. Thus, an ARIMA(p, d, q) is a model of the form:
yt=c+ϕ1yt−1+⋯+ϕpyt−p+θ1εt−1+⋯+θqεt−q
where the elements, εi, represent the forecast errors at the ith time step. Elements c, ϕ1, …, ϕp, θ1, …,θq can be estimated through maximum likelihood estimation. As influenza in the US has strong yearly seasonality, a seasonal ARIMA model may often, though not always, be a better fit. Seasonal ARIMA models are specified with three additional parameters P, D, Q where D denotes seasonal differencing and P, Q are analogous to p, q, respectively, as defined above.
We used an implementation of an iterative method proposed by Hyndman and Khandakar [36] from the R [37] forecast [38] package to find an appropriate order for the ARIMA models. Briefly, the Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test [39] and extended Canova-Hansen test [40] are used to determine an appropriate d and D respectively. To find values for the remaining parameters, an iterative process is initiated with the model that has the lowest Akaike’s Information Criterion (AIC) [41] amongst a small default set of models, as the candidate model. In each subsequent step, the parameters of the candidate model are varied by ±1 within a pre-specified parameter space (p, q: (0, 5); P, Q: (0, 2)) and the variant with the lowest AIC becomes the new candidate model. The process is terminated when the parameter space is exhausted or all variants of the candidate model result in a higher AIC.
Retrospective near-term forecasts were generated for US National and the 10 HHS regions during the 2010/11 to 2014/15 influenza seasons for MMWR weeks 41 through week 20. Traditionally an influenza season is considered to run from MMWR week 40 thru MMWR week 39 of the following calendar year. Late spring and summer weeks (MWWR week 20 onwards) experience low incidence and hence were excluded in this study. Separate models were fit for each location and week. Models for each location are isolated as they do not use observations from any other location.
Let Xi and Zi denote the log transformed ILI rates and GFT estimates at week i respectively. All ILI/GFT values less than 2 (per 100000) were rounded up to 2 before log transformation. As described in a previous section, when forecasts are generated operationally at the end of week t, X1,⋯,Xt and Z1,⋯,Zt+1 would be available; i = 1 indicates MMWR week 40 of 2009/10 season. For a given week w ≤ t, Xw is ILIp if w and t belong to the same season, and ILIf otherwise. Corrected GFT, Z^t+1, is estimated using a random forest [41–44] regression model with explanatory variables Xt,Xt−1,Xt−2,Zt+1,Zt,Zt−1,Zt−2 and response variable Xt+1. S4 Fig shows corrected GFT at the US national level, and its error with respect to ILIf.
To generate near term forecasts, two models were developed: the first using ILIp only (ILIp), and the second using ILIp and corrected GFT (ILIp+GFT). For a given week t, the ILIp models were trained on the time series X1, …, Xt and used to forecast rates for weeks t+1, …, t+4, denoted by X^t+1…X^t+4. The corresponding ILIp+GFT ARIMA models were fit using the time series X1,…,Xt,Z^t+1 and forecast rates for X^t+2…X^t+4.Z^t+1 doubles as the 1-week ahead forecast. For both models MSE, MAE and MAPE, as defined above, were calculated with ILIf as reference.
For example, the ILIp models for week 46 of 2011/12 season were fit using ILIf from the 2009/10 and 2010/11 seasons and ILIp from weeks 40 to 46 of the 2011/12 season, and were used to forecast rates for weeks 47 through 50. The GFT correction model for week 46 was fit using training instances compiled with ILI and GFT through week 46 of 2011/12 season and used to estimate, Z^47 with test instance (X46,X45,X44,Z47,Z46,Z45,Z44). ILIp+GFT models used Z^47 as an additional observation, and forecast rates for weeks 48 to 50. Therefore, the week 50 forecast from ILIp ARIMA model was a 4-week ahead forecast but a 3-week ahead forecast for the ILIp+GFT ARIMA model. Forecast errors for both model forms were then calculated using ILIf for weeks 47 to 50 as reference.
Table 1 shows that the MSE of GFT is on average 2.5 times that of ILIp with considerable variability by location. Region 9, where the mean squared errors were nearly equal, had the smallest difference between GFT and ILIp, whereas Region 4 had the largest difference, with GFT error about 7.6 times as large as that of ILIp. Similar variability was observed across seasons, with the largest difference by far occurring during the 2012/13 season, and the smallest during 2009/10. As previously reported [14], GFT estimates for weeks around the peak of the 2012/13 season were large over-estimates, which contributed considerably to the high mean errors.
The corresponding difference in MAPE (S1 Table) is slightly smaller overall (GFT error 1.8 times ILIp error), with the GFT error actually lower than that of ILIp for Region 9. In reporting Table 1 (and S1 Table) we excluded season 2012/13 for Overall and regional aggregations; see S2 Table for aggregations across all seasons.
Fig 2 shows MSE with the final version of the GFT system for the 2014/15 season and the average GFT errors in all regions (denoted by the black triangle) are larger than corresponding ILIp errors. But as indicated by the data points above the diagonal, ILIp does not consistently have lower errors for all weeks. As supported by S2 Fig, during weeks very early (blue data points) or towards the end (red data points) of the season, the difference between GFT and ILIp is relatively small (data points closer to the diagonal). The larger errors for both ILIp and GFT occur during weeks of increased ILI activity around the peak week (green and grey data points). S1 Fig has the corresponding MAPE errors for the 2014/15 season. On the whole, errors during the 2014/15 season are in line with some of the previous seasons, and the final GFT model was not a marked improvement over previous models.
Looking at GFT errors at the state-level (Fig 3, S3 Fig), the errors are much larger than the errors at the corresponding HHS regions (black horizontal mark). Overall (top left panel), states with low (< 2 million) and medium (2–6 million) population sizes, tend to have larger GFT errors than high population states. We know from previous work [5] with search trends from Google's Health Trends API, that terms/queries whose search frequencies do not meet a predetermined threshold limit are reported as 0. If GFT used a dataset that was based on similar criteria, low population states where search volumes are smaller, would have had sparser feature spaces.
Similar patterns were seen when the errors are disaggregated by season. It is interesting to note that among all seasons studied, the season with the smallest differential in MSE between state and regional errors was the anomalous 2012/13 season, where the large increase in GFT regional errors was not accompanied by a proportionate increase in errors for states. In a few cases, the errors for a state were smaller than the errors at the corresponding region.
As shown in Table 2 and S4 Fig, considerable reduction in GFT MSE was achieved through regression on lagged data. An overall reduction of 44% was observed across the 11 locations and 4 seasons. The large reduction during 2012/13 reiterates the utility of this additional step as a check against extreme failures of GFT. It is also interesting to note that this step reduces GFT errors below that of ILIp i.e. the use of search trend data can not only provide an estimate of incidence a week earlier than ILINet, but can do so more accurately than ILINet's own initial estimate of incidence. S3 Table shows the corresponding overall reductions in MAE and MAPE, and the findings noted with MSE hold.
There is considerable variability in the magnitude of improvement in nowcast quality across locations and seasons, and with a few exceptions the decrease in errors was significant (P < 0.05) per a paired Wilcoxon signed rank test [45–47].
Table 3 shows the MSE for near-term forecasts generated with ILIp alone and using both ILIp and corrected GFT (ILIp+GFT). At all targets (1- to 4-week ahead estimates) and seasons, and for all but two regions (Fig 4), inclusion of GFT lowered MSE. The overall MAPE with the ILIp+GFT models is also lower (S4 Table, S5 Fig), although the relative advantage over ILIp with different regional or seasonal disaggregation criteria is more mixed. The overall reduction in errors when aggregated by target or region is not limited to reduction from the anomalous 2012/13 season; ILIp+GFT errors continue to be lower and significant when the 2012/13 season is excluded (S5 Table).
For all three measures, the accuracy of the regression model's nowcast either matches or exceeds that of the 1-week ahead ARIMA forecast. Reduction of errors at longer horizons is larger and this is quite likely due to the k week ahead forecast of the ILIp+GFT model being lined up with the k+1 week ahead forecast of the ILIp model, as ARIMA errors tend to increase with increasing horizon.
The increasing availability of big data has naturally led to the development of experimental applications in several domains, including those such as public health surveillance that have traditionally relied on more robust, but also labor intensive, data collection processes. Google Flu Trends was developed as an alternative method to measure ILI in the general population, to be used in conjunction with traditional surveillance methods when and where they exist. Given its prospects for use (and misuse) GFT appropriately received wide attention; but it is our belief that it has been adjudged wanting against goals it was not designed to meet.
Reporting errors of ILIp rates alongside GFT errors, helps quantify the transient errors in ILINet due to delayed reporting and provides a more appropriate baseline for comparing the accuracy of GFT (and alternative nowcast models) in operational settings. The use of ILINet rates as ground truth, here and in previous studies, is appropriate simply because these are the targets GFT was designed to estimate and a more reliable system for estimating ILI broadly in the US does not exist. However, when assessing the validity of alternatives methods for influenza estimation, we must remain cognizant of the deficiencies of ILINet in capturing influenza transmission at metapopulation scales–for instance, its passive data collection process, broad symptom definition that is geared towards ILI rather than influenza, and estimation of incidence from visit counts without a requirement for virologic confirmation.
The opening up of Google Trends API directly addresses one major obstacle in improving nowcasts over the GFT models, namely, the non-availability of public search trends data. Additionally, US state level ILINet rates were not available prior to the 2017/18 season, and previously required some form of extrapolation from regional ILI rates to state ILI rates in order to build state-level nowcast models. With these data now being released in real time, nowcast models for states should be able to identify more reliable predictor variables, and the accuracy of these nowcast estimates can be expected to improve over GFT estimates. Furthermore, fine-grained nowcast estimates, say at city or county scales, or for large hospital settings, are possible when reliable ILI rates exist.
Our results show that a regression model with lagged ILIp and GFT predictors can adequately correct errors in search trend based nowcasts and thereby avoid catastrophic failures, and the model estimated rates are at least as accurate as partially observed surveillance rates in the US. Indeed, during the 2017/18 and 2018/19 influenza seasons, which saw atypical, large, sustained outbreaks, our search trends based nowcasts did not exhibit large errors. Use of this data source alongside other data sources like twitter, electronic medical records, Wikipedia logs etc. [3, 4], can further reduce the risk of such failures by making the nowcasts less reliant on any single source.
Results with the near-term forecasts show that the provision of an additional week of observation to the ARIMA models considerably improves forecast quality. Forecasts generated with ILIp and corrected GFT also improve over those generated with ILIp and uncorrected GFT (S6 Table). Both the random forest and ARIMA models used here were standard implementations from open source statistical packages with no domain specific tailoring, and we have no reason to believe that these improvements and the ensuant findings are specific to the models used. Other mechanistic or time series models may offer similar improvements in accuracy, and some recent results are suggestive of such improvements [48, 49]. Our choice of ARIMA as the forecast model should not be construed as a vote in favor of its optimality in forecasting ILI; on the contrary, as ARIMA is not informed by any of the transmission dynamics of ILI, we include it as a non-naïve reference method. Researchers proposing alternative methods tailored for ILI should be expected to show that they do at least as well as ARIMA.
Overall, we believe that the results presented here provide sufficient evidence to encourage continued efforts to improve search trend based nowcasts for influenza and make a case for their more wide-spread adoption in operational forecasting systems. At a minimum, they show that reports of the failure of GFT are not unequivocal and they should not deter use of Google Trends API in areas other than ILI estimation.
|
10.1371/journal.pbio.2003467 | Disentangling metabolic functions of bacteria in the honey bee gut | It is presently unclear how much individual community members contribute to the overall metabolic output of a gut microbiota. To address this question, we used the honey bee, which harbors a relatively simple and remarkably conserved gut microbiota with striking parallels to the mammalian system and importance for bee health. Using untargeted metabolomics, we profiled metabolic changes in gnotobiotic bees that were colonized with the complete microbiota reconstituted from cultured strains. We then determined the contribution of individual community members in mono-colonized bees and recapitulated our findings using in vitro cultures. Our results show that the honey bee gut microbiota utilizes a wide range of pollen-derived substrates, including flavonoids and outer pollen wall components, suggesting a key role for degradation of recalcitrant secondary plant metabolites and pollen digestion. In turn, multiple species were responsible for the accumulation of organic acids and aromatic compound degradation intermediates. Moreover, a specific gut symbiont, Bifidobacterium asteroides, stimulated the production of host hormones known to impact bee development. While we found evidence for cross-feeding interactions, approximately 80% of the identified metabolic changes were also observed in mono-colonized bees, with Lactobacilli being responsible for the largest share of the metabolic output. These results show that, despite prolonged evolutionary associations, honey bee gut bacteria can independently establish and metabolize a wide range of compounds in the gut. Our study reveals diverse bacterial functions that are likely to contribute to bee health and provide fundamental insights into how metabolic activities are partitioned within gut communities.
| Honey bees are important pollinators that harbor a relatively simple gut microbiota with striking parallels to the mammalian system. This makes them relevant models to study gut microbiota functions and their impact on host health. We applied untargeted metabolomics to characterize metabolic changes induced by the gut microbiota and to characterize the contributions of the major community members. We find that the gut microbiota digests recalcitrant substrates derived from the pollen-based diet of bees. Most metabolic changes could be explained by the activity of individual community members, suggesting substrate specificity and independent metabolic functions. We did identify some cross-feeding interactions between species, including for pyruvate. Our study provides novel insights into the functional understanding of the bee gut microbiota and provides a framework for applying untargeted metabolomics to disentangle metabolic functions of gut bacteria.
| Metabolic activities of the microbiota are key for symbiotic interactions in the gut and impact health and disease of the host in manifold ways. Gut bacteria facilitate the breakdown of refractory or toxic dietary compounds [1–3], produce metabolites that promote host growth and physiology [4–7], and modulate immune functions in the gut [8] and other tissues [9,10]. Moreover, metabolic activity is the basis for energy and biomass production, resulting in bacterial growth and the occupation of ecological niches conferring colonization resistance against pathogenic microbes [11]. Substrates of gut bacteria predominantly originate from the diet of the host [2,12], making diet the major modulator of the composition and metabolic activity of the gut microbiota [13,14].
The substantial metabolic potential of the animal gut microbiota has been profiled by the direct sequencing of functional gene content (i.e., shotgun metagenomics) [15–18]. However, it is challenging to predict functional metabolic output from such sequencing data. With recent advances in the coverage and throughput of untargeted screening metabolomics [19–21], it has become feasible to quantify metabolic changes in microbiota or host tissues at large coverage and throughput. Besides identifying metabolites connected to human health and disease [22–30], untargeted screening metabolomics holds considerable promise to unravel metabolic functions of individual microbiota members in animals with divergent dietary preferences. However, such mono-colonization studies are complicated by the highly variable and species-rich composition of most animal microbiota. Thus, gut communities of reduced complexity are valuable models to disentangle metabolic functions of the constituent species.
Like mammals, honey bees harbor a highly specialized gut microbiota. However, in contrast to mammals, the honey bee gut microbiota is surprisingly simple and consistent, with seven species (categorized by clustering at 97% sequence identity of the 16S rRNA) accounting on average for >90% of the entire gut community in bees sampled across continents [31]. This microbiota is composed of four Proteobacteria (Gilliamella apicola, Snodgrassella alvi, Frischella perrara, and Bartonella apis), which mostly reside in the ileum, and two Firmicutes (Lactobacillus spp. Firm-4 and Firm-5) and one Actinobacterium (B. asteroides), which are predominantly found in the rectum. These specific locations suggest that bacteria occupy different metabolic niches in the bee gut and potentially engage in syntrophic interactions [32,33].
The honey bee gut microbiota has marked effects on the host. It promotes host weight gain and hormone signaling under laboratory settings [34] and stimulates the immune system of the host [35,36]. In addition, honey bees are ecologically and economically essential pollinators that have experienced increased mortality in recent years [37,38], which could in part be due to disturbances of their microbiota composition [39–42].
Genomic analyses and in vitro experiments have shown that fermentation of sugars and complex carbohydrates (e.g., pectin) into organic acids [15,32,43,44] is a prominent metabolic activity of the gut microbiota [34]. Lacking, however, is a detailed understanding of the consumption of diet-derived substrates and how individual community members contribute to the metabolic activities in vivo. For instance, it is elusive whether analogously to mammals, recalcitrant dietary compounds (especially from pollen) are broken down by the microbiota in the hindgut (i.e., large intestine composed of ileum and rectum), while more accessible compounds are reportedly absorbed by the host in the midgut (i.e., small intestine) [45–47].
To profile the metabolic output of the honey bee gut microbiota and its individual members, we employed gnotobiotic bee colonizations and in vitro experiments in conjunction with untargeted metabolomics (Fig 1). We first characterized robust metabolic differences between microbiota-depleted bees and bees colonized with a reconstituted community composed of the seven major bacterial species of the gut microbiota. Subsequently, we analyzed bees colonized with each community member separately to assay their potential contribution to the overall metabolic output of the gut microbiota. Finally, we recapitulated our results in vitro using pollen-conditioned medium. Our systematic approach provides unprecedented insights into the metabolic activities of the honey bee gut microbiota and demonstrates the possibility to use metabolomics in combination with gnotobiotic animal models to disentangle functions of individual gut microbiota members.
To characterize the metabolic output of the honey bee gut microbiota, we colonized newly emerged bees with selected bacterial strains previously isolated from the bee gut. The reconstituted bacterial community consisted of 11 strains (S1 Table) covering the seven predominant species of the bee gut microbiota described above. We used two strains for G. apicola and four strains for Firm-5 in order to cover the extensive genetic diversity within these species [44,48]. Exposure of newly emerged adult bees to this community resulted in the successful establishment of all seven species, with a total of approximately 109 bacterial cells per gut after 10 d of colonization; hereafter, these are referred to as CL bees (Fig 2A). In contrast, non-colonized bees had total bacterial loads of <106 cells per gut, an observation consistent with previous studies [32,49]. In the following, we will refer to these bees as microbiota-depleted (MD) because they were not colonized with detectable levels of typical honey bee gut bacteria as determined with species-specific qPCR primers (<105 bacterial cells, except for one bee that was slightly above this cut-off for Firm-5) (Fig 2B). However, these bees may have harbored low levels of environmental microbes, as they were not kept under sterile conditions, especially in cases in which the bacterial loads were slightly above our detection limit of 105 bacterial cells (Fig 2A). It is also important to point out that newly emerged bees can occasionally be contaminated with specific bee gut bacteria, resulting in “MD” bees that in fact are colonized. Therefore, to be able to exclude such bees from further analysis, it is essential to determine the microbiota status of gnotobiotic bees using the qPCR assays presented in this study or an equivalent method.
Compared to hive bees of the same age, bacterial abundances of most species were slightly elevated in CL bees. However, in both groups the Firm-5 species was consistently the most abundant community member, while B. apis colonized at relatively low levels. This is in line with recent 16S rRNA gene-based community analyses [31,50,51], and we thus conclude that the selected strains assembled into a structured community resembling the native honey bee gut microbiota. Overall, this analysis validates our gnotobiotic bee system as a tool for microbiota reconstitution experiments and enables the study of microbiota functions under controlled laboratory conditions.
To reveal microbiota-induced metabolome changes in the gut, we dissected the combined mid- and hindgut of MD, CL, and hive bees and analyzed water-extracted homogenates of these gut samples by untargeted metabolomics [21]. In total, we detected 24,899 mass-to-charge features (ions), 1,079 of which could be annotated by matching their accurate mass-to-sum formulas of compounds in the full Kyoto Encyclopedia of Genes and Genomes (KEGG) database (S2A Data). These 1,079 ions putatively correspond to 3,270 metabolites (S2B Data), since this method cannot separate isobaric compounds. For statistical analysis, we continued with the annotated ions, and for ion changes with multiple annotations, we provided the most likely annotation based on information from literature and genomic data.
Principal component analysis on the ion intensities revealed that CL and MD bees separate into two distinct clusters, which suggests colonization-specific metabolic profiles (S1 Fig). In two independent experiments, a total of 372 ions exhibited significant changes between CL and MD bees (Welch’s t test, Benjamini and Hochberg adjusted [BH adj.] P ≤ 0.01, S2C Data). A subset of 240 ions (65%) were more abundant in MD bees, suggesting that the cognate metabolites are utilized by the gut microbiota. These ions are hereafter referred to as bacterial substrates. Conversely, 132 ions were more abundant in CL bees and are hereafter referred to as bacterial products, indicating that they are produced either by the microbiota or by the host in response to the microbiota. To facilitate the biological interpretation of these multitude metabolic changes, we carried out two analyses. First, we looked at whether certain compound classes were overrepresented among the subsets of bacterial substrates and products (S3A Data). Second, we sorted ion changes based on their ability to explain the difference between the CL and MD metabolome profiles in an Orthogonal Projection of Least Squares-Differentiation Analysis (OPLS-DA) (Fig 3) [52].
We first focused on the 240 substrate ions that were more abundant in MD versus CL bees and potentially correspond to metabolites utilized by the microbiota (S2C Data). We found 3 compound classes to be strongly enriched: “flavonoids” (20 of 36 annotated ions, one-sided Fisher’s exact test, P < 0.001) and both “purine nucleosides and analogues” and “pyrimidine nucleosides and analogues” (in total eight of nine annotated ions, both P < 0.01). Seven flavonoids, three nucleosides, and a nucleoside precursor (orotate, m/z 155.009) were also among the 28 substrate ions with the most discriminatory power for distinguishing CL versus MD bees as based on OPLS-DA (Fig 3). Other ions among these most discriminatory substrates included two ω-hydroxy acids (m/z 315.254 and m/z 331.248) and three phenolamides (m/z 582.260, m/z 630.245, and m/z 233.129) from the outer pollen wall, as well as quinate (m/z 191.056) and citrate (m/z 191.019), both of which had previously been predicted to be utilized by certain community members of the honey bee gut microbiota [53,54]. Because they are the most remarkable groups among the identified substrates, nucleosides, flavonoids, and pollen wall-specific compounds will be discussed in more detail.
We next looked into the 132 ions that were more abundant in CL versus MD bees and thus represent possible metabolites produced by the microbiota (S2C Data). Again, we used enrichment analyses and OPLS-DA (Fig 3) to prioritize the most important product ions. Three compound classes were to some extent enriched among the bacterial products (S3A Data): “carboxylic acids and derivatives” (seven of 26 annotated ions, one-sided Fisher’s exact test, P < 0.03), “fatty acids and conjugates” (seven of 29 annotated ions P < 0.05), and “eicosanoids” (five of eight annotated ions, P < 0.01).
To assess how much of the total metabolic output can be identified in hive bees under natural conditions, we analyzed the gut metabolome of 10-d-old hive bees that were exposed to social interactions and natural dietary resources and were colonized by the native gut microbiota. Principal component analysis revealed that hive bees clustered separately from CL bees (S1 Fig). This may be explained in part by the different diet of hive bees, the presence of multiple strains in a bee colony, and the impact of the environment on the gut metabolism. However, we found that 27 of the 28 most discriminatory substrate ions and 15 of the 22 most discriminatory product ions showed qualitatively the same changes in hive bees as in CL bees (S2D Data, Welch’s t test, BH adj. P ≤ 0.01). On the substrate side, this included most flavonoid ions, all nucleosides, quinate, and citrate, as well as the ions annotated as ω-hydroxy acids and phenolamides from the outer pollen wall. On the product side, we found four of the five prostaglandins and one of the juvenile hormone derivatives to be significantly increased in hive bees relative to MD bees, suggesting that these host-derived metabolites are also induced under natural conditions. Moreover, ions corresponding to fermentation products were either significantly increased (sebacic acid and valerate) or showed a trend towards increased levels (succinate and pimelate) in hive bees. The same was the case for the four ions corresponding to possible degradation products of flavonoids (hydroxy- and dihydroxyphenylpropionate, maleylacetate, and hydroxy-3-oxoadipate; S2A Data). We conclude that the remarkable overlap of metabolic changes between hive bees and CL bees highlights the relevance of our findings.
We thus far presented evidence for substrates and products of the complete microbiota in the honey bee gut. To elucidate which community members might be responsible for these transformations, we conducted mono-colonizations of MD bees with all seven bacterial species (again using a mix of four and two strains together for Firm-5 and G. apicola, respectively). All species successfully established in the gut of MD bees, with other community members being generally below the limit of detection (<105 bacterial cells) (S5 Fig). We again extracted metabolites from the mid- and hindgut of individual bees to address how many of the 372 robust ion changes can be explained by one or multiple mono-colonizations (S2A Data), i.e., show qualitatively the same change as in CL bees (analysis of variance [ANOVA] followed by Tukey honest significant difference [HSD] post hoc test at 99% confidence, P ≤ 0.05) (S7 Data and S8 Data). Extended results of this analysis can be found in S2A Data.
Remarkably, using these significance cutoffs, 299 of the 372 (80%) robust changes between MD and CL bees could be explained by one or multiple mono-colonizations. This included 201 (84%) substrate and 98 (74%) product ions. The two Lactobacilli species (Firm-5 and Firm-4) explained most changes, followed by B. asteroides and the two Gammaproteobacteria (Fig 4A). Interestingly, the relative contribution to substrate conversion and product accumulation varied between mono-colonizations. For example, B. asteroides contributed relatively little to the conversion of substrates but seemed to be responsible for the production of a relatively large fraction of bacterial products. The Firm-4 species showed the opposite pattern, explaining relatively many bacterial substrates but a small fraction of bacterial products. Ion changes identified in CL bees but not in any of the mono-colonizations (in total 20%) may be due to our strict significance cutoffs, additive metabolic activities, or concerted functions of the community members, such as cross-feeding or interspecies metabolic feedback.
Our in vivo results strongly suggest that specific gut bacteria utilize distinct substrates from the pollen diet of bees. This prompted us to test (1) whether the bacterial species could grow in vitro on a pollen-based culture medium and (2) whether this would result in the metabolic conversions of the same compounds as was observed in vivo. To this end, we water-extracted metabolites from the same pollen batch that was used for feeding the bees and analyzed the metabolic composition of this extract using untargeted and targeted metabolomics. Detailed results are presented in S4 Text and show that pollen extract contains physiologically meaningful levels of nutrients and is expectedly enriched in “amino acids and derivatives,” “flavonoids,” “monosaccharides,” and “carboxylic acids and derivatives” (S6 Fig). Strikingly, all community members, except for S. alvi, showed substantial growth in the presence of this pollen extract compared to the nutrient-limited base media in which little or no growth was observed after 16 h of incubation (Fig 5A).
We then profiled the metabolome of growth media before and after bacterial incubation in a separate metabolomics experiment. We annotated a total of 1,031 ions (S9 Data), of which 427 (41%) were also present among the 1,079 ions from the in vivo dataset. In line with their growth profiles, the largest number of depleted metabolites (log2(FC) ≥ |1| and Welch’s t test BH adj. P ≤ 0.01) was found for the growth cultures of Firm-5, followed by G. apicola, Firm-4, F. perrara, B. asteroides, B. apis, and S. alvi (Fig 5A).
Using strict criteria, we identified 17 ions (13 pollen-derived substrates and four bacterial products), which were explained in vivo and in vitro by the same species (S7 Fig and S3 Table). Seven of these 13 substrates belonged to the most discriminatory substrate ions for CL versus MD bees (Fig 3): three flavonoids (quercitrin, afzelin, and rutin), one nucleoside (inosine), and ions annotated as quinate, citrate, and 2-fuorate. The fact that different community members were responsible for the conversion of some of these substrates (B. asteroides, Firm-4, Firm-5, F. perrara, B. apis, and G. apicola) demonstrates that our in vitro cultures allowed us to recapitulate metabolic activities covering the entire community.
We found remarkably overlapping substrate specificity for four flavonoids in vitro and in vivo, with the Firm-5 species being the only member capable of converting rutin and scolymoside, while quercitrin and afzelin were also utilized by Firm-4, and quercitrin was additionally used by B. asteroides and B. apis (Fig 5A). Among the four in vitro recapitulated products were three of the four ions corresponding to putative breakdown products of flavonoids (S7 Fig and S3 Table). These ions accumulated in vivo and in vitro in the presence of Firm-4 and/or Firm-5, providing further evidence for breakdown of the polyphenolic ring structure of flavonoids. However, we also found that deglycosylated flavonoids (i.e., aglycones) significantly accumulated in cultures of Firm-5 and showed a trend towards accumulation for Firm-4 and B. asteroides (Fig 5A). Based on these results, we propose that flavonoid degradation involves two steps (Fig 5B): (1) deglycosylation of sugar residues and their subsequent fermentation and (2) the breakdown of the polyphenol backbone. The second step could be relatively slow, explaining why aglycones accumulated in vitro (16 h), but not in vivo (10 d). Alternatively, the accumulation of some of these aromatic compound degradation intermediates could also result from the metabolism of other substrates such as aromatic amino acids.
An obvious difference in our in vitro experiments compared to the in vivo situation is the absence of the host, which may predigest pollen grains before gut bacteria utilize pollen-derived metabolites. For example, certain sugars and amino acids are expected to be present in lower amounts in vivo because of host absorption. Conversely, the host may also provide physicochemical conditions that support the growth of some community members. This could explain the poor growth of S. alvi in vitro, especially as in vivo S. alvi is tightly associated with the gut epithelium and other gut bacteria such as G. apicola [77].
Microbial species in gut communities can organize into food chains, where one species provides metabolites that can be utilized by others. Such metabolites may be released from insoluble dietary particles via bacterial degradation or can be generated as waste products of metabolism [2]. To identify possible metabolic interactions between community members of the bee gut microbiota, we focused on ions that in vivo significantly accumulated in some mono-colonizations and were depleted in others (S2A Data). A total of 27 ions showed such opposing changes between two or several mono-colonizations (S4 Table).
An example of a potential metabolic interaction identified in our dataset is the liberation and consumption of one of the major bacterial substrates in CL bees, 9-10-18-trihydroxystearate (m/z 331.248), originating from the outer pollen wall. In our mono-colonization experiments, the corresponding ion was depleted in Firm-4 and B. asteroides but accumulated in the case of Firm-5 and G. apicola (Fig 4C). This suggests that the latter two species facilitate the release of this ω-hydroxy acid from the outer pollen wall, possibly rendering it more accessible for further degradation by Firm-4 and B. asteroides.
A second example is pyruvate (m/z 87.008), which substantially accumulated in the gut of bees mono-colonized with G. apicola but was utilized as a substrate by other bacteria such as S. alvi and Firm-5 (Fig 6A). A syntrophic interaction between G. apicola and S. alvi had previously been suggested, because these bacteria are colocalized on the epithelial surface of the ileum [77] and harbor complementary metabolic capacities [32]. To test for potential cross-feeding of pyruvate from G. apicola to S. alvi, we supplemented the pollen-conditioned medium of S. alvi with culture supernatant of G. apicola cultures. The growth of S. alvi was slightly but significantly improved in the conditioned medium compared to the control medium (Fig 6B). Metabolome analysis (S10 Data) revealed six ions that accumulated during the growth of G. apicola and were utilized from the conditioned medium by S. alvi (Fig 6C). Besides pyruvate, these were ions corresponding to three putative fermentation products, a nucleoside derivative, and hydroxyphenylpropionate. We determined the concentration of pyruvate biochemically and showed that G. apicola indeed produces high levels of pyruvate (approximately 4 mM) and that this is subsequently utilized by S. alvi (S8 Fig). These results confirm our predictions from the in vivo dataset and show that bee gut bacteria engage in cross-feeding interactions. While not essential for gut colonization in itself as based on our mono-colonization experiments, such interactions may be important for community assembly and resilience and reflect the longstanding coexistence among these gut bacteria.
The simple composition and experimental amenability of the honey bee gut microbiota facilitated our systems-level approach. We reconstituted the honey bee gut microbiota from cultured strains, characterized the metabolic output of the complete microbiota, identified the contributions of individual community members in vivo, and recapitulated their activities in vitro. Our results provide unprecedented insights into the metabolic functions of bee gut bacteria.
As in the mammalian and termite gut ecosystem [1,68], we conclude that most substrates utilized by the bee gut microbiota are indigestible compounds that originate from the diet of the host and accumulate in the hindgut where bacterial density is the highest (Fig 7). Such compounds include plant metabolites from the outer pollen wall, such as ω-hydroxy acids, phenolamides, and flavonoid glycosides. While one of the bee gut bacteria had previously been identified as utilizing a major pollen polysaccharide (pectin) [5], our data provides, to our knowledge, the first evidence for a role of the gut microbiota in breaking down outer pollen wall components. Bacterial fermentation of these pollen-derived compounds resulted in the accumulation of organic acids (e.g., succinate) and putative polyphenol degradation products, which are likely to impact the physicochemical conditions in the colonized gut. In addition, we found that host-derived signaling molecules are induced by B. asteroides, suggesting a specific interaction of this gut symbiont with the host.
Based on the mono-colonization experiments, we conclude that most metabolic output of the bee gut microbiota can be explained by the metabolic activities of individual community members. This suggests different metabolic niches in the gut, which could be in part explained by the distinct distribution of bacteria along the gut [77]. While we have found evidence for cross-feeding (e.g., between G. apicola and S. alvi), the metabolic exchange between bacteria seems not to be essential for gut colonization, as each community member was able to colonize on its own. This may also be the case for gut bacteria of other animals, as they cannot rely on the presence of specific interaction partners in the highly dynamic gut environment but rather adapt to diet-derived nutrients. However, the bee gut microbiota is relatively simple, and interspecies metabolic exchanges may be essential to establish in more complex communities such as those of the termite or mammalian gut.
The metabolic activities identified in this study are likely underlying the symbiotic functions of the bee gut microbiota and thus may be directly linked to the microbiota’s impact on bee health and physiology [34,41]. The large metabolic overlap between colonized and hive bees demonstrates the relevance of our findings and validates our gnotobiotic bee model. Moreover, our study highlights the versatility of high-throughput untargeted metabolomics to disentangle metabolic functions in microbial ecosystems. We believe that this systematic approach can be extended to other gnotobiotic animals to enable a better understanding of the diversity of metabolic activities and functions that are present in microbial communities.
Newly emerged bees were obtained from a healthy-looking colony of Apis mellifera carnica located at the University of Lausanne. In short, dark-eyed pupae were carefully removed from capped brood cells with sterile tweezers and transferred to sterilized plastic boxes as described previously [36]. Boxes with pupae were kept with a source of sterile sugar water (50% sucrose solution, w/v) at 35°C with 80%–90% humidity for 2 d until the bees emerged, followed by a reduction in temperature to 32°C. For each box, one or two newly emerged bees were dissected, and their homogenized hindguts (in 1 ml 1x PBS) cultured on growth media as described below. To minimize the chance of including contaminated bees in colonization experiments, we excluded cages for which bacterial growth was observed for the tested bees. For the colonization of newly emerged bees, bacterial strains were inoculated from glycerol stocks and restreaked twice. Details on bacterial strains and culture conditions can be found in S1 Table. Bacterial cells were harvested and resuspended in 1x PBS/sugar water (1:1, v/v) at an OD600 of 1. For colonization, bacterial suspensions were added to a source of sterilized pollen and provided to the newly emerged bees (for details, see S5 Text). MD bees were kept under the same conditions, with the same food sources, but without being exposed to bacteria. The mid- and hindgut (S9 Fig) of gnotobiotic bees were dissected at day 10 post colonization and stored at −80°C until further use.
To obtain age-controlled hive bees, several brood frames without adult bees were transferred from the hive to a ventilated Styrofoam box that was kept in an incubator at 32–34°C overnight. The next morning, the newly emerged bees were collected, marked on the thorax with a pen, and reintroduced into the hive. These bees were recollected 10 d later, and their mid- and hindguts were dissected and stored at −80°C until further use.
The colonization experiment was repeated at two different time points of the year (spring and fall, referred to as experiment 1 and experiment 2 in this study). Whenever possible, we included bees from both experiments in our analysis, such as for CL and MD bees. However, this was not possible for all mono-colonizations because of bacterial contaminations (as detected by qPCR) or in a few cases because of the presence of above-threshold viral loads. The precise numbers of bees included per condition are listed in S5 Table.
Bacterial loads were determined by qPCR using universal bacterial and species-specific 16S rRNA primers on DNA samples obtained from the gut tissues used for metabolomics analysis. Details on DNA/RNA extraction methods are given in S5 Text. Each DNA sample was screened with 11 different sets of primers targeting the actin gene of A. mellifera, the universal 16S rRNA region, and the species-specific 16S rRNA region of nine bacterial species, including the seven species used in this study and two non-core species frequently found in the gut of A. mellifera: Alpha-2.1 and Lactobacillus kunkeei. Primers used for this qPCR analysis are listed in S2 Table. We also screened all gut samples for the presence of viruses. Samples that were contaminated with other bacteria than the desired ones (i.e., >105 bacteria cells detected by qPCR) or that had high virus titers were excluded from the analysis where possible (S5 Table, S5 Fig and S5 Text). The minimum information for publication of qPCR experiments (MIQE) guidelines were followed throughout the data analysis of the qPCR experiments [78]. Details on the qPCR analysis can be found in S5 Text.
Bacteria were precultured on solid media from −80°C glycerol stocks before liquid cultures were inoculated for in vitro growth experiments. For G. apicola ELS0169, S. alvi wkB2, F. perrara PEB0191, and B. apis PEB0149, we used a modified M9 minimal medium supplemented with casamino acids and vitamins (http://dx.doi.org/10.17504/protocols.io.kdqcs5w). For B. asteroides ESL0170, the Firm-5 strains, and Firm-4 Hon2N, we used carbohydrate-free MRS (cfMRS) medium [79]. Bacteria were harvested from plates or spun down from overnight liquid cultures (the latter only for Lactobacilli and B. asteroides) and resuspended in the corresponding minimal medium. Freshly prepared liquid cultures were supplemented with either 10% (v/v) ddH2O or pollen extract and inoculated at a final OD600 of 0.05 (see S5 Text for details on pollen extract preparation). Half of the culture was immediately processed to determine colony-forming units (CFUs) and to harvest supernatants for metabolomics at time point 0 h, i.e., before growth. The other half of the culture was incubated for 16 h according to the conditions listed in S1 Table and then processed in the same way as the sample taken at time point 0 h. For CFU counting, serial dilutions were plated on solid media and incubated under the species-specific culturing conditions. For metabolomics analysis, the remaining bacterial culture was spun down at 20,000x g at room temperature for 10 min, and 300 μl of the culture supernatant was transferred to a fresh tube stored at −80°C until further processing. Five replicates were included for each species and treatment group.
For the cross-feeding experiment, G. apicola strain ESL0169 was grown for 8 h in pollen-supplemented M9 medium as described above to an OD600 of 0.11–0.15. Cultures were subsequently sterile filtered and mixed with fresh pollen-supplemented M9 medium 1:1 (v/v) in a total volume of 3 ml in a 12-well plate. For the control condition, non-inoculated pollen-supplemented M9 medium was incubated for 8 h, sterile filtered, and mixed with fresh pollen-supplemented M9 medium 1:1 (v/v). Then, S. alvi wkB2 was added to each well at a final OD600 of 0.05. The growth of S. alvi was assessed by OD600 measurements of a 100-μl aliquot in a 96-well plate with FLUOstar Omega microplate reader (Huberlab, Switzerland). As S. alvi tends to form aggregates, each culture was thoroughly mixed by pipetting up and down before transferring the aliquot and recording the OD600. For metabolomics analysis, supernatants were sampled at time points 0 h and 8 h for the G. apicola cultures and at time points 0, 16, 36, and 72 h for the S. alvi cultures. For biochemical quantification of pyruvate, we used a Pyruvate Assay Procedure kit (K-PYRUV, Megazyme, United States) according to the manufacturer’s microplate assay instructions. Samples of 8, 6, 4, 2, 1, 0.5, 0.25, and 0.125 mM pyruvate were used to generate the standard curve (slope = 0.062, intercept = 0.023, R2 = 0.987). Standards and samples were measured in triplicate at 340 nm with an Infinite M200PRO microplate reader (Tecan, Switzerland).
Metabolites from gut and pollen samples were water-extracted after mechanical disruption, and supernatants from the in vitro experiments were harvested by centrifugation. Gut samples were preselected based on their wet-weight (arithmetic mean 55.1 mg, standard deviation 9.9). Ten times more water than the gut wet weight (v/w) was added, and the samples were homogenized with 0.1 mm zirconia beads (0.1 mm dia. Zirconia/Silica beads; Carl Roth) in a Fast-Prep24 5G homogenizer (MP Biomedicals) at 6 m/s for 45 s. While most of the homogenate was snap-frozen in liquid nitrogen for subsequent DNA/RNA extraction, aliquots of 100 μl were diluted 1:1 with water for metabolite extractions. To do so, the diluted aliquots were incubated in a preheated thermomixer at 80°C and 1,400 rpm for 3 min. After each minute, the samples were vortexed for 10 s. Subsequently, the samples were centrifuged at 20,000x g and 4°C for 5 min, and 150 μl of the resulting supernatant was transferred to a new tube and centrifuged again at 20,000x g for 30 min. Samples for untargeted metabolomics analysis were further diluted 10x in water. All samples were stored at −80°C before metabolomics analysis.
For untargeted analysis, samples were injected into an Agilent 6550 time-of-flight mass spectrometer (ESI-iFunnel Q-TOF, Agilent Technologies) operated in negative mode, at 4 Ghz, high resolution, and with a mass / charge (m/z) range of 50−1,000 [21]. The mobile phase was 60:40 isopropanol:water (v/v) and 1 mM NH4F at pH 9.0 supplemented with hexakis(1H, 1H, 3H- tetrafluoropropoxy)phosphazine and 3-amino-1-propanesulfonic acid for online mass correction. After processing of raw data as described in [21], m/z features (ions) were annotated by matching their accurate mass-to-sum formulas of compounds in the KEGG database with 0.001 Da mass accuracy and accounting for deprotonation [M-H+]-. The complete KEGG database was used because it has broad coverage of plant, bacterial, and insect metabolic pathways. Notably, this metabolomics method cannot distinguish between isobaric compounds, e.g., metabolites having identical m/z values, and in-source fragmentation cannot be accounted for. The raw data of samples from the three sets of experiments (bee gut samples, in vitro supernatants, and cross-feeding supernatants) were processed and annotated separately to accommodate their different sample matrices or times of measurement. These data can be explored in S2 Data, S9 Data and S10 Data. Raw data processing and annotation took place in MATLAB (MATLAB 2015b, The Mathworks, Natick) as described previously [21], and downstream processing and statistical tests were performed in R (version 3.3.2, R Foundation for Statistical Computing, Vienna, Austria).
Selected metabolite samples were measured in targeted fashion using ultra-high-pressure chromatography-coupled tandem mass spectrometry as described before [72]. Metabolite quantifications were performed by interpolating observed intensities to a standard curve of the metabolite using a linear model (R2 ≥ 0.95). Metabolites with standard curves of R2 ≤ 0.95 or in which intensities had to be extrapolated can be interpreted as relative changes only and were labeled in grey in all plots (S6B Fig). We used the weights of the extracted material to express the concentrations in millimole per gram of gut or gram of pollen. The dataset can be found in S6 Data.
Flavonoid ions were targeted for MS/MS fragmentation as [M-H+]- electrospray derivatives with a window size of ± 4 m/z in Q1. Fragmentation of the precursor ion was performed by collision-induced dissociation at 0, 10, 20, and 40 eV collision energy, and fragment-ion spectra were recorded in scanning mode by high-resolution time-of-flight MS. Spectra were interpreted using MetFrag [80], and spectral cosine similarity scores were calculated between reference spectra that were obtained in-house or library spectra from MassBank of North America (MoNA, http://mona.fiehnlab.ucdavis.edu/). For further details, see S5 Text.
All steps of the downstream data analysis were performed in R (R Foundation for Statistical Computing, Vienna, Austria). Samples from double injections (technical replicates) were confirmed to be highly similar and averaged. Subsequent analyses were performed on these averaged ion intensities, which are available in S2D Data, S9 Data and S10 Data.
Principal component analysis (pca function in R) on ion intensities was used for the multivariate inspection of co-clustering of samples from different groups. For reasons of transparency, we carried out four different PCAs on all annotated ions or the subset of ions with robust changes between CL and MD bees and on log2-normalized or Z-score normalized ion intensities. Z-score transformation was used to remove the domination of high-intensity ions.
Ions that were deemed robustly different between the CL and MD bees were those that were significantly different (Welch t test with Benjamini and Hochberg correction ≤ 0.01, t.test and p.adjust(x, method =“BH”) in R) between CL and MD in both independent experiments. Differences between MD and CL samples were expressed as log2(fold change) values for both experiments separately and for pooled data of both experiments (see S2A Data). Fold changes were based on the arithmetic mean of the CL samples divided by the arithmetic mean of the MD samples. The standard error of the log2(fold change) was computed as the square root of the sum of the squared standard errors of the log2-transformed intensities of both CL and MD.
Enrichment analyses were computed on compound class categories from KEGG (in-house database), which are added in the column “compound class” in S2A Data. Some ions with ambiguous annotations had a compound class associated with multiple of these annotations. However, supported by the observation that compound classes between alternative annotations were often the same (or highly related), only the compound class of the first annotation was used as input for one-sided Fischer’s exact tests (fisher.test(x, alternative =“greater”) in R) on a 2 x 2 contingency table for every compound class. Compound classes associated with a single ion were removed from the results because they were deemed not biologically meaningful.
Ions were sorted based on to what extent they are responsible for explaining the separation between the CL and MD groups from experiment 2. To do this, these datasets were selected as the input for an OPLS-DA (opls from the ropls R package). The correlation and covariance between the log10-transformed ion intensities of included samples and the opls “scoreMN” output were computed with the cor and cov functions in R, respectively. The resulting scores were plotted in a so-called S-plot (Fig 3). The substrate and product ions most responsible for the separation were selected based on an absolute correlation ≥0.8 and an absolute covariance of ≥5. Because such analyses can be prone to overfitting, we tested the sensitivity of the most discriminatory ion selection by implementing a “leave one out strategy” and concluded that the selected ions are robust (>80% present in all 1,000 permutations).
One-way ANOVA (aov adjusted with TukeyHSD(x, conf.level = 0.99) in R) was performed between all bee gut samples after selecting the relevant samples from the data matrix and normalizing the intensities to ion standard (Z-) scores (i.e., by row) by subtracting for every ion its arithmetic mean intensity and dividing the resulting values by the standard deviation of its respective ion intensity. The results of the full ANOVA analysis can be explored in S7 Data. For this study, the focus was on differences between any group and MD bees, which were considered significant when having a Tukey HSD post hoc adjusted P value ≤ 0.05. When for a specific mono-colonization group this significance cut-off was met and the direction of the change was the same as that for CL versus MD, the ion was considered to be “explained” by this group.
In order to enrich for pollen ions, we only considered ions with an arithmetic mean intensity of ≥10,000 in the pollen samples, in addition to being highly significantly different from water-matrix control samples (Welch t test with BH correction ≤ 0.001, t.test and p.adjust(x, method = “BH”) in R combined with log2(fold change) difference of ≥ 2).
For the in vitro data (S9 Data), the goal was to identify pollen substrates and bacterial products for which changes in levels were observed in vivo and in vitro. Pollen ions were mapped by matching the top annotation formula of both datasets. For all media-strain combinations, we performed a statistical comparison (Welch t test with BH correction, t.test and p.adjust(x, method =“BH”) in R) between the time points 16 h and 0 h and considered only those ions with a log2(fold change) of ≥|1| and BH adj. P value of ≤0.01 as significant in vitro products or substrates. In order to be certain that only pollen-derived substrates were included, for every strain only ions that displayed a significant negative log2(fold change) exclusively in the base medium supplemented with pollen extract were considered as in vitro pollen substrates.
To identify ions that might be cross-fed between G. apicola and S. alvi, ions were selected that increased during the growth of G. apicola and were depleted when S. alvi was grown in this conditioned medium mixed 1:1 with fresh base medium. To do this, all ion intensities for both strains (S10 Data) were split and transformed to log2(fold change) with respect to the first time point of sampling. Ions that had a log2(fold change) of ≥1 during G. apicola growth and a log2(fold change) of ≤−1 during S. alvi growth were selected. The raw data and R code for recapitulating the metabolomics data analysis can be found in S11 Data.
|
10.1371/journal.ppat.1004110 | Timed Action of IL-27 Protects from Immunopathology while Preserving Defense in Influenza | Infection with influenza virus can result in massive pulmonary infiltration and potentially fatal immunopathology. Understanding the endogenous mechanisms that control immunopathology could provide a key to novel adjunct therapies for this disease. Here we show that the cytokine IL-27 plays a crucial role in protection from exaggerated inflammation during influenza virus infection. Using Il-27ra−/− mice, IL-27 was found to limit immunopathology, neutrophil accumulation, and dampened TH1 or TH17 responses via IL-10–dependent and -independent pathways. Accordingly, the absence of IL-27 signals resulted in a more severe disease course and in diminished survival without impacting viral loads. Consistent with the delayed expression of endogenous Il-27p28 during influenza, systemic treatment with recombinant IL-27 starting at the peak of virus load resulted in a major amelioration of lung pathology, strongly reduced leukocyte infiltration and improved survival without affecting viral clearance. In contrast, early application of IL-27 impaired virus clearance and worsened disease. These findings demonstrate the importance of IL-27 for the physiological control of immunopathology and the potential value of well-timed IL-27 application to treat life-threatening inflammation during lung infection.
| Annual epidemics of influenza result in 3 to 5 million cases of severe illness and approximately 300,000 deaths around the world. Although most patients infected with normal circulating influenza A viruses recover from the illness, complications arise during infections with highly pathogenic strains of the virus, resulting in increased mortality associated with severe immunopathology and acute respiratory distress. Previous studies suggested a major contribution of the vigorous immune response to lung damage. How the immune system constrains the negative impact of inflammation might therefore be of significant importance for future therapies. Our study in a mouse model of influenza shows that the cytokine IL-27 plays a crucial role in survival by protecting against lung damage. Its actions include regulation of innate (neutrophil influx) and adaptive (inflammatory cytokine production of T cells) arms of immunity during the acute respiratory infection. The data also suggest a therapeutic potential of IL-27, as mice treated with recombinant cytokine at later stages of infection exhibited decreased immunopathology and showed improved survival. The findings uncover an important role of IL-27 in limiting the collateral damages of anti-viral immunity and provide initial evidence that these mechanisms might be exploited for the management of severe immunopathology after infection.
| Infection with highly pathogenic strains of influenza viruses, such as the pandemic 1918 Spanish flu, which resulted in 30–50 million deaths, is still a major threat to health [1], [2]. Pathological findings suggest that the vigorous mobilization of innate and adaptive arms of host immunity upon infection leads to uncontrolled inflammation and potentially fatal lung injury [3]. Rapid leukocyte infiltration of the lung and a subsequent cytokine storm involving the excessive production of inflammatory cytokines and chemokines have been strongly implicated in mediating lung immunopathology [3]–[5]. A better understanding of the factors that regulate the balance between viral clearance, tissue damage and resolution of inflammation is therefore necessary [6].
Interleukin 27 (IL-27) might be one important player in this context. The heterodimeric IL-27 belongs to the IL-12 superfamily and is composed of the Epstein-Barr virus inducible gene-3 (EBI3) and the IL-27p28 subunit [7]. The IL-27 receptor complex consists of IL-27Rα (WSX-1, TCCR) and the gp130 subunit, and is expressed by a wide range of cell types including T cells, monocytes, and neutrophils [8]. Initially, IL-27 was thought to promote TH1 responses because of its ability to induce T-bet expression, thereby triggering the upregulation of IL-12β2 receptor and IFN-γ under some conditions [9]–[11]. However, a series of subsequent studies using in vivo models of infection or autoimmune diseases provided evidence that its dominant function is rather to limit immune-mediated pathology [12].
Mice deficient in IL-27 receptor displayed increased immunopathology associated with overwhelming TH1 responses following infection with a number of parasites and intracellular bacteria [13]–[18]. Moreover, the lack of IL-27 receptor signaling resulted in augmented IL-17 production by CD4+ T cells in several animal models, including experimental autoimmune encephalomyelitis (EAE) [15], [19], [20]. Most studies have characterized the ability of IL-27 to suppress CD4+ T cell responses, but accumulating evidence suggests that the regulatory function of IL-27 also extends to cells of the innate immune system [8]. Consistent with its regulatory function, IL-27-mediated activation of STAT1, STAT3, STAT4 or BLIMP-1 promotes IL-10 and suppresses IL-17 production by CD4+ T cells [15], [21]–[23]. Additionally, IL-27 induced the expression of SOCS3 in CD4+ T cells, resulting in reduced IL-2 secretion in these cells [24].
The role of IL-27 in influenza has not been comprehensively studied. Liu et al. reported that influenza virus infection of epithelial cells or leukocytes induced IL-27, which correlates with increased serum levels of IL-27 in influenza patients [25]. Additionally, they showed a STAT1-dependent antiviral action of IL-27 in vitro [25]. Mayer et al. reported that IL-27 induces IFN-γ in transgenic CD8+ T cells [26]. In contrast, Sun et al. found no effect on T cell derived IFN-γ but a reduced IL-10 production by CD8+ T cells and increased leukocyte infiltration in infected Ebi-3−/− or conditional Prdm1−/− mice [23], [27].
We therefore investigated the impact of IL-27 and its receptor IL-27Rα on immunopathology using the highly mouse pathogenic [28] strain A/PR/8 (H1N1) and subsequently explored the therapeutic potential of recombinant IL-27 (rIL-27) to treat inflammatory lung disease in influenza. We found that, Il-27ra−/− mice exhibited increased mortality after influenza virus infection due to exaggerated immunopathology, in conjunction with augmented numbers of IFN-γ or IL-17-producing CD4+ and CD8+ T cells, and a strongly increased neutrophil infiltration. These effects were only partially attributed to diminished IL-27-induced IL-10. Thus, IL-27 plays an important role in limiting destructive inflammation, notably in the resolving phase of infection. Well-timed treatment with rIL-27 improved lung injury and accelerated recovery without affecting viral clearance. Our findings suggest that therapeutic application of rIL-27 predominantly suppresses innate cell recruitment but hardly affects the T cell response in the local tissue. These data demonstrate that IL-27 has a unique role in controlling immunopathology without impacting on host defense, and might therefore represent a promising candidate for immunomodulatory therapy of viral pneumonia.
To determine the role of IL-27 in shaping the immune response against influenza virus, we first examined the kinetics of Il-27p28 and Ebi3 mRNA expression in the lungs of sublethally infected C57BL/6 mice (Fig. 1A). While Ebi3 was constitutively expressed and not significantly upregulated in the lungs and other organs (Fig. 1B), Il-27p28 expression displayed a pronounced peak on day 7 post-infection (d.p.i.), two days after the peak of the viral load (Fig. 1C). Coinciding with the peak of Il-27p28 expression was the maximal expression of Il-10 mRNA, which is consistent with the assumption that IL-27 is an important inducer of IL-10 [15]. These mRNA data were confirmed at the protein level where IL-27 and IL-10 peaked at 7 d.p.i (Fig. 1D). In contrast, the inflammatory cytokines IL-12 and IL-23 were maximal already at 3 d.p.i (Fig. 1D). Thus, the expression kinetic of IL-27 in the infected lungs follows, with some delay, the kinetic of the virus load, being highest when virus is already declining and coming down when immunopathology has resolved. This is compatible with its role for dampening uncontrolled inflammation in a late phase while initially allowing for a rapid start of immune defense.
To assess the impact of IL-27 on survival during influenza, we challenged wild-type (WT) C57BL/6 or IL-27 receptor-deficient (Il-27ra−/−) mice with 3000 egg infectious dose (EID) influenza virus. Il-27ra−/− mice displayed accelerated weight loss and increased mortality following infection (Fig. 1E, F). Accordingly, Il-27ra−/− mice displayed a more severe lung pathology compared to control mice at 7 d.p.i (using a slightly lower virus dose, 2500 EID, to allow survival of all mice) (Fig. 1G). Furthermore, Il-27ra−/− mice had increased capillary leakage in the respiratory tract, leading to increased protein content in the bronchoalveolar lavage (BAL) fluid of these mice (Fig. 1H). A higher neutrophil, but not NK cell infiltration was observed in the lungs of Il-27ra−/− mice at 8 d.p.i (Fig. 1I). Remarkably, the increase in immunopathology and mortality in Il-27ra−/− mice was not due to a compromised viral elimination, as virus load was not significantly different between Il-27ra−/− and control mice (Fig. 1J). These findings demonstrate that IL-27 plays a critical role in limiting immunopathology during the later stages of infection.
Consistent with the ability of IL-27 to suppress TH1 and TH17 responses [15], [29], influenza virus infected Il-27ra−/− mice exhibited significantly increased IFN-γ levels in BAL fluid, lung homogenate (Fig. 2A), and supernatants of enriched lymphocytes from Il-27ra−/− mice after polyclonal stimulation using PMA/ionomycin (Fig. 2B). Accordingly, we detected increased numbers of CD4+ and CD8+ T cells able to produce IFN-γ upon re-stimulation in the BAL and lungs of the Il-27ra−/− mice (Fig. 2C). In contrast, IFNα levels in the BAL of infected Il-27ra−/− mice were not different from WT animals (Fig. S1A). The results suggest that IL-27 dampens IFN-γ-production by T cells during influenza.
Similar to the increased numbers of IFN-γ+ T cells, we observed augmented numbers of IL-17+CD4+, IL-17+CD8+ (Fig. 2D) and TNF-α+CD4+ T cells (Fig. S1B) in the lungs of infected Il-27ra−/− mice. A slight but not significant increase was also found for IL-4+CD4+ T cells (Fig. S1C). Total numbers of CD4+ and CD8+ T cells in the lungs at 9 d.p.i were not different between Il-27ra−/− and WT mice (Fig. 2E). Taken together, these data demonstrate that endogenous IL-27 limits the magnitude of effector cytokine production by T cells during influenza.
Consistent with the role of IL-27 in inducing IL-10 production by CD4+ T cells in other models [15], [30]–[32], infected Il-27ra−/− mice had decreased levels of IL-10 in the BAL fluid (Fig. 3A) and in the supernatant of PMA/ionomycin-restimulated lymphocytes (Fig. 3B). This decrease correlated with the impaired ability of Il-27ra−/− CD4+ T cells from infected mice to produce IL-10 after in vitro restimulation with PMA/ionomycin (Fig. 3C). Moreover, Il-27ra−/− mice had reduced numbers of IL-10+IFN-γ+ double-positive cells, while total IFN-γ+CD4+ T cells were increased, resulting in a significantly reduced IL-10:IFN-γ ratio in CD4+ T cells of Il-27ra−/− mice compared to WT animals (Fig. S2). Although reduced, IL-10-producing CD4+ T cells were not completely lacking in Il-27ra−/− mice, suggesting that other factors besides IL-27 contribute to the induction of IL-10 [33].
Numerous studies have established that IL-27 is signaling via STAT factors such as STAT1, STAT3 and STAT4. Among these, STAT4 has been shown to be involved in the induction of IL-10 production by CD4+ T cells [22], [34]. STAT4 is, however, also the main intermediate of IL-12 signaling. Sublethal influenza virus infection of Stat4−/− mice resulted in significantly fewer lung-infiltrating IL-10+IFN-γ+CD4+ T cells; yet this was not the case in IL-12p40−/− mice. Thus, IL-27 but not IL-12 is responsible for STAT4 mediated induction of IL-10 (Fig. S3A, B). Viral loads in the lungs of infected Stat4−/− mice at 9 d.p.i were not significantly different to WT animals (Fig. S3C).
Thus, IL-10 becomes induced in IFN-γ+CD4+ T cells during influenza by IL-27, in part mediated via STAT4.
We next determined whether IL-10 mediates the anti-inflammatory effects of IL-27. To this end, we infected Il-10−/− mice with influenza virus and analyzed influenza peptide-specific IL-17 or IFN-γ-producing T cells. Indeed, Il-10−/− mice had elevated numbers of IL-17+CD4+ (Fig. 3D) and a slightly increased numbers of IL-17+CD8+ T cells in the lungs (Fig. S4), similar to that of infected Il-27ra−/− mice. These results indicate IL-17 suppression is largely mediated via IL-10.
In contrast, the increased numbers of IFN-γ-producing CD4+ or CD8+ T cells in the respiratory tract of infected Il-27ra−/− mice were not observed in Il-10−/− mice (Fig. 3D and Fig. S4). These findings were verified by blocking IL-10 signaling in vivo by administration of an anti-IL-10 receptor antibody (αIL-10R; Fig. S5A, B). Viral titers in the lungs of Il-10−/− mice were not different to that of WT mice (Fig. S5C).
The ability of IL-27 to directly modulate IFN-γ production in CD8+ T cells was confirmed in vitro, where addition of rIL-27 to the cultures strongly suppressed IFN-γ and IL-2 production by activated IFN-γ+CD8+ T cells (Tc1 cells), even when IL-10R signaling was blocked (Fig. 3E, F and Fig. S6). An IL-27-dependent suppression of IFN-γ and IL-2 in CD4+ T cells had already been described previously [21].
These results demonstrate that IL-10-dependent effects only partially account for IL-27 mediated suppression. Notably the IL-10-independent effects on IFNγ-production, but also on distinct recruitment events (see below) might explain why deficiency in IL-27 signaling has a strong impact on the disease course in influenza while IL-10 deficiency has not (Fig. S7; for the latter see also [35]).
Having demonstrated the pronounced role of IL-27 in regulating immunopathology, we wondered whether this property could be exploited for therapeutic purposes. In accordance with the delayed endogenous production of IL-27, we administered exogenous rIL-27 from 5–10 d.p.i. Indeed, this treatment regimen resulted in decreased weight loss and accelerated recovery (Fig. 4A), a striking improvement in lung immunopathology (Fig. 4B), and in reduced capillary leakage as indicated by a lower BAL fluid protein content (Fig. 4C). Notably, rIL-27 therapy did not impair viral clearance (Fig. 4D). In line with this, the numbers of CD8+ (Fig. 4E) or CD4+ (Fig. S8) T cells from the infected respiratory tract producing either TNF, IL-17 or IFN-γ upon antigen-specific stimulation were not changed. Only a slight decrease of secreted IL-17 (Fig. 4F) and increase of IL-10 levels (Fig. 4G) was found in the BAL fluid. The protective effect of treatment with rIL-27 was also found when mice were infected with a lethal dose of influenza virus (Fig. 4H). We did not observe significant effects of IL-27 treatment on the activity of virus-specific CTLs as measured by the CD107 mobilization assay or the fraction of IFN-γ-producing CD8+ cells (Fig. S8).
To test whether the delayed kinetics of endogenous IL-27 is relevant for an unhindered initial response to infection, we applied exogenous rIL-27 from 1–7 d.p.i. (early phase). Mice treated under this regimen exhibited stronger weight loss (Fig. 5A) and reached the limits for euthanasia at 7 d.p.i. To assess the impact of treatment on immunopathology and other parameters, all animals were sacrificed at this time point. Although a diminished lung histopathology was observed (Fig. 5B), a significantly higher viral load was found (Fig. 5C). Impaired viral clearance was not due to a suppressed T cell cytokine response as treated mice had unchanged numbers of influenza peptide specific IL-17+ or IFN-γ+ T cells in the respiratory tract (Fig. 5D). However, mice treated in the early phase with rIL-27 had significantly reduced frequencies of neutrophils (Fig. 5E) and monocytes, but not NK cells (Fig. 5F). In our model, NK cells played a minimal role in viral clearance, as NK1.1 cell-depletion did not influence viral loads (Fig. S9).
These findings suggest that systemic rIL-27 treatment during the early stages of influenza has little impact on the local antigen-specific T cell response, but suppresses neutrophil and monocyte influx that are crucial for the control of infection at this stage [28], [36]. In contrast, treatment at a later time-point, starting at the peak of viral load, did not impair viral clearance but immunopathology and disease course were markedly improved.
Therapeutic application of rIL-27 from 5–9 d.p.i. had surprisingly little impact on the T cell compartment, but strongly suppressed the accumulation of neutrophils (Fig. 6A), monocytes and partially NK cells in the lung (Fig. 6B). We therefore conclude that IL-27 can regulate innate cell trafficking independently of any effect on T cell responses. Reduction of neutrophils, but not of NK cells was mediated via IL-10, as is the reduction of some chemokines in the BAL (Fig. S12)
Leukocyte accumulation in sites of inflammation is regulated by chemokines and adhesion molecules governing both the entry and exit from tissue. Indeed, rIL-27-treatment during influenza reduced the levels of multiple chemokines in the BAL fluid (Fig. 6C). To identify the cellular targets of IL-27-dependent chemokine suppression, different leukocyte subsets from the lungs of influenza virus infected mice were isolated and cultured overnight in the absence or presence of rIL-27. IL-27 suppressed chemokine production by CD11b+ or CD11c+ cells but had little impact on NK cells or neutrophils (Fig. 6D). Especially the chemokines KC (CXCL1), MIP-1β (CCL4) and RANTES (CCL5), which are prototypic attractors of neutrophils, monocytic, and lymphocytic cells, were suppressed. A similar suppression by IL-27 was found for IL-1β or IL-6-induced chemokine production by endothelial cells isolated from the lungs (Fig. S10).
To confirm the ability of IL-27 to suppress leukocyte recruitment in the absence of a significant contribution from T cells, we determined the impact of IL-27 treatment in the T cell-independent zymosan-induced peritonitis model. Similar to influenza, rIL-27 significantly reduced the levels of chemokines in the peritoneal lavage as well as the numbers of neutrophils in peritoneum and blood (Fig. 6E, F). As only minimal levels of IL-17 are present in the peritoneal lavage (Fig. S11), this effect does not rely on IL-17 suppression by IL-27. These data unravel a novel mode of action of IL-27 that is based on suppression of innate cell recruitment into sites of inflammation.
The tight regulation of both the induction and subsequent down-regulation of inflammatory responses during influenza is imperative in minimizing severe immunopathology. Infection with highly pathogenic strains of influenza viruses results in increased leukocytic pulmonary infiltrates and leads to the exaggerated production of inflammatory cytokines (“cytokine storm”) that causes massive inflammation with increased mortality [4], [5], [37]. Therefore, understanding the regulatory pathways during infection not only sheds light on the mechanisms controlling the delicate balance of efficient viral clearance and disastrous immunopathology, but also reveals potential therapeutic approaches to target resolution of inflammation [6]. Few studies have evaluated the therapeutic potential of anti-inflammatory agents in influenza; while broad-acting immunosuppressants such as corticosteroids were found to worsen the disease, a combination of antiviral therapy and anti-inflammatory non-steroidals inhibiting cyclooxygenases (COX) improved survival in mice [38]. Similarly, targeting inhibitory pathways such as macrophage CD200, PAR2 and endothelial S1P1 receptors have been found to reduce immunopathology in influenza infection models [39]–[41]
Our findings suggest that IL-27 is a potential candidate for the treatment of immunopathology, as endogenous IL-27 was found to play a major role in dampening of exaggerated inflammation in influenza while having little impact on virus elimination. The absence of IL-27Rα signaling during acute virus infection worsened immunopathology and disease course; this ultimately resulted in increased mortality, despite controlled viral loads. Additionally, increased neutrophil accumulation and augmented IFN-γ or IL-17 production by T cells were observed in the infected Il-27ra−/− mice while local IFNα levels appeared not to be affected. These data are in agreement with a number of in vivo models of bacterial or parasitic infection that underline a crucial role of IL-27 in dampening inflammation [14], [16]–[18], [42].
That IL-27 acts in vivo predominantly as an anti-inflammatory cytokine was not foreseen in the beginning, as several studies demonstrated activating effects of IL-27, e.g. on the production of IFN-γ in vitro [9], [10], [43]. In an influenza model, Mayer et al. reported that WT CD8+ T cells displayed higher IFN-γ production than IL-27Rα-deficient cells [26]. In this chimera model, non-hematopoietic and half of the hematopoietic cells responded to IL-27 so that only T cell- intrinsic effects of deficiency were effective. In contrast, under the conditions of global absence of IL-27Rα as used here, we observed increased IFN-γ levels and two-fold higher numbers of IFN-γ+ T cells in the infected respiratory tract of Il-27ra−/− mice, in line with the findings of the above-mentioned parasite infection models. We assume, that the global effect of IL-27 in vivo involves a complex network of cell types including myeloid cells or even non-hematopoietic cells. In addition, timing and conditions might be crucial for the quality of IL-27 effects, as we also found a direct, IL-10-independent suppression of IFN-γ and IL-2 in activated Tc1 cells by IL-27 in vitro. Thus, the environmental context plays a significant role for the action of IL-27 in vivo, and its impact on the innate response might dominate over effects restricted to the T cell compartment. Indeed, the strong increase in the number of lung-infiltrating neutrophils in the absence of IL-27 signaling was one of the most impressive findings and appears to be crucial for the worsened immunopathology.
Major effects of compromised IL-27 signaling were also found on the number of IL-17 producing T cells. Both IFN-γ and IL-17 have been reported to play a significant role for lung injury during influenza [4], [44], [45]. IL-17 has been described as a major factor boosting expansion, recruitment and activation of neutrophils by inducing hematopoietic growth factors, chemokines and other activating signals [46]–[48]. IL-27-dependent regulation of TH17 responses was reported to occur through a number of mechanisms [15], [20], [49]–[51]. Here we provide evidence that suppression of IL-17 is largely dependent on IL-10 acting as an intermediate, since infected Il-10−/− mice displayed augmented numbers of IL-17+ T cells, similar to that observed in Il-27ra−/− mice. This is in agreement with a previous study in which blocked IL-10R signaling during high dose influenza virus infection resulted in elevated numbers of IL-17+CD4+ T cells [35]. In contrast, IFN-γ-producing T cells were not affected by absence of IL-10.
Based on these data demonstrating the important role of IL-27 in controlling inflammation, we reasoned that application of rIL-27 might be of value in situations in which exaggerated immunopathology, rather than virus elimination, becomes a critical issue for host survival as it is often the case in severe influenza. Indeed, systemic application of daily doses of rIL-27 at 5–9 d.p.i accelerated recovery and alleviated immunopathology by suppressing the influx of neutrophils, monocytes and, to a lesser degree, NK cells into the infected lungs of mice. Reduced infiltration appears to be the major cause of the improved overall status of treated mice, as large numbers of these cells can contribute to lethal lung damage by producing inflammatory cytokines, chemokines and reactive oxygen species, which results in the amplification of inflammatory signals [4], [5], [37]. Again, the reduction in infiltrating neutrophils upon IL-27 therapy was largely dependent on IL-10 (Fig. S12). In contrast, the reduced infiltration of NK cells was not dependent on IL-10, underlining that not all effects of IL-27 are mediated by induced IL-10 and that IL-27 has a broader suppressive effect than its downstream-mediator IL-10. This latter conclusion is supported by the finding that infected Il-27ra−/−, but not Il-10−/− mice exhibited a more severe disease course compared to WT animals.
Surprisingly, the number and cytokine profile of influenza virus-specific T cells in the lung was not significantly affected by treatment with rIL-27. Moreover, virus elimination was not impaired, if not even improved, upon treatment in the late phase. Whether this is due to the reported induction of antiviral activity by IL-27 that activates an interferon-induced antiviral protein kinase-R (PKR) via STAT1 in human lung epithelial cells [25], or whether destructive inflammation counteracts an efficient antiviral defense, remains to be shown.
Although some reduction in the level of IL-17 and increase in IL-10 was found in the BAL fluid upon treatment, these findings suggest that rIL-27 applied systemically predominantly regulates innate cell accumulation in the lungs rather than limiting the activity of the adaptive arm of the immune system such as IFN-γ or IL-17 producing T cells within the inflamed tissue. An explanation could be that local levels of IL-27 calculated for the lung tissue of infected WT animals are two orders of magnitude higher than plasma levels after systemic application of rIL-27 and are therefore hardly increased upon treatment (Fig. S13). We therefore propose that systemically applied rIL-27 predominantly acts on cells exposed directly to blood or plasma exudate and/or on innate cells before or during their journey to the inflamed lung.
To test the hypothesis that IL-27 treatment is able to suppress the accumulation of neutrophils independent of T cells, we applied rIL-27 in an acute model of TLR-induced sterile inflammation, the zymosan-induced peritonitis model. In this model, T cells are virtually absent in the inflammatory site, and IL-17 is hardly detectable. Indeed, rIL-27 inhibited the accumulation of neutrophils also under these conditions.
As leukocyte trafficking is controlled by adhesion molecules and chemokines presented on endothelial cells, we tested whether IL-27 affects key molecules involved in the recruitment or retention of innate leukocytes in influenza. Consistent with a role for IL-27 in modulating the trafficking of neutrophils and monocytes, treatment with rIL-27 reduced the production of neutrophil and monocyte chemoattractants KC (CXCL1), MIP-1α (CCL3), and RANTES (CCL5) produced in vitro by pulmonary monocytes (CD11b+), alveolar macrophages/dendritic cells (CD11c+) or NK cells isolated from infected lungs and resulted in reduced chemokine levels in the BAL fluid of infected mice. Similar effects were found with lung endothelial cells. These data complement recent findings that IL-27 suppresses the response of macrophages to TNF-α and IL-1 [52].
While the altered levels of chemokines in the BAL might affect the retention of leukocytes in the alveolar space, the deposition of chemokines on the endothelial surface by macrophages lining the blood vessels would directly affect the adhesion and transmigration of circulating leukocytes. Indeed, a major fraction of monocytic cells in the lung is not situated in the parenchyma but sitting within the vessel wall (“marginal pool”), rendering these cells sensitive to the cytokines in the blood, including exogenously administered cytokines [53], [54]. In addition we found that the chemokine production of endothelial cells upon stimulation with IL-1β or IL-6 was suppressed by IL-27. Moreover, IL-27 has been reported to directly affect adhesion and activation of neutrophils [55].
The expression of the IL-27p28 subunit in the influenza virus-infected respiratory tract peaks at the later phase of infection when viral titers are at a decline, which is consistent with the suggested role of IL-27 in limiting the immune response. Interferons can elicit IL-27 production as the Il-27p28 gene promoter contains an IFN-stimulated response element region (ISRE), which becomes activated through IRF-1 [56]–[58]. In contrast to the inflammatory cytokines IL-12 or IL-23, which are rapidly produced by myeloid cells, e.g. upon triggering TLR receptors, and accordingly found in early time points in the influenza infection, the expression of IL-27 is turned on in a delayed fashion by the inflammatory microenvironment and serves as a negative feedback mechanism, thereby dampening the immune response in the later phase when adaptive immunity is established and the risk of severe immunopathology comes to the fore.
In line with this concept, we observed protective effects when rIL-27 was administered in a later phase of infection, starting at the peak of viral load when also the endogenous IL-27 production is near its highest level. To test whether timing is crucial, we additionally applied rIL-27 in the early phase of infection, starting 1 day after infection. Indeed, under these conditions IL-27 treatment also reduced leukocyte infiltration and immunopathology, but simultaneously impaired virus elimination, resulting in a worsened disease course. This suggests that interference with leukocyte recruitment in the early phase of influenza aggravates the infection, and the low level of endogenously produced IL-27 in this early phase is appropriate to allow their unhindered rapid activity in virus defense.
Indeed, previous studies have demonstrated that neutrophils are essential for early host protection in influenza infection, as neutrophil depletion before infection led to increased viral titers and accelerated mortality [28], [36]. The same was found for alveolar macrophages [36]. While their mode of action is still unknown, these studies suggest that neutrophils and macrophages contribute to protection in the early phase. In the late phase, depletion of neutrophils or macrophages was not affecting the disease course, while even infection with a low pathogenic strain was fatal in RAG-2γc−/− mice lacking NK, T and B cells [28]. On the other side, recruited leukocytes have a major role in immunopathology, e.g. by inducing apoptosis in epithelial cells [59]. These findings are compatible with the paradigm that the innate system contributes to early protection in viral infection, while at later time points, antigen-specific T cells take over and eliminate the virus.
In conclusion, our study shows that endogenous IL-27 has a crucial role in preventing a fatal disease course in influenza where it acts to limit and resolve the inflammatory process while allowing an unimpaired antiviral response (Fig. S14). Based on its physiological role as a master factor regulating IL-10-dependent as well as -independent anti-inflammatory mechanisms, we here demonstrate that well-timed therapeutic application of recombinant IL-27 can successfully counteract detrimental immunopathology while keeping the antiviral response intact. Combination of IL-27 treatment with anti-viral or anti-microbial treatment might further expand the applicability of this concept, especially when the role of IL-27 in secondary bacterial infection [60] is appropriately taken into account.
These data suggest that strategies to target natural multifunctional pathways involved in the resolution of inflammation might be a valuable alternative for the treatment of inflammation-caused immunopathology and complement current therapeutic approaches focused on the inhibition of isolated effector mechanisms.
Wild-type (WT) C57BL/6, BALB/c, Il-10−/−, Il-12p40−/− and Stat4−/− mice were purchased from Taconic Farms, Charles River, or The Jackson Laboratory. Il-27ra−/− mice were backcrossed more than nine generations to C57BL/6 mice [11] and housed in the Department of Pathobiology at The University of Pennsylvania, Philadelphia, USA. Mice were infected with influenza A/PR/8/34 virus intranasally (i.n) while under dexdomitor (0.5 mg/mL, Pfizer) and anti-mepazole (5 mg/mL, Pfizer) anesthesia, or with light isofluorane. For sublethal infection with influenza virus, BALB/c and Stat4−/− mice were infected with 12.5 plaque forming units (PFU), while C57BL/6, Il-10−/− and Il-27ra−/− mice were infected with 25 PFU or 2500 egg infectious dose (EID). Survival assays with Il-27ra−/− mice was performed with 3000 EID. For experiments addressing the therapeutic potential of rIL-27, 45 PFU of the virus were used as sublethal and 55 PFU as lethal dose. Infected mice were weighted daily and assessed for clinical symptoms of infection. When reaching a weight loss of >30%, mice were euthanatized. Animal care and experiments were performed in accordance with the institutional guidelines of German Federal Law and local authorities of Berlin (LAGESO), or Animal Care and Use Committee of the University of Pennsylvania.
Animal procedures were performed in accordance with the German “Tierschutzgesetz in der Fassung vom 18. Mai 2006 (BGBI.IS.1207)” and the guideline 2010/63/EU from the European Union and the European Convention for the protection of vertebrate animals used for experimental and other scientific purposes. Animal protocols were approved by the ethics committee and the Berlin state authorities (LAGeSo Registration # G03310/08). Experiments performed at the University of Pennsylvania were carried out in accordance with the guidelines in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The animal protocol (# 802004) was approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Pennsylvania, Philadelphia PA (Federal assurance # FWA00004028; Office of Laboratory Animal Welfare assurance # A3079-01).
Influenza A/PR/8/34 virus was grown in the allantoic cavaties of 11-day old embryonated chicken eggs or was purchased from Charles River.
Bronchoalveolar lavage (BAL) was obtained by flushing the airways 3× with 1 mL sterile PBS. Lungs (approximately 200 mg of tissues) were mashed through a 70 µm cell strainer and suspended in 10 mL of PBS/BSA. Lung lymphocytes were enriched using a percoll gradient 40∶70 (vol/vol). Erythrocytes were lysed with an erylysis buffer (Sigma). BAL fluid and lung homogenates were collected and stored at −80°C for ELISA and BCA.
CD11b+, CD11c+, NK1.1+CD3− (NK cells) and Ly6G+CD11b+ (neutrophils) cells were isolated from infected lungs of C57BL/6 mice at 5 d.p.i. and sorted by FACS. Cells (1×105/50 µL cRPMI) were incubated in the presence or absence of rIL-27 (50 ng/mL) for 24 hours (h) without further stimulation and chemokine concentrations in supernatants analyzed.
Total BAL fluid protein was measured using a bicinchoninic protein assay (BCA) kit (Pierce Chemical). Cytokine or chemokine concentration in BAL fluid, lung homogenates or cell cultures were measured by enzyme-linked immunosorbent assay (BD Biosciences or eBiosciences) or FlowCytomix multiple detection kit (eBioscience).
C57BL/6 mice were injected with zymosan (1 mg in 1 mL PBS) intraperitoneally (i.p) with or without co-injection of rIL-27 (200 ng, R&D Systems). After 24 h, blood or 5 mL peritoneal lavage (PBS) was obtained.
Naive CD8+ T cells (CD8+CD62L+) were isolated from pooled splenocyte and lymph nodes using magnetic bead separation (Miltenyi). For activation, purified CD8+ T cells (1×106 cells per mL) were cultured with plate-bound anti-CD3 (3 µg/mL; clone 145-2C11) and anti-CD28 (3 µg/mL; clone 37.51). CD8+ T cells were supplemented with murine rIL-12 (5 ng/mL; eBioscience), rIFN-γ (20 ng/mL; eBioscience), and rIL-2 (10 ng/mL; R&D Systems) plus anti-mouse IL-4 antibody (5 µg/mL; clone 11B11) in complete RPMI media (RPMI 1640 (Gibco®) plus 10% FCS (vol/vol) and antibiotics) for 3 days (d). After 3 d in culture, Tc1 cells were transferred to a plate without anti-CD3 and anti-CD28 and supplemented with fresh medium plus rIL-2 and cultured for 2 d for a total of 5 d in culture. In cases where rIL-27 (rIL-27) was added or IL-10 receptor was blocked, the media were supplemented with rIL-27 (50 ng/mL; eBioscience) and/or anti-IL-10 receptor blocking antibody (αIL-10R) (40 µg/mL; 1B1-2). In some experiments, cells were labeled with CFSE (Sigma) and cultured under the same conditions as mentioned above.
Intracellular cytokine staining (ICS) was performed as previously described (Hamada et al., 2009). Briefly, enriched lymphocyte samples were restimulated either with phorbol 12-myristate 13-acetate (PMA; 100 ng/mL) plus ionomycin for 4 h, or a combination of immunodominant influenza virus peptides (Anaspec; Table S1) for 6 h in the presence of Brefeldin A (10 µg/mL, Sigma). For analysis of CD107 expression, anti-CD107a antibody (10 µg/mL) and anti-FcRγ (10 µg/mL) was added to the culture during 6 h peptide restimulation. Cells were incubated with anti-FcRγ receptor antibody prior to staining for surface markers. Surface-labeled cells were fixed for 20 minutes using 2% paraformaldehyde. Cytokines were stained using fluorochrome-labeled anti-mouse monoclonal antibodies (mAbs) in 0.1% saponin buffer for 20 minutes. Cells were washed and resuspended in PBS/BSA then analyzed on a FACS Canto II (BD Biosciences). Data were analyzed using Flowjo analysis software (TreeStar). For FACS antibodies used, see supplemental methods.
RNA was extracted using RNeasy Mini Kit with oncolumn DNase digestion (Qiagen). cDNA was synthesized using SuperscriptII Reverse Transcriptase (Invitrogen) with random hexamers and oligo(dT) primers (Qiagen). Quantitative reverse-transcription PCR (qRT-PCR) was performed using Platinum SYBR Green qPCR SuperMix-UDG (Invitrogen) on a Stratagene MX3000 thermo cycler. For qRT-PCR primers used, see Table S2.
Blockade of IL-10 signaling in vivo was achieved by administration of an anti-IL-10 receptor-specific mAb (αIL-10R; clone 1B1-2) after 3 d (i.p., 1 mg in 200 µL PBS), 4 d (i.n., 0.15 mg in 30 µL PBS) and 6 d.p.i (i.p., 1 mg in 200 µL PBS) [27]. rIL-27 (200 ng in 100 µL PBS; eBioscience) was injected i.p. from 1–7 d.p.i. (early) or 5–9/10 d.p.i. (late treatment). In parallel, mice were injected solely with PBS (NT) as control. NK cell depletion was performed by i.p injection of anti-NK1.1 depleting antibody (500 µg in 500 µL; clone PK136) at −1, 1 and 5 d.p.i.
Lung samples were fixed with 4% formaldehyde, embedded in paraffin and stained with hematoxylin and eosin (H&E). Images were acquired using an AxioImager Z1 microscope equipped with a charge-coupled device (CCD) camera (AxioCam MRm) and processed with AxioVision software (all purchased from Carl Zeiss MicroImaging, Inc.). H&E stained lung sections were scored in a blinded manner as follows: (0) normal, (1) minor perivascular inflammation around large blood vessels, (2) moderate perivascular and peribronchial inflammation, (3) increased perivascular and peribronchial inflammation, (4) severe formation of perivascular, peribronchial, and interstitial inflammation.
The following murine monoclonal antibodies (mAbs) were purchased from BD Biosciences or eBioscience: CD4 (RM4-5), Ly6-G (RB6-8C5), CD11c (N418), CD49b (pan-NK), CD11b (M1-70), CD8a+ (53-6.7), CD62L (16A/MEL-14), CD3 (145-2C11), IFN-γ (AN18.17.24), IL-10 (JES5-A6E3), IL-17A (TC11-18H10), TNF-α (MP6-XT22), CD107 (1D4B). CD31 (MEC13.3), ICAM-1 (KAT1) and MHCII (M5/114) were obtained in house. In certain cases, PI was added at 40 µg/ml to the cells immediately prior to cell acquisition.
Data are means ± s.d. or s.e.m. Statistical tests used include Kaplan-Meier log-rank survival test, and unpaired two-tailed Students t test. All P values>0.05 are considered not to be significant.
|
10.1371/journal.pntd.0004911 | Alternatively Activated Mononuclear Phagocytes from the Skin Site of Infection and the Impact of IL-4Rα Signalling on CD4+T Cell Survival in Draining Lymph Nodes after Repeated Exposure to Schistosoma mansoni Cercariae | In a murine model of repeated exposure of the skin to infective Schistosoma mansoni cercariae, events leading to the priming of CD4 cells in the skin draining lymph nodes were examined. The dermal exudate cell (DEC) population recovered from repeatedly (4x) exposed skin contained an influx of mononuclear phagocytes comprising three distinct populations according to their differential expression of F4/80 and MHC-II. As determined by gene expression analysis, all three DEC populations (F4/80-MHC-IIhigh, F4/80+MHC-IIhigh, F4/80+MHC-IIint) exhibited major up-regulation of genes associated with alternative activation. The gene encoding RELMα (hallmark of alternatively activated cells) was highly up-regulated in all three DEC populations. However, in 4x infected mice deficient in RELMα, there was no change in the extent of inflammation at the skin infection site compared to 4x infected wild-type cohorts, nor was there a difference in the abundance of different mononuclear phagocyte DEC populations. The absence of RELMα resulted in greater numbers of CD4+ cells in the skin draining lymph nodes (sdLN) of 4x infected mice, although they remained hypo-responsive. Using mice deficient for IL-4Rα, in which alternative activation is compromised, we show that after repeated schistosome infection, levels of regulatory IL-10 in the skin were reduced, accompanied by increased numbers of MHC-IIhigh cells and CD4+ T cells in the skin. There were also increased numbers of CD4+ T cells in the sdLN in the absence of IL-4Rα compared to cells from singly infected mice. Although their ability to proliferate was still compromised, increased cellularity of sdLN from 4x IL-4RαKO mice correlated with reduced expression of Fas/FasL, resulting in decreased apoptosis and cell death but increased numbers of viable CD4+ T cells. This study highlights a mechanism through which IL-4Rα may regulate the immune system through the induction of IL-10 and regulation of Fas/FasL mediated cell death.
| In areas endemic for schistosomiasis, repeated exposure to infective cercariae is a frequent occurrence, and repeated exposure of murine skin to Schistosoma mansoni resulted in CD4+ T cells becoming hypo-responsive. Here potential contributory mechanisms were investigated. In the skin infection site, three mononuclear phagocyte populations were identified (tissue macrophages, dendritic cells, and macrophages) which exhibited up-regulation of genes associated with alternative activation, in particular the gene encoding RELMα. However, in repeatedly infected mice deficient in RELMα, there was no change in the abundance of mononuclear phagocytes in the skin, and CD4+ cells in the skin draining lymph nodes remained hypo-responsive. In mice deficient for IL-4Rα, required for alternative activation, levels of dermal regulatory IL-10 were reduced and there was an increase in the abundance of antigen presenting MHC-IIhigh cells, which was accompanied by increased numbers of CD4+ T cells. Although the absence of IL-4Rα did not translate into increased CD4+ cell responsiveness, they exhibited lower expression of Fas/FasL, resulting in decreased apoptosis/cell death and increased cell viability. This study highlights a mechanism through which IL-4Rα may regulate the immune system through the induction of IL-10 and regulation of Fas/FasL mediated cell death.
| Schistosomiasis is a debilitating disease that develops following percutaneous infection with the parasitic helminth Schistosoma spp [1, 2]. The disease affects approximately 230 million people worldwide with infection occurring when the skin is exposed to the free-swimming tissue invasive cercariae [3]. Since the infective stage of the parasite is often present in water used for domestic purposes, individuals living in areas endemic to schistosomiasis are at risk of repeated infections.
In order to investigate the effect of repeated infection with schistosome cercariae on the immune response, we developed an experimental model whereby mice were exposed via their pinnae once (1x), or repeatedly (4x), to doses of infective S. mansoni cercariae [4]. It was found that after 4x, compared to 1x, exposures there were major changes in the cell populations within the skin site of infection such that eosinophils, macrophages, dendritic cells (DCs), neutrophils, mast cells, CD4+ T cells and keratinocytes were all increased after 4x infections [4–7]. Moreover, repeated infections resulted in the development of CD4+ T cell hypo-responsiveness in the skin-draining lymph nodes (sdLN), as well as decreased immunopathology in the liver generated in response to eggs released by adult worms [4]. CD4+ T cells in the sdLN from 4x mice had reduced ability to proliferate and secrete cytokines in response to larval schistosome antigens, and this was shown to be IL-10 dependent [6, 8].
Exposure of the skin to repeated doses of schistosome cercariae caused major changes in the local cytokine environment, particularly the levels of IL-4, IL-13 and IL-10, and it was proposed that immune responses in the skin involved mononuclear phagocytes that were alternatively activated [4]. The term alternative activation conventionally describes macrophages under the influence of IL-4 and IL-13 [9–12]. Alternatively activated macrophages have been given the term M2, or subgroupings thereof M2a-c [13], whilst the term alternative activation has also been used in the context of DCs [14]. Parasitic helminth infections are often associated with the development of alternatively activated cells and polarized Th2-type immune responses [15]. Together, these cell populations are linked to the development of wound healing, which accompanies the response to tissue invasive helminths [16, 17]. Therefore, it is likely that there will be substantial wound healing following the repeated percutaneous invasion of the skin by schistosome cercariae.
In the current study, we set out to investigate mononuclear phagocyte cell populations in the skin infection site following repeated exposure to schistosome larvae, and found evidence for a gene expression profile associated with alternative activation. Subsequently, we focused our investigation upon the impact of IL-4Rα which is a receptor for both IL-4 and IL-13 essential for the alternative activation of macrophages [9, 10], and Resistin-like molecule α (RELMα) which is a key marker of alternative activation and is abundant during Th2 immune responses in allergic lung inflammation and helminth infection [18–21]. Consequently, we hypothesized that IL-4Rα and RELMα might play an important role in the response following repeated exposures to infectious S. mansoni cercariae.
Experiments were carried out in accordance with the United Kingdom Animals Scientific Procedures Act 1986 and with the approval of the University of York Ethics Committee (PPL 60/4340).
C57BL/6 wild-type (WT), IL-4Rα deficient (Il-4rα-/-; IL-4RαKO) [22], and Resistin-like molecule α deficient (Retlnα-/-; RelmαKO) [23], mice were bred and housed at the University of York. Both IL-4RαKO and RelmαKO mice on C57BL/6 background were kind gifts from Professor Judith Allen, University of Edinburgh. Age and sex-matched mice (between 6 and 10 weeks) were used for all experiments.
The life cycle of a Puerto Rican strain of S. mansoni was maintained at the University of York. Both pinnae of mice were percutaneously exposed to 150 S. mansoni cercariae either once only (1x), or four times in total (4x) on a once-a-week basis between day 0 to day 21, as described previously [4, 24]. Pinnae and skin-draining lymph nodes (sdLN; auricular lymph nodes) were harvested 4 days after the final infection. Skin infection via the pinnae results in a 50% penetration rate [24], therefore a dose of approximately 75 cercariae per pinna is achieved. Pinnae inflammation was measured using a dial gauge micrometer (Mitutoyo, Japan).
Pinnae were collected and briefly exposed to 70% ethanol to sterilize before being air-dried. They were then split along the central cartilage into two halves, and floated on the surface of complete RPMI media (RPMI-1640 (Gibco, Paisley, UK) containing 10% heat inactivated FCS (Biosera, Uckfield, UK), 2 mM L-Glutamine, 1% Pen/Strep (both Gibco) and 50 μM 2-mercaptoethanol (Sigma-Aldrich, Gillingham, UK) overnight at 37°C 5% CO2, to obtain the dermal exudate cells (DEC) as previously described [4, 24, 25]. After overnight in vitro culture, the tissue samples were discarded, while the culture supernatants were spun at 1000 xg for 7 minutes at 4°C to recover the DEC. The skin biopsy culture supernatants were frozen at -20°C prior to subsequent analysis by ELISA. DEC were re-suspended in complete RPMI, numbers determined by trypan blue exclusion, and then subjected to either analysis by flow cytometry, or separated into groups by fluorescence-activated cell sorting (FACS).
The amounts of released IL-4, IL-10, IL-12p40 present in the skin biopsy culture supernatants were determined using DuoSet enzyme-linked immunosorbent assay (ELISA) kits (R&D Systems, Abingdon, UK).
Mice received 1 mg bromodeoxyuridine (BrdU; Sigma-Aldrich) via daily intraperitoneal (i.p.) injection for the final 4 days prior to harvest of the sdLN in order to determine in vivo CD4+ cell proliferation [8]. Cells recovered from the sdLN were initially blocked with anti-CD16/32 monoclonal antibodies (mAbs; eBioscience, Hatfield, UK,) in goat serum (Sigma-Aldrich), and later labelled using anti-CD3 and anti-CD4 mAbs (both eBioscience) in PBS supplemented with 1% FCS (FACS Buffer). Cells were then washed in FACS Buffer, incubated in 1x Fixation/Permeabilization buffer (eBioscience) for one hour at 4°C, washed again, and then incubated at 37°C in 100 μg DNase (Sigma-Aldrich) for 1 hour. After a final wash in FACS buffer, cells were labelled for 45 minutes at room temperature with anti-BrdU mAb, or rat IgG1 isotype control mAb (both eBioscience), in 1x permeabilization buffer, according to the manufacturer’s instructions.
DEC were first incubated using Fixable Live/Dead Aqua stain (Life Technologies, Paisley, UK), blocked with anti-CD16/32 mAbs (eBioscience) in goat serum (Sigma Aldrich), and subsequently labeled with the following mAbs conjugated to fluorescent labels: anti-CD45, anti-F4/80, anti-MHC-II (IA-IE), anti-Fas, anti-FasL, anti-CD3 and anti-CD4 (all eBioscience). Flow cytometry data was acquired on the Cyan ADP, or the BD LSR Fortessa analyzer (Beckman Coulter, London, UK). Data was analyzed using FlowJo Software v7.6.5 (Tree Star Inc, Oregon Bio, Oregon US).
After surface staining for anti-CD3 and anti-CD4, sdLN cells were washed in cold PBS supplemented with 1x annexin V binding buffer (eBioscience), and incubated for 15 minutes with anti-annexin V FITC at room temperature. Cells were then washed in annexin V binding buffer and resuspended for analysis in annexin V binding buffer. Propidium iodide (PI) (eBioscience) was added directly before acquiring the data.
DEC obtained from infected pinnae in three independent experiments (12–18 mice each) recovered on day 4 after infection with S. mansoni cercariae were pooled and labelled with anti-MHC-II (IA-IE) (clone # M5/114) and anti-F4/80 (clone # BM8) mAbs. Cell populations which were F4/80+MHC-IIhigh, F4/80+MHC-IIlo, or F4/80-MHC-IIhigh were recovered by FACS (MoFlo Astrios, Beckman Coulter) and RNA extracted from sorted DEC populations using TRIzol (Life Technologies). RNA was quantified using a Nanodrop (Thermo Scientific, Waltham, USA) and quality checked using a Bioanalyzer (Agilent Technologies, Santa Clara, USA). Staff at the Technology Facility at the University of York, York, UK, prepared and carried out microarray analysis of purified RNA, including sample labelling and hybridisation, using the Agilent SurePrint system (Agilent Technologies).
GeneSpring (Agilent Technologies) was used to normalize microarray data, calculate fold differences, prepare dendrograms of cell populations and establish significant (p< 0.05) differentially expressed genes.
Statistical analyses were performed using a one-way analysis of variance (ANOVA) and Tukey’s multiple comparisons test using GraphPad Prism v6 software (GraphPad Software Inc, San Diego, California. USA).
Dermal exudate cells (DEC) were recovered from in vitro cultured skin biopsies after exposure to either a 1x dose, or 4x doses, of S. mansoni cercariae in order to characterize genes that were differentially regulated four days after the final infection. Cells which were F4/80+MHC-II-, shown to be SiglecF+ eosinophils [4, 12] were excluded from the analysis as we previously showed that the absence of these cells had no effect on the development of CD4+ cell hypo-responsiveness in the sdLN [7]. Therefore, our microarray analysis focused upon cells of the mononuclear phagocyte system (i.e. DCs and macrophages). These cells distributed into three discrete populations: F4/80-MHC-IIhigh cells (denoted ‘R4’) were classed as DCs, F4/80+MHC-IIhigh cells (denoted ‘R4A’) were termed tissue resident macrophages, whilst F4/80+MHC-IIint cells (denoted ‘R3’) were classified as macrophages [6] as shown in Fig 1A.
Clustering analysis of microarray data obtained from biological replicates of the three sorted DEC populations is shown as dendrograms with corresponding selected heat maps for cells from 1x and 4x infected groups of mice (Fig 1B and 1C). The heat maps displayed highlight a section from the start of the entire microarray heat map, whilst the dendrograms are based upon analysis of all identified genes to yield an overall comparison. This revealed clear clustering patterns within each of the defined cell populations (i.e. R4, R4A, and R3) validating our gating strategy. Based on the dendrogram analysis of the microarray data from the 1x and 4x DEC populations, gene expression profiles within the R4A tissue resident macrophages were more closely associated to R3 macrophages than to R4 DCs (Fig 1B and 1C). Transcriptional distinctions between the R4A and R3 populations were reduced after 4x infection (Fig 1C).
Many identified genes found in the three sorted DEC populations were differentially up-regulated in the 4x compared to the 1x samples, and were linked to alternative activation (e.g. retnla, chi3l3, chi3l4, ccl24, cd209 and cd163 [11, 26–28]) (Fig 1D). Retnla (encoding RELMα) was one of the most highly up-regulated genes in 4x DCs (R4; x121-fold), 4x tissue macrophages (R4A; x16-fold) and 4x macrophages (R3; x80-fold), compared to their 1x counterpart cell populations. Similarly, chi3l3 was up-regulated in both 4x tissue macrophages (R4A; x31-fold) and 4x macrophages (R3; x31-fold), whilst ccl24 was up-regulated in all three cell populations 24-33-fold. The gene for IL-4 was up-regulated in 4x macrophages (R3; x22-fold), whilst il4ra was also up-regulated in these cells (R3; x6-fold). Other genes which featured in the top 20 up-regulated genes included those associated with tissue destruction/wound healing such as igf1 (x8-30-fold) and fn1 (x14-23-fold).
Given the significant changes in gene expression associated with IL-4Rα and RELMα, their roles in the early stage immune response were investigated using Il-4rα-/-, or Retlnα-/- mice. After 4x exposures to infective S. mansoni cercariae, pinnae thickness of WT mice, as well as those deficient in IL-4Rα, was significantly increased compared to 1x exposure (Fig 2A; p<0.0001 and p<0.05 respectively). However, the thickness of the pinnae infection site was comparable between WT and IL-4RαKO mice, irrespective of the infection regime (Fig 2A; p>0.05). Similarly, whilst RelmαKO mice exposed to 4x doses of cercariae displayed significantly enhanced levels of skin inflammation compared to naive and 1x RelmαKO animals (all p<0.0001), the levels of inflammation were comparable with those of their 4x WT cohorts (Fig 2B; all p>0.05).
Increased levels of IL-4, IL-12p40 and IL-10 released by skin biopsies cultured in vitro in the absence of added antigen were detected in the pinnae of 4x compared to 1x WT mice (Fig 2C–2H; p<0.05–0.0001). In the absence of IL-4Rα, the levels of IL-4 in 4x IL-4RαKO mice were comparable to those in 4x WT samples (Fig 2C), on the other hand, the levels of pro-inflammatory IL-12p40 in 4x IL-4RαKO mice were significantly increased compared to 4x WT samples (Fig 2D; p<0.001). The levels of regulatory IL-10 were low in 4x IL-4RαKO mice (Fig 2E), resulting in a significant reduction in the quantities of IL-10 being released compared to 4x WT mice (Fig 2E; p<0.0001). The absence of RELMα had no effect on the production of IL-4, or IL-12p40 in 1x and 4x mice compared to their WT cohorts (Fig 2F and 2G; p>0.05), however similar to 4x IL-4RαKO mice, IL-10 production was significantly lower in 4x RelmαKO skin compared to 4x WT skin (Fig 2H; p<0.05). Therefore, although IL-4RαKO and RelmαKO mice produce similar amounts of IL-4 after 1x and 4x infections, IL-4RαKO mice have a more pro-inflammatory phenotype (i.e. increased IL-12p40) and both KO strains had decreased IL-10 production compared to WT mice.
As expected, significantly greater numbers of DEC were recovered from 4x WT compared to 1x WT mice (Fig 3A and 3B; p<0.01), although as with pinnae thickness, DEC numbers in 1x and 4x infected IL-4RαKO mice were similar compared to their WT cohorts (Fig 3A; p>0.05). The number of DEC recovered from pinnae biopsies of 4x RelmαKO compared to 4x WT mice was also not significantly different (Fig 3B; p>0.05).
The abundance of F4/80+ antigen presenting cells may be critical to the development of CD4+ T cell hypo-responsiveness, particularly as it has been shown that F4/80+ alternatively activated macrophages depress CD4+ cell responses following infection with filarial parasites [29]. Here, we show there were no significant changes in the number of F4/80-MHC-IIhigh (R4), F4/80+MHC-IIhigh (R4A), or F4/80+MHC-IIint (R3) cells in 4x WT compared to 1x WT mice (Fig 3C–3E; all p>0.05). However, in the absence of IL-4Rα signaling, there were large increases in the numbers of these three cell types in 4x IL-4RαKO compared to 1x IL-4RαKO mice (Fig 3C–3E; p<0.05–0.01). Moreover, the number of R4 and R4A cells, which express high levels of MHC-II, were significantly greater in 4x IL-4RαKO than in 4x WT cohorts (Fig 3C and 3D; p<0.05–0.01), although the change in the number of R3 macrophages which express intermediate levels of MHC-II in 4x IL-4RαKO and 4x WT mice was not statistically different (Fig 3E; p>0.05). In RelmαKO mice, the numbers of R3, R4 and R4A DEC were similar to those in WT cohorts, and there was no significant difference in the number of these two cell types after 4x compared to 1x infection (Fig 3F–3H; p>0.05).
Increased numbers of MHC-IIhigh R4 and R4A cells in 4x infected IL-4RαKO mice could support a stronger T cell response, both in the skin site of infection and down-stream in the draining lymphoid tissue [30, 31]. This might be particularly relevant when coupled with low levels of IL-10 production in the skin, a cytokine that we recently showed to be fundamental in regulating CD4+ T cell numbers [6, 8]. Indeed, in the absence of IL-4Rα signaling, significantly greater numbers of CD3+CD4+ T cells were present in the skin infection site of 4x IL-4RαKO, compared to 4x WT mice (Fig 3I; p<0.0001). Therefore, we subsequently investigated the changes in the numbers, viability, and/or responsiveness of CD4+ cells in the sdLN.
After 4x infections, the sdLN of WT mice had significantly fewer cells compared to 1x WT mice (Fig 4A; p<0.05). However, in the absence of IL-4Rα, cellularity was increased, resulting in comparable cell numbers between 1x and 4x infected IL-4RαKO mice which were not significantly different (Fig 4A; p>0.05). Similarly, total sdLN cellularity was also increased in 4x RelmαKO compared to 4x WT mice (Fig 4B; p<0.01), resulting in the number of sdLN cells from 4x and 1x RelmαKO mice being not significantly different (Fig 4B; p>0.05).
There were also decreased numbers of CD4+ T cells in the sdLN of 4x compared to 1x WT mice (Fig 4C; p<0.05), although the number of CD4+ T cells in 4x compared to 1x IL-4Rα KO mice was not significantly reduced (Fig 4C; p>0.05). Therefore, in comparison there were significantly more CD4+ T cells in 4x IL-4RαKO than in their 4x WT counterparts (Fig 4C; p<0.05). The numbers of CD4+ T cells in the sdLN of 1x and 4x RelmαKO mice was also similar (Fig 4D; p>0.05) leading to significantly greater numbers of CD4+ T cells being detected in the sdLN of 4x RelmαKO compared to 4x WT mice (Fig 4D; p<0.05). This showed that the absence of either IL-4Rα, or RELMα, leads to increased number of CD4+ T cells in the sdLN following repeated infection comparable to the levels seen in 1x mice.
It was proposed that the increased number of CD4+ T cells in sdLN of 4x IL-4Rα and 4x RelmαKO mice could be due to a reversal in the development of hypo-responsiveness normally seen in 4x WT mice. In fact, the total number of BrdU+CD4+ cells was slightly greater in 4x IL-4RαKO mice compared to the 4x WT cohort group (Fig 4E; p<0.05), although the number in 4x RelmαKO versus 4x WT mice was not significantly different (Fig 4F; p>0.05). While ~15–20% of the CD4+ T cells in the sdLN from 1x WT mice were BrdU+ and had therefore proliferated in vivo, only ~7% from 4x WT mice were BrdU+ (Fig 4G; p<0.0001) confirming the establishment of CD4+ T cell hypo-responsiveness. However, although there were a slightly greater number of BrdU+ cells in 4x IL-4RαKO mice (Fig 4E), the proportion of BrdU+ cells in 4x IL-4RαKO mice remained significantly lower than in 1x IL-4RαKO mice (Fig 4G; p<0.05) demonstrating that the absence of IL-4Rα does not restore proliferation of CD4+ T cells in the sdLN of 4x mice to the levels seen in 1x WT mice. A similar situation was observed in 4x RelmαKO mice, where the proportions of BrdU+ cells in 4x RelmαKO compared to 4x WT mice were not significantly different (Fig 4H; p>0.05) and both exhibited CD4 T cell hypo-responsiveness in vivo compared to their 1x cohorts (Fig 4H; p<0.0001). This shows that the absence of RELMα also does not restore CD4+ T cell proliferation in the sdLN after repeated exposures of the skin to S. mansoni cercariae.
Previously, we reported that CD4+ T cell hypo-responsiveness in the sdLN observed after repeated infection is dependent on IL-10 [8]. This was due to increased CD4+ T cell activation accompanied by decreased death and apoptosis of the CD4+ T cell population in the sdLN. In the current study, we show that both IL-4RαKO and RelmαKO mice had decreased IL-10 production in the skin and increased cellularity in the sdLN after 4x infection. As RELMα expression is dependent on IL-4 signaling [32], we restricted further analysis of the CD4+ cell population to cells in the sdLN of 1x versus 4x IL-4RαKO mice. We observed that surface protein expression of both Fas and FasL increased in CD4+ T cells recovered from 4x WT compared to 1x WT mice but was significantly decreased in 4x IL-4RαKO compared to 4x WT mice (Fig 5A and 5B; p<0.05 and p<0.0001). In addition, whereas in WT mice, there was a significant decrease in the proportions of AnnV-PI- viable CD4+ T cells in 4x compared to 1x mice (Fig 5C; p<0.01), the absence of IL-4Rα signaling resulted in significantly greater proportions of AnnV-PI- viable CD4+ T cells in 4x IL-4RαKO compared to 4x WT mice (Fig 5C; p<0.01). This caused there to be no significant difference between the viability of CD4+ T cells from 4x IL-4RαKO mice compared to either 1x WT, or 1x IL-4RαKO mice (Fig 5C; p>0.05). The increase in the proportion of AnnV-PI- viable CD4+ T cells in 4x IL-4RαKO mice was accompanied by the detection of significantly fewer AnnV+PI- apoptotic CD4+ T cells (Fig 5D; p<0.05), as well as fewer AnnV+PI+ dead CD4+ T cells, compared to 4x WT mice (Fig 5E; p<0.01). Thus, it appears that a reduction in IL-4Rα signaling facilitates CD4+ T cell survival. This increased survival could explain why the number of CD4+ T cells in the sdLN of 4x IL-4RαKO mice is not significantly reduced after repeated exposure to schistosome cercariae which contrasts with the situation in 4x WT mice where there is a significant reduction in number of CD4+ T cells when compared to their 1x WT counterparts.
Here we demonstrate that three discrete mononuclear phagocyte cell populations were present in DEC recovered from the skin infection site of mice exposed to repeated doses of schistosome cercariae. Moreover, all three populations exhibited an alternatively activated phenotype as judged by microarray analysis of DEC. For example, Retnla (encoding RELMα) which is a key marker of alternatively activated macrophages [33, 34] was one of the most highly up-regulated genes in 4x compared to 1x DCs, tissue macrophages, and macrophages. Several other genes also linked to alternative activation were up-regulated in the three discrete 4x DEC populations including chi3l3 and chi3l4 (encoding chitinase-like molecules YM-1 and YM-2 [12, 33]). The gene for IL-4 was also up-regulated in the macrophage DEC population after repeated infections, and IL-4 has been shown to be produced by alternatively activated/type-II macrophages using both human and mouse cells [35, 36]. Other up-regulated genes linked to alternative activation were mrc1 and cd209 which encode the C-type lectins for the mannose receptor CD206 and DC-SIGN CD209 [9], whilst expression of cd163 encoding the scavenger receptor CD163 when used in combination with chi3l3 and Retnla [11, 12, 37] also supports the identification of the three discrete 4x DEC populations as being alternatively activated. Conversely, Arg1 which we had previously reported to be up-regulated in DEC from 4x mice [4], was not >2 fold up/down-regulated in the any of the 3 sorted cell populations, whilst Clec10a of the C-type lectin/C-type lectin-like domain (CTL/CTLD) superfamily, associated with alternative activation [9] was only 2.2 fold up-regulated in the DC population. On the other hand, wound healing associated genes such as igf-1 encoding Insulin-like growth factor 1 suggested to be involved in resolving tissue damage following helminth infection [38], and fn1 encoding fibronectin produced during the resolution of tissue damage [39] were both highly up-regulated. This provides evidence supporting an evolutionary link between alternative activation in mononuclear phagocytes and the resolution of tissue damage caused by helminth infection [16, 17, 38]. Indeed, we have shown that schistosome cercariae cause significant tissue disruption as they invade through the skin [3], and therefore tissue repair of the skin following repeated exposure to invasive schistosome cercariae should be expected.
The discrete DEC mononuclear phagocyte populations all express genes typically associated with alternative activation [9–12]. This provides a more in-depth analysis of DEC from 4x mice in which mononuclear phagocytes appeared to switch from classically-activated to alternatively-activated commensurate with up-regulated mRNA transcripts for Ym1 and RELMα but only low levels of iNOS and IFNγ [4]. Nevertheless, whilst the R4, R4A, and R3 mononuclear phagocytes all appear to be alternatively activated, subtle qualitative differences in the expression of specific genes between the three DEC populations underline the likely heterogeneity/plasticity of these mononuclear phagocytes which may have different and/or overlapping functional roles in vivo [40–42].
In the context of repeat infection of the skin with S. mansoni cercariae, DEC recovered from the site of infection had a pronounced increase in the expression of RELMα but its absence had no effect on the extent of skin inflammation (i.e. pinnae thickness), or on the numbers of mononuclear phagocyte DEC populations recovered from the skin infection site (Fig 3F–3H). The absence of RELMα did however result in a slight reduction in IL-10 production in the skin alongside an increase in the numbers of CD4+ cells in the sdLN such that they were as abundant as in 1x RELMαKO mice and were more numerous than in 4x WT cohorts. However, the absence of RELMα did not affect the development of CD4+ T cell hypo-responsiveness in the sdLN of 4x mice. In contrast, a separate study showed that RELMα KO mice infected with Nippostrongylus brasiliensis exhibit enhanced intestinal and pulmonary pathology accompanied by expulsion of the parasite indicating a regulatory role for RELMα by dampening normally protective strong Th2-dependent responses [32, 43]. Moreover, pulmonary immune granulomas to injected S. mansoni eggs were enhanced in the absence of RELMα [44], as were hepatic immune granulomas formed at the chronic phase of schistosome infection [32]. Nonetheless, these previous studies were performed when the immune response was skewed towards a Th2 phenotype [21]. In contrast, the response investigated in our study occurs at an early phase of infection, prior to egg-induced Th2 biased immunopathology, and is accompanied by cytokines with a mixed Th1 and Th2 type response [4]. Therefore, RELMα may only have a regulatory role during strong Th2 responses during the chronic phase of infection and we suggest that it has less of a role when the immune response has a mixed Th1/Th2 phenotype.
Signaling through IL-4Rα is well defined as being important in the development of alternative activation [10, 12] and in the maintenance of Th2 responses [45]. Whilst the absence of IL-4Rα did not have an impact on inflammation at the skin site of infection, nor on the number of DEC, its absence significantly increased the numbers of cells expressing high levels of MHC-II. In addition, increased release of pro-inflammatory IL-12p40 in the absence of IL-4Rα was accompanied by decreased levels of regulatory IL-10. Therefore, we considered it possible that the increased numbers of cells with antigen presenting potential in the skin might lead to enhanced CD4+ cell activity downstream in the sdLN. However, whilst the absence of IL-4Rα resulted in increased numbers of CD4+ cells in the skin infection site and the sdLN, the cells in the sdLN remained hypo-responsive in vivo to stimulation with parasite antigen similar to 4x WT mice. The use of cell specific gene deficient mice, such as those expressing a Cre recombinase from the lysozyme M-encoding locus, which has been widely used in the context of IL-4Rα function in macrophages and in the context of schistosome infection [22], would have been desirable to fully interrogate the role of particular genes on mononuclear phagocytes. However, such mice were not available during the current project. In addition, it has been reported that Il4rα excision in these mice is incomplete during inflammatory conditions [46], raising doubts about interpretation of data obtained using these mice and suggests that alternative cell specific gene animals should be sought.
Further analysis of CD4+ cells in the sdLN revealed that there was a significant elevation in the expression of Fas and FasL on the CD4+ T cells in the sdLN in 4x infected WT mice, whereas in the absence of IL-4Rα the expression levels of these two molecules was not elevated and was not different from those in 1x IL-4Rα mice. We also showed that skin biopsies from 4x IL-4RαKO mice, in contrast to 4x WT cohorts, released negligible quantities of IL-10 suggesting a possible role for IL-4Rα in the promotion of IL-10 production. Indeed, IL-4Rα signaling is required for the production of IL-10 derived from Th2 cells following infection with N. brasiliensis, thereby resulting in increased levels of regulation via IL-10 [47], although others show that IL-10 production following chronic schistosome infection can be IL-4Rα-independent [48]. Here, in our study we found that IL-10 production in the skin was significantly reduced in the absence of IL-4Rα, potentially resulting in decreased regulation via IL-10.
Several studies have identified a link between IL-10 and Fas/FasL expression. For example in systemic lupus erythematosus (SLE), IL-10 is directly able to induce the expression of Fas on the surface of T cells resulting in increased levels of apoptosis [49–51]. Moreover, we recently showed that in IL-10KO mice, there was reduced Fas expression on CD4+ T cells in the sdLN following repeated schistosome infection which led to a reduction in CD4+ T cell death in the sdLN and consequently may have contributed to the alleviation of CD4+ T cell hypo-responsiveness in the absence of IL-10 [8]. Here, we suggest a novel mechanism for the regulation of the immune response through IL-4Rα, which impacts both IL-10 production and antigen presenting cells numbers, which would subsequently regulate Fas and FasL expression on CD4+ T cells in the sdLN of 4x infected mice. Thus, IL-4Rα signaling results in increased IL-10 production, increased levels of apoptotic and/or dead T cells in 4x mice and a dampening of the immune response. This link between IL-4Rα and IL-10 and Fas/FasL-induced apoptosis could be a potential novel mechanism through which IL-4Rα regulates the immune system.
|
10.1371/journal.ppat.1006391 | Microbiota-induced peritrophic matrix regulates midgut homeostasis and prevents systemic infection of malaria vector mosquitoes | Manipulation of the mosquito gut microbiota can lay the foundations for novel methods for disease transmission control. Mosquito blood feeding triggers a significant, transient increase of the gut microbiota, but little is known about the mechanisms by which the mosquito controls this bacterial growth whilst limiting inflammation of the gut epithelium. Here, we investigate the gut epithelial response to the changing microbiota load upon blood feeding in the malaria vector Anopheles coluzzii. We show that the synthesis and integrity of the peritrophic matrix, which physically separates the gut epithelium from its luminal contents, is microbiota dependent. We reveal that the peritrophic matrix limits the growth and persistence of Enterobacteriaceae within the gut, whilst preventing seeding of a systemic infection. Our results demonstrate that the peritrophic matrix is a key regulator of mosquito gut homeostasis and establish functional analogies between this and the mucus layers of the mammalian gastrointestinal tract.
| When a female mosquito takes a blood meal from a human, the bacteria residing within its gut grow significantly. Following a blood meal, female mosquitoes produce a barrier within their gut, known as the peritrophic matrix, which physically separates the blood meal from the cells of the epithelium. Here, we show that the presence of bacteria in the gut is required for the synthesis of the peritrophic matrix. By experimentally disrupting this barrier, we find that this structure plays a role in limiting the extent to which bacteria of one particular family are able to grow and persist in the mosquito gut. We also find that the peritrophic matrix ensures that bacteria remain within the gut, preventing them from invading the mosquito body cavity. These results will be useful in designing disease control strategies that depend on the ability of bacteria to colonize and persist in relevant tissues in the mosquito host.
| Mosquitoes of the Anopheles genus are responsible for the transmission of Plasmodium parasites, the causative agents of malaria. The study of the Anopheles gut microbiota has recently emerged as an important field in an effort to characterize mosquito-parasite interactions in greater depth and to develop new methods to stop disease transmission. The microbiota have been shown to trigger a constitutive immune response in the mosquito gut epithelium that enhances resistance to parasite infection [1,2]. Furthermore, specific gut bacteria have been found to directly impact parasites, compromising their infectivity [3,4]. Finally, a promising transmission-blocking intervention is paratransgenesis, which aims to use vector-associated bacteria as a delivery tool for antimalarial effectors [5]. The success of such an approach would require the persistence of genetically-modified bacteria at sufficient abundance within the gut ecosystem, and potentially their successful dissemination throughout the mosquito body. As such, deeper understanding of mosquito-microbiota interactions may highlight mechanisms by which bacteria can be utilized to block malaria transmission.
The balance between immune resistance and tolerance is key to bacterial persistence within the gut environment. Resistance refers to bacterial killing or the prevention of bacterial growth, whilst tolerance encompasses the prevention or repair of host tissue damage caused by pathogens or immune responses [6]. In Drosophila, commensals are controlled largely by the production of reactive oxygen species (ROS) by the dual oxidase (DUOX) enzyme [7,8], whilst the other main resistance mechanism in the Drosophila gut, the Imd pathway, is under strong negative regulation to prevent its stimulation by commensals [9–11]. In mosquitoes, blood feeding triggers substantial microbiota proliferation [12,13] and induces high levels of oxidative stress, potentially precluding further production of ROS for immune control [14]. In A. gambiae, commensals are known to induce the Imd pathway, and suppression of the Imd pathway receptor PGRPLC and its transcription factor REL2 causes microbiota overgrowth [1,2]. The transcription factor Caudal, which is specifically expressed in the gut, down-regulates REL2-dependent expression of antimicrobial peptides (AMPs), facilitating microbiota tolerance [15].
Some tolerance mechanisms are based on the strengthening of physical barriers between the microbiota and the host. Notably, an A. gambiae heme peroxidase is induced by blood feeding and, together with DUOX, forms a network of dityrosine bonds that is thought to protect the gut epithelium from immune elicitors, thus mediating bacterial persistence [12]. The peritrophic matrix has also been identified as playing a role in host-bacteria interactions in a number of insects. It is an acellular structure composed of chitin, proteins and glycoproteins located between the gut lumen and the epithelium. The mosquito type I peritrophic matrix is produced by adult female midgut cells during blood feeding and physically surrounds the blood bolus, whilst the type II peritrophic matrix is permanently produced by the cardia in the anterior larval gut. The type II peritrophic matrix of the hematophagous tsetse fly provides infectious Serratia bacteria with a protective niche in which they can proliferate without inducing a gut immune response, increasing susceptibility to infection [16]. In the tick Ixodes scapularis, the gut microbiota induces the formation of a peritrophic matrix whose presence facilitates colonization of the spirochete bacterium Borrelia burgdorferi, possibly by protecting the pathogen from blood meal pro-oxidants and cellular immunity [17]. In Drosophila, oral bacterial infection induces the expression of genes encoding proteins with chitin binding domains (CBDs) [18], and a protein of the type II peritrophic matrix is shown to reduce both local and systemic Imd pathway stimulation and to protect epithelial cells against pore-forming toxins [19,20].
The mosquito peritrophic matrix is often considered as a barrier to parasite infection, though one that parasites have evolved to overcome. Secretion of chitinase by Plasmodium effectively facilitates traversal of the peritrophic matrix [21–25]. A constitutive peritrophic matrix protein, fibrinogen-related protein 1 (FREP1), has recently been proposed to be exploited by invading parasites, serving as an anchor that facilitates P. falciparum invasion [26]. More generally, the Aedes aegypti peritrophic matrix is thought to play a role in blood meal detoxification, sequestering large quantities of heme released during blood bolus digestion [27]. The role of the mosquito peritrophic matrix in bacterial pathogenesis and microbiota homeostasis in the gut has not yet been explored.
Here, we use RNA sequencing to explore the microbiota-dependent gene expression in the midgut of the A. coluzzii mosquito (until recently known as A. gambiae M form). We find that the gut microbiota induce the expression of several components of the peritrophic matrix, and that the microbiota are necessary for the synthesis of a structurally complete peritrophic matrix. We also show that the peritrophic matrix plays a role in resistance to the Enterobacteriaceae bacteria present in the gut microbiota, both reducing the extent to which this family of bacteria grows and persists within the gut, and precluding this family of bacteria from seeding a systemic infection.
To explore the transcriptional response to the dynamic changes in microbiota load over the blood feeding cycle, we sequenced RNA extracted from A. coluzzii midguts at five time points: 2–3 day old mosquitoes that had accessed only fructose since emergence (sugar-fed, ‘SF’), 5h, 24h and 72h after a human blood meal and 24h (96h) after a second human blood meal that was given at the 72h time point. This time course was performed with conventionally-reared mosquitoes that harbored their native microbiota, and a cohort of mosquitoes that were fed an antibiotic cocktail (50μg/ml gentamicin, 60μg/ml streptomycin, 60U/ml penicillin) in both sugar and blood meals. This antibiotic treatment was effective in substantially depleting mosquito guts of bacteria as detected by qRT-PCR against 16S rRNA (Fig 1A). Each sample consisted of a pool of 20 midguts, and four independent replicates were performed using four independent batches of mosquitoes, as there is evidence that the microbiota of laboratory-reared mosquitoes varies between generations [1]. The resulting cDNA libraries were sequenced across four lanes of an Illumina flowcell on an Illumina Hiseq 1500, resulting in a total of 893,247,801 pairs of 100bp reads across the forty samples. After quality control, an average of 85.6% of input sequences per sample aligned uniquely to the A. gambiae PEST genome (AgamP4). A total of 6753 genes (49.9% of all annotated genes) had non-zero counts in all forty samples. Principal component analysis (PCA) indicated that the samples clustered according to their blood feeding status, with no obvious outliers (Fig 1B).
Soft clustering analysis indicated that the oral antibiotic treatment had an overall relatively minor effect on the general transcriptional changes occurring over the blood feeding cycle (S1 Fig). Nevertheless, we identified 889 genes that were significantly differentially regulated at one or more time points by antibiotic treatment (S1 File). Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis implicated these genes in diverse processes, including carbohydrate, protein and lipid metabolism, folate biosynthesis, oxidation-reduction processes and immunity (S1 Table). Interestingly, the 0h and 72h samples exhibited the greatest number of differentially regulated genes (Fig 1C), despite having the lowest bacterial load in the control samples. We hypothesised that this could be indicative of the microbiota playing a more significant role in midgut physiology at these time points, or of the existence of highly effective tolerance mechanisms in the gut following blood feeding.
As observed previously [1], we noted that several microbiota-regulated genes encoded proteins containing CBDs, a signature of the structural components of the peritrophic matrix. Although the precise structure of the peritrophic matrix remains under-explored, a proteomic analysis has previously identified its most abundant protein components [28]. Of genes encoding 24 of the top candidate proteins identified in that study, 12 were significantly differentially regulated in our dataset, with 11 of these being down regulated following antibiotic treatment (S2 Table). The microbiota-induced genes included AgAPER1 (AGAP006795; Fig 2A), which encodes a chitin-binding A. gambiae peritrophic matrix component [29] and was the most abundant CBD-containing protein identified by mass spectrometry [28]. We also noted the microbiota-dependent expression of ICHIT (AGAP006432) that is known to be transcriptionally induced by both P. berghei and bacterial infections and encodes two CBDs and a proline-rich domain that may be involved in protein-protein aggregation [30] (Fig 2A). Two genes (AGAP009313 and AGAP006194) encoding proteins identified in the peritrophic matrix proteomic study [28] were significantly microbiota regulated at all five time points (Fig 2A); neither of these genes encode CBD-containing proteins.
In addition to the protein components of the peritrophic matrix, the main structural constituent is chitin, a polymer of N-acetylglucosamine. Insects are able to synthesize chitin from glucose in a multistep reaction (Fig 2B); fructose-6-phosphate, derived from glucose, is converted to glucosamine-6-phosphate in a rate limiting step catalyzed by glucosamine-fructose-6-phosphate aminotransferase (GFAT) [31]. Glucosamine-6-phosphate is then metabolized to UDP N-acetylglucosamine, which is subsequently polymerized to chitin fibers by chitin synthase. The A. gambiae genome encodes two chitin synthase enzymes, CHS1 (AGAP001748) and CHS2 (AGAP001205), of which CHS2 is expressed in the midgut and responsible for the synthesis of peritrophic matrix-associated chitin [32]. Here, we identified GFAT and CHS2, the enzymes catalyzing two rate-limiting steps of the chitin synthesis pathway, as being microbiota-regulated at the transcript level at one or more of the time points examined (Fig 2C). Following antibiotic treatment the expression of these enzymes is either significantly reduced or temporally delayed.
We speculated that this transcriptional response in the production of both chitin and peritrophic matrix proteins could be bacterially induced either directly, through detection of bacterial elicitors, or indirectly as a result of bacteria causing thinning of the peritrophic matrix (e.g. through chitinases) that results in compensatory transcription to produce a peritrophic matrix of normal thickness. We used thin abdominal sections 24h post blood meal stained with hematoxylin and eosin (H&E) to observe the effect of oral antibiotic treatment on the structure of the gut tissue (Fig 2D). In the control samples, a thick layer of dark pigment was observed surrounding the blood bolus and effectively separating the contents of the lumen from the epithelial cells. We hypothesized that this dark layer is likely an accumulation of heme pigment at the surface of the peritrophic matrix. Indeed, heme released during hemoglobin digestion is shown to bind CBD-containing peritrophic matrix proteins in Aedes aegypti mosquitoes [27,33,34]. In the antibiotic-treated mosquitoes, several regions exhibited disruption of this layer, suggestive of a similar disruption of the peritophic matrix, resulting in red blood cells (RBCs) coming into direct contact with the epithelial cells (Fig 2D). Quantification of such instances confirmed that RBC contact with the epithelium is significantly increased in antibiotic-fed (92%, n = 12) compared to control guts (6%, n = 18; p<0.0001, Chi-square test). To confirm the disruption of the peritrophic matrix, we stained abdominal sections with the chitin specific stain calcofluor white (Fig 2E). In both control and antibiotic treated samples, chitin specific staining of the cuticle was observed. In control mosquitoes, we additionally observed a prominent layer of chitin staining surrounding the blood bolus, which corresponds to the peritrophic matrix. In the antibiotic treated group this staining was either absent or fragmented. These observations suggest that, indeed, the presence of the microbiota is required for the synthesis of a structurally complete peritrophic matrix.
The RNAseq data indicated that the microbiota play a significant role in regulating antimicrobial peptide (AMP) expression in the gut, with seven characterized immune effector-encoding genes being upregulated at one or more time points by the presence of the microbiota (S3 Table). These AMPs include three cecropins (CEC1, CEC2 and CEC3), one defensin (DEF1), gambicin (GAM1) and two C-type lysozymes (LYSC1 and LYSC7). We therefore sought to explore whether the peritrophic matrix plays a role in mediating the mosquito immune response to the microbiota. We supplemented the blood meal with 100μM polyoxin D, a chitin synthase inhibitor that has previously been demonstrated to abolish synthesis of the A. gambiae type I peritrophic matrix [24]. Staining of abdominal mosquito sections with calcofluor white at 24h post blood meal confirmed the absence or fragmentation of the peritrophic matrix upon treatment with polyoxin D (S2A Fig). In the midguts of polyoxin D-fed mosquitoes, we also observed increased RBC contact with the epithelium (47%, n = 17) compared to control guts (19%, n = 32; p<0.05, Chi-square test; S2B Fig).
We next investigated whether the peritrophic matrix plays a role in modulating the midgut epithelium response to the microbiota. We selected two AMP reporter genes, CEC1 and GAM1, which showed strong microbiota-dependent expression at all time points examined (Fig 3A and S3 Table). The midgut expression of the two AMP genes was monitored at 24h after blood meal supplementation with 100μM polyoxin D or an equal volume of water as a control. We observed a significant increase in the expression of GAM1 in the midguts of polyoxin D-treated mosquitoes (Fig 3B), whilst the increase in CEC1 expression between polyoxin D treated and untreated mosquitoes was not statistically significant (S3A Fig). Treatment of mosquitoes with antibiotics revealed that the increase of GAM1 expression was microbiota-dependent (Fig 3B). In order to confirm that the effect of polyoxin D on the immune response in the gut was due to disruption of the peritrophic matrix as opposed to any direct effect of polyoxin D on gut bacteria or the cells of the epithelium, we sought an independent method of peritrophic matrix disruption. To this end, we silenced by RNAi the APER1 gene that encodes an abundant peritrophic matrix component. The results showed an elevated immune response in the midguts of APER1 knock down mosquitoes, corroborating the polyoxin D feeding experiments (S3B Fig); in this case, CEC1 expression was significantly increased, whilst GAM1 expression also showed non-significant up-regulation. This effect was again microbiota dependent (S3B Fig).
These data raised the question whether disruption of the peritrophic matrix affects tolerance or resistance mechanisms. In the former case, the elevated AMP expression in peritrophic matrix-disrupted midguts would reflect increased access of bacteria and immune elicitors to the innate immune receptors found on the epithelial cells, which could consequently result in decreased bacterial growth. In the latter case, disruption of the peritrophic matrix could relieve bacterial growth from biochemical and/or physical constraints, which may result in a higher bacterial load, consequently increasing AMP induction. We quantified the total bacterial load in the polyoxin D-treated compared to control midguts as well as the specific load of three bacterial families commonly found in Anopheles midguts: Enterobacteriaceae, Flavobacteriaceae and Acetobacteraceae [35]. We did not detect a significant difference in total bacterial load nor in the load of bacteria of the Flavobacteriaceae and Acetobacteraceae families between the two groups (S3C Fig). However, a significant increase was detected in the load of the Enterobacteriaceae, which was highly variable between midgut pools (Fig 3C). Corroborating this, we also observed substantial Enterobacteriaceae overgrowth in a subset of the APER1 knock down mosquito cohorts compared with the LACZ controls (S3D Fig). These data point to a role of the peritrophic matrix in resistance to the Enterobacteriaceae.
To further characterize the role of the microbiota in AMP regulation, we examined the correlation between the bacterial loads and GAM1 expression in each of the mosquito pools used in the polyoxin D experiments described above. In control mosquitoes, GAM1 expression positively correlated with both the Enterobacteriaceae and Acetobacteraceae loads, but not with the total bacterial load or the Flavobacteriaceae load (Fig 3D). These data suggest that the Enterobacteriaceae and Acetobacteraceae families are primarily responsible for the induction of GAM1. In the polyoxin D-fed mosquitoes, the positive correlation between Enterobacteriaceae load and GAM1 expression was maintained, but not that between Acetobacteraceae load and GAM1 expression (Fig 3D). These data are consistent with a model whereby formation of the peritrophic matrix serves as a resistance mechanism, limiting the growth of the Enterobacteriaceae after blood feeding and the extent to which this family of bacteria induces a local immune response.
By 72h after the blood meal, we observed that gut microbiota load had been restored to pre-blood feeding levels (Fig 1A) [36]. We sought to understand the mechanisms underlying this re-establishment of homeostasis following blood feeding, hypothesizing that the excretion of the blood bolus may additionally facilitate the physical removal of bacteria from the gut. To investigate this, we monitored individual mosquitoes 48h after blood feeding, dividing them into two groups according to whether or not they had excreted their blood bolus, and analyzed the gut bacterial load in each group (Fig 4A). We found that, indeed, mosquitoes that had excreted their blood bolus had 98% lower bacterial load than those that still retained their blood bolus at this time point. These results strongly suggested that bacteria are excreted with the blood bolus, thus contributing to the restoration of gut homeostasis.
Upon completion of blood digestion, the peritrophic matrix is believed to be excreted with the blood bolus [37]. To investigate whether the peritrophic matrix plays a role in mediating bacterial excretion, we monitored the effect of peritrophic matrix disruption on the bacterial load at 72h post blood feeding, a time point at which all individuals have excreted their blood bolus. The polyoxin D-fed cohort of mosquitoes harbored significantly higher loads of Enterobacteriaceae and Acetobacteraceae than the control cohort (Fig 4B), as well as non-significantly higher loads of Flavobacteriaceae and total 16S rRNA (S4 Fig), suggesting that the peritrophic matrix prevents bacteria from occupying niches within the gut that cannot be cleared upon excretion of the blood bolus. We performed immunohistochemistry against lipopolysaccharide (LPS), a major component of the outer membrane of Gram-negative bacteria, to investigate bacterial localization in the mosquito gut. We observed the majority of staining in the periphery of the gut, suggesting possible co-localization of the gut bacteria with the peritrophic matrix (Fig 4C).
We next used Gram staining to investigate this localization further in both control and polyoxin D treated guts. In the control guts, we observed bacteria localizing between the blood bolus and the epithelial cell layer (Fig 4D). In the polyoxin D treated guts, bacterial localization was more diffuse, with bacteria being observed at the periphery of the blood bolus, and indeed proximally to the cells of the epithelium, as well as within the gut lumen, amongst the blood bolus. Together, these data are suggestive of a model whereby the presence of an intact peritrophic matrix facilitates the efficient clearance of bacteria from the gut after blood bolus digestion, while co-localization of the gut bacteria and the peritrophic matrix may be a pre-requisite of this.
In Drosophila and other insects, one function of a local gut immune response is to prevent or minimize systemic immune induction arising from an oral infection. We investigated whether the peritrophic matrix plays a role in preventing the induction of a systemic response to microbiota growth within the gut following a blood meal. For each pool of dissected midguts we collected the associated carcass samples, consisting of the abdominal cuticle with the fat body attached, after removal of all other organs, namely the gut, ovaries and malpighian tubules. At 72h post blood feeding, we observed that the gut microbiota induced considerable systemic induction of CEC1 in a subset of the polyoxin D-fed mosquitoes (Fig 5A), with GAM1 expression exhibiting a similar trend (S5A Fig). The same effect was also observed in APER1 knock down mosquitoes compared to the LACZ double stranded RNA-injected control, though in this case at 24h after the blood meal, where a significant increase in the expression of both GAM1 and LYSC1 was observed (S5B Fig). Again, this systemic immune response was fully dependent on the presence of the microbiota (S5B Fig).
Mechanistically, we considered that this could occur either via translocation of live bacteria from the gut to the hemocoel, or via bacteria or mosquito-derived molecules signaling from the gut to the hemocoel. We focused our analysis on peritrophic matrix disruption by polyoxin D feeding, as the APER1 knock down cohort had sustained damage to the carcass during the injection process, providing a possible route of entry for exogenous bacteria. To investigate the former scenario, we attempted to amplify 16S rRNA from the carcass samples of the control and polyoxin D fed cohorts (Fig 5B). In the antibiotic-treated mosquitoes, 16S rRNA amplification was insignificant, at the level of the qRT-PCR negative controls, whereas 16S rRNA was amplified above this level in a subset of both the non-antibiotic-treated control and polyoxin D-fed samples. We observed no significant difference in the relative total 16S rRNA that was amplified from the carcasses of the control or polyoxin D-fed mosquito cohorts at 72h post blood feeding (Fig 5B).
Given that we observed an increase in the Enterobacteriaceae load in the gut at 24h post blood feeding in the polyoxin D fed cohort, and no difference in the overall bacterial load detected in the control and peritrophic matrix disrupted carcasses, we hypothesized that the systemic immune induction could be due specifically to translocation of this bacterial taxon. Indeed, at 72h post blood feeding we found a significant increase in the Enterobacteriaceae load in the peritophic matrix disrupted cohorts, with this family only being confidently detected in polyoxin D-fed pools of mosquito carcasses and not in the control pools (Fig 5C). No significant difference was observed in the incidence or load of the Flavobacteriaceae or Acetobacteraceae (S6C Fig). Furthermore, in the polyoxin D-fed group, Enterobacteriaceae detection in the carcass correlated significantly with CEC1 induction, which was not the case in the control group (Fig 5D). Flavobacteriaceae and Acetobacteraceae abundance did not correlate with CEC1 expression (S5C Fig). The clear relationship between Enterobacteriaceae load and CEC1 expression, together with the fact that this family of bacteria is detected only in the carcasses of polyoxin D-fed mosquitoes, strongly suggests that this family of bacteria is able to translocate from the gut to the hemocoel upon disruption of the peritrophic matrix, seeding a systemic infection.
The data presented here reveal a complex and dynamic relationship between the midgut microbiota, the type I peritrophic matrix and local and systemic immune responses in adult female mosquitoes. In mammals, the inner and outer mucus layers of the gastrointestinal tract are composed of mucin glycoproteins and form a physical and biochemical barrier between the gut flora and the epithelial cells [38]. The mucus layer is at its thickest in the distal colon, the region of highest bacterial colonization, where it functions as a scaffold for AMPs and immunoglobulin A, and acts to protect against microbiota contact with the epithelium [39,40]. The outer mucus layer is known to interact with intestinal microbes [41], providing a habitat for O-glycan foraging taxa [42]. Thus, the defensive nature of mucus is dependent on the entrapment of microbes in the outer layer, but this equally facilitates microbial colonization in acting as a source of nutrition. The mosquito peritrophic matrix is structurally analogous to the vertebrate mucus layer, containing heavily glycosylated proteins, though in this case cross-linking chitin. Here, we reveal additional functional analogies between these two evolutionarily diverse biological structures in limiting gut microbiota growth and precluding bacterial invasion of the intestinal epithelia.
The type I peritrophic matrix is specifically produced by adult female midgut cells upon blood feeding and physically surrounds the blood bolus where blood digestion takes place. We show that synthesis of a structurally complete type I peritrophic matrix is dependent on the presence of the gut bacteria. Similar observations have been reported in other insects. In adult Drosophila, oral infection by Erwinia carotovora carotovora (Ecc15) has been shown to induce the expression of genes encoding proteins with CBDs, with the induction of a proportion of these genes being dependent on the Imd pathway, one of the main immune signalling pathways in the fly gut [18]. Similarly, a protein constituent of the adult Drosophila peritrophic matrix, Drosocrystallin, is induced by oral bacterial infection [19]. The presence of the gut microbiota in the tick I. scapularis has also been shown to be necessary for production of a peritrophic matrix of proper thickness, with this being dependent on STAT signalling [17]. Finally, a proteomic analysis of the type II peritrophic matrix of the tsetse fly identified 27 proteins derived from the secondary endosymbiont Sodalis glossinidius, suggesting similar interaction of this bacterium with the peritrophic matrix [43].
In the mosquito gut, the microbiota is also known to induce the Imd pathway [1,2]. The JAK/STAT pathway, which is activated in response to viral infections [44,45], is known to be responsive to bacterial infection [46] but has not yet been characterized as being microbiota responsive. It remains unclear which signalling pathway(s) are responsible for the microbiota-dependent induction of the A. coluzzii peritrophic matrix, though it is noteworthy that a number of regulated genes have canonical STAT binding sites in their upstream regions (S4 Table), indicating a potential role for the JAK/STAT pathway. Importantly, we observed significant microbiota-dependent regulation of the FoxO signalling pathway, which has recently been shown to facilitate bacteria-dependent synthesis of AMPs in Drosophila enterocytes [47] and may therefore also be considered a candidate pathway for peritrophic matrix induction.
Disruption of peritrophic matrix synthesis resulted in elevated load of the Enterobacteriaceae family of bacteria. Given that we observed bacteria co-locating with the peritrophic matrix, it is likely that the peritrophic matrix has antibacterial properties, whether by direct interaction or by sequestration within a hostile niche. Indeed, there is evidence that peritrophins from other organisms are able to interact directly with bacteria and have antibacterial functions [48,49], while a properly structured peritrophic matrix can also be a scaffold maintaining AMPs and other immune factors in the gut [38,50]. Interestingly, we found that AMP expression in the midgut correlated with the load of the Enterobacteriaceae and the Acetobacteraceae families, but not with the Flavobacteriaceae. This could suggest that the Flavobacteriaceae occupy niches within the gut that are not surveyed by the immune system, whilst the Enterobacteriaceae and Acetobacteraceae live more proximally to the epithelium, likely within or upon the peritrophic matrix. It remains unclear why the load of the Enterobacteriaceae but not the Acetobacteraceae is limited by the presence of the peritrophic matrix after blood feeding. This observation is, however, consistent with a previous study that found Asaia, a major constituent of the Acetobacteraceae family in our mosquito colony, to be resistant to the mosquito immune response [51].
Intriguingly, we observed bacteria present throughout the ectoperitrophic space, including proximally to the epithelium. This could suggest that the bacteria that are directly associated with the peritrophic matrix are efficiently excreted with the blood bolus, whilst those located in the ectoperitrophic space remain in the gut, seeding the bacterial population in the next gonotrophic cycle.
In addition to local effects in the gut, we observe the induction of a systemic immune response upon disruption of the peritrophic matrix and show that this is associated with the break down in the integrity of the gut barrier and translocation of bacteria of the Enterobacteriaceae family into the body cavity. In humans, bacterial translocation is thought to be a common occurrence in healthy individuals and can present as a complication in the critically ill [52]. It is understood to be a consequence of bacterial overgrowth [53], disruption of the mucosal barrier [54] and impaired immune defense [55]. In Drosophila, it has recently been demonstrated that the enteric nervous system controls peritrophic matrix permeability, and that when permeability is compromised flies succumb to bacterial dissemination throughout the body following oral bacterial infection [56]. In concurrence with this, our data suggest that the peritrophic matrix is a key barrier to prevent or limit translocation of microbiota-derived Enterobacteriaceae into the mosquito body cavity.
In our A. coluzzii colony, abundant genera of the Enterobacteriaceae family include Cedecea, Enterobacter, Ewingella and Serratia [36]. Species of Enterobacter have been demonstrated to invade epithelial cells [57] and Serratia marcescens is a model pathogen in Drosophila that, when introduced orally, is able to traverse the gut epithelium [58]. More generally, Enterobacteriaceae have been associated with the induction of colitis, or inflammation of the colon, in mice [59]. Bacterial translocation has not been formally demonstrated in the mosquito, but certain lines of evidence point to this phenomenon. Importantly, a GFP-tagged Asaia strain is known to be able to colonize both the reproductive organs and salivary glands when fed to adults in a blood or sugar meal [60]. Furthermore, knock down of immune effectors can result in bacterial proliferation in the hemolymph even in the absence of infection, though in both of these cases the cuticle was damaged by RNAi injection, providing a possible route of entry for exogenous bacteria [61,62]. Taken with the results presented here, these data suggest that at least some native constituents of the mosquito gut microbiota are able to disseminate throughout the body, and that the peritrophic matrix plays a key role in limiting this dissemination.
The A. coluzzii Ngousso colony was used in all experiments described here. Eggs were hatched in 0.1% salt water and larvae fed Tetramin or Nishikoi fish food. All adults were allowed ad libitum access to 5% w/v fructose solution and females were maintained on human blood. The insectary was maintained at 27°C (±1°C), 70–80% humidity with a 12h light/dark cycle.
Human blood for mosquito feeding was acquired from the NHS blood service. During feeding, blood was maintained at 37°C on a membrane-feeding device or in a parafilm-covered Petri dish warmed with a handwarmer. Mosquitoes were allowed to feed for 1h and non-engorged mosquitoes were removed within 24h. Mosquitoes were offered egg dishes for oviposition the night before each subsequent blood meal. Antibiotics (60U/ml penicillin, 60μg/ml streptomycin and 50μg/ml gentamicin) or an equal volume of water were supplemented in the sugar solution offered from emergence, in the blood meal and in the egg dish provided for oviposition. For polyoxin D feeding, 0.01M stock solution was prepared from powder in water and added to human blood at a final concentration of 100μM immediately before feeding. An equal volume of water was added to the control blood meal.
Double stranded RNA (dsRNA) was used for transient in vivo knock down of target genes by RNAi. The target region was amplified from total A. gambiae cDNA using primers flanked with the T7 RNA polymerase promoter sequence (sequences are listed in S5 Table). dsRNA was synthesised from the PCR product by overnight incubation at 37°C with T7 polymerase and dNTPs from the MEGAscript RNAi kit, according to the manufacturer’s instructions. dsRNA was purified using the Qiagen RNeasy kit, adjusted to a concentration of 6000 ng/μl, and stored in aliquots at -20°C. 69 nl of 6000 ng/μl dsRNA (totalling 414 ng) was injected into the thorax of CO2- anaesthetised 0–2 day old female mosquitoes using the Nanoject II (Drummond Scientific). dsRNA against a region of the lac operon (LACZ), not present in the A. gambiae genome, was injected as a control for the injection process.
Prior to dissection mosquitoes were ‘surface sterilized’ by immersion in 75% ethanol for 3–5 min and washed three times in phosphate buffered saline (PBS) to minimize environmental contamination from cuticle bacteria into dissected midgut samples. Midguts were removed under a dissecting microscope, frozen immediately on dry ice in pools of 20 (for RNA sequencing), 3–5 (for excretion experiment and APER1 experiments) or 8–10 (for all other experiments), and stored at -20°C until processing. For carcass dissections, the abdominal carcass and attached fat body tissue was dissected, ensuring that all other organs (ovaries, gut, malpighian tubules) were removed. Carcasses were frozen immediately on dry ice in pools of 3–5 (for APER1 experiments) or 8–10 (for polyoxin D experiments) then stored at -20°C until processing.
Frozen tissues were homogenized in TRIzol (Invitrogen) and chloroform using a Precellys24 tissue homogenizer with bead beating (Bertin). RNA was precipitated from the aqueous phase with isopropanol, washed twice in 70% ethanol and resuspended in water. For RNA sequencing experiments, samples underwent a further column purification using the Qiagen RNeasy kit. For qRT-PCR experiments, cDNA was synthesized from up to 500 ng RNA using the Takara reverse transcriptase kit, according to the manufacturer’s instructions.
Libraries for sequencing were prepared in accordance with the Illumina TruSeq RNA sample preparation v2 guide (Part # 15026495, rev.D, September 2012) for Illumina Paired-End Indexed Sequencing. PolyA mRNA first underwent two rounds of purification using Illumina poly-T oligo-attached magnetic beads. During the second elution, the polyA mRNA was fragmented and primed with random hexamers for cDNA synthesis. After first strand cDNA synthesis, the RNA template was removed and a replacement strand was synthesized to generate double stranded cDNA. Ends were then repaired, dA base added and Illumina indexing adapters ligated. cDNA fragments with adapters on both ends underwent 15 cycles of PCR. Libraries were validated with the Agilent 2100 bioanalyzer to check size distribution. Samples were quantified by qRT-PCR, the concentrations normalized and samples pooled according to biological replicate. Pools were loaded at 10 pM onto four lanes of an Illumina flowcell v3 and sequenced using the Illumina HiSeq 1500, 2 X 100 bp paired-end run. Sequences are deposited in the NCBI Sequence Read Archive under the BioProject ID PRJNA385903.
Quality control, filtering and alignment were conducted in the Galaxy platform [63–65]. Groomed FASTQ files underwent adapter clipping (ILLUMINACLIP with Truseq3 adapter sequences) and were then trimmed by sliding window, averaging a minimum Phred quality score of 20 over 4 bases (Trimmomatic tool version 0.32.1). Only reads with both mate pairs being longer than 20 bp were processed further. These were aligned by Bowtie2 (version 0.4) to a custom built index to filter out non-mRNA reads, composed of all sequences annotated as A. gambiae in the SILVA rRNA database (release 119) [66], plus all sequences annotated as A. gambiae tRNAs and mitochondrial rRNAs in the AgamP4.2 geneset in Vectorbase. The splice aware aligner Tophat2 v2.0.9 was used to align paired-end reads to the A. gambiae PEST genome AgamP4. The mean inner distance between mate pairs was set to -25 with a standard deviation of 60. Default settings were used for alignment with the following exceptions: the maximum number of mismatches allowed between a read and the reference sequence was 5 to allow for the highly polymorphic nature of the A. gambiae genome, and the minimum intron length was set to 30 bp. The accepted hits were filtered such that only reads that were uniquely mapped were accepted for downstream analysis.
Aligned reads were converted to gene count data using HTSeq, specifying the union mode [67]. The input gtf file was AgamP4.2 after removal of all features annotated as rRNA. Differential expression analysis was conducted with the DESeq2 package [68], using HTSeq count tables as input files. For each gene, the DESeq2 package fits a generalized linear model (GLM) with a negative binomial distribution. For pairwise comparisons at each time point the input parameters were “replicate” and “treatment” (i.e., plus/minus antibiotics), and “treatment” was removed in the reduced model. DESeq2 applies the Wald test to assess statistical significance followed by the Benjamini-Hochberg adjustment for multiple testing. Genes with adjusted p-values <0.1 were considered significantly differentially expressed.
Variance stabilizing transformation of count data was performed in DESeq2 prior to clustering. A median-transformed value of the four replicates was calculated for each condition and soft clustering performed in Mfuzz [69]. Soft clustering does not require a priori gene filtering, is noise robust and allows genes to be placed in more than one cluster, making it ideal for time-course data. The fuzzifier m was chosen with the mestimate function, and the optimal number of 12 clusters was selected based on when the minimum distance between cluster centroids (Dmin) declines at a reduced rate.
KEGG and GO term enrichment analysis for differentially expressed genes were performed in g:Profiler [70]. Genes were ordered by their fold change for input to the software and a Bonferroni adjustment was made for multiple testing.
qRT-PCR was used to quantify A. gambiae mRNA levels, as well as bacterial load by amplification of the 16S rRNA gene, employing primers that anneal to a region of the sequence that is common to all eubacteria or specific to bacterial families examined. Primer sequences are listed in S5 Table. In each case, the A. coluzzii ribosomal protein encoding gene S7 (AGAP010592) was used as an internal control of the quantity of input RNA. Expression ratios were calculated using primer efficiencies that were determined by amplification of serially diluted targets. qRT-PCR amplifications were performed in duplicate using the SYBR premix ex Taq kit (Takara) in a total volume of 10μl on a 7500 Fast Real Time PCR machine (Applied Biosystems).
For hematoxylin and eosin (H&E) staining, whole mosquito abdomens were dissected 24h post blood feeding, fixed overnight at 4°C in Duboscq’s-Brasil fixative (0.4% w/v picric acid, 53% ethanol, 27% formalin, 7% glacial acetic acid) and washed in 70% ethanol. Abdomens were then processed to paraffin wax and sections cut to 4μm onto Superfrost Plus slides (VWR). For Gram staining, immunostaining and calcofluor white staining, whole mosquito abdomens were dissected 24h post blood feeding, fixed overnight at 4°C in 4% formalin, washed in Aedes saline (150mM sodium chloride, 1.4mM calcium chloride, 2mM potassium chloride, 1.2mM sodium hydrogen carbonate, pH 7.2), and embedded and sectioned as described. For Gram staining, sections were dewaxed and stained according to the Gram/Twort protocol. Slides were observed under a Leica DMR microscope. For immunostaining, sections were stained with E. coli polyclonal antibody at 1:400 (bs-2351R Bioss Antibodies) and a goat anti-rabbit antibody bound to Alexa 647 (Thermofisher A21245) was used as secondary (1:1000). Slides were mounted in Prolong Gold antipode (Invitrogen) and observed under a Zeiss Widefield Axio Observer Microscope. For calcofluor white staining, sections were dewaxed and rinsed in distilled water before being stained for 2h in calcofluor white solution in the dark (Sigma 18909). Sections were observed under a Zeiss Widefield Axio Observer Microscope.
qRT-PCR data (including gene expression and 16S analyses) were analyzed by generalized linear mixed models (GLMMs) in R (version 3.1.2). GLMMs fit both fixed-effect parameters and random effects in a linear predictor via maximum likelihood. Mixed effect models were used to account for the use of multiple sample pools per condition within each independent replicate, avoiding issues of pseudoreplication. Statistical significance was assessed by an ANOVA test on a linear mixed effect regression model (lmer, in the lme4 package). Correlation was analyzed by a Spearman rank-order correlation test.
|
10.1371/journal.ppat.1004869 | Host Transcriptional Response to Influenza and Other Acute Respiratory Viral Infections – A Prospective Cohort Study | To better understand the systemic response to naturally acquired acute respiratory viral infections, we prospectively enrolled 1610 healthy adults in 2009 and 2010. Of these, 142 subjects were followed for detailed evaluation of acute viral respiratory illness. We examined peripheral blood gene expression at 7 timepoints: enrollment, 5 illness visits and the end of each year of the study. 133 completed all study visits and yielded technically adequate peripheral blood microarray gene expression data. Seventy-three (55%) had an influenza virus infection, 64 influenza A and 9 influenza B. The remaining subjects had a rhinovirus infection (N = 32), other viral infections (N = 4), or no viral agent identified (N = 24). The results, which were replicated between two seasons, showed a dramatic upregulation of interferon pathway and innate immunity genes. This persisted for 2-4 days. The data show a recovery phase at days 4 and 6 with differentially expressed transcripts implicated in cell proliferation and repair. By day 21 the gene expression pattern was indistinguishable from baseline (enrollment). Influenza virus infection induced a higher magnitude and longer duration of the shared expression signature of illness compared to the other viral infections. Using lineage and activation state-specific transcripts to produce cell composition scores, patterns of B and T lymphocyte depressions accompanied by a major activation of NK cells were detected in the acute phase of illness. The data also demonstrate multiple dynamic gene modules that are reorganized and strengthened following infection. Finally, we examined pre- and post-infection anti-influenza antibody titers defining novel gene expression correlates.
| Gene expression profiling of human blood cells might uncover the complex dynamics of host response to ARIs such as pandemic H1N1. However, only limited data are available on the system level response to naturally acquired infections. To understand the molecular bases and network orchestration of host responses, we prospectively enrolled 1610 healthy adults in the fall of 2009 and 2010, followed the subjects with influenza-like illness (N = 133) for 3 weeks, and examined changes in their peripheral blood gene expression. We discovered distinct phases of the host response spanning 6 days after infection, and identified genes that differentiate influenza from non-influenza virus infection. We then moved the focus from gene expression patterns to gene co-expression patterns. We detected gene modules that are related to core features of regulatory networks and found a substantial increase in the connectivity of the influenza responsive genes. Finally, we identified a molecular signature that correlated significantly with antibody response to pH1N1 virus. Taken together, our findings offer insights into the molecular mechanisms underlying host response to influenza virus infection, and provide a valuable foundation for investigation of the global coordinated responses to ARIs. Molecular correlates of the immune response suggest targets for intervention and improved vaccines.
| Influenza viruses are highly contagious respiratory pathogens that cause about three to five million cases of severe illness, and about 250 000 to 500 000 deaths worldwide each year [1]. In the US, influenza affects an estimated 5% to 20% of the population yearly [2], and is responsible for an average of 3.1 million hospitalized days, and 31.4 million outpatient visits. Direct medical costs are estimated to be at least $10.4 billion annually [3]. A new influenza virus appeared in Mexico and the United States in April 2009 and caused extensive outbreaks of influenza in the population. The virus was promptly identified as a swine-like influenza A (H1N1) virus and shown to be a triple reassortant virus containing genes from swine, human, and avian influenza A viruses [4]. Pandemic swine influenza (pH1N1) peaked in the United States in October 2009, with minimal activity during the subsequent winter period of influenza and reappeared during the winter of 2010–2011. Our recent studies showed that preexisting antibody to the seasonal A/H1N1 virus reduced pH1N1 influenza virus infection and illness in healthy young adults [5, 6].
Complex coordinated responses are triggered in the host following an acute respiratory viral infection. Many aspects of host-pathogen interactions after influenza infection have been studied [7–12]. Blood transcriptome profiling provides a ‘snap shot’ of the systematic host immune networks, as blood circulates throughout the body, carrying naive and educated immune cells, whose transcriptional activity can be influenced by environmental stimuli such as a respiratory virus illness [13]. Transcriptional signatures have been described in the context of ARIs caused by different etiological agents, including influenza, rhinovirus (HRV), and respiratory syncytial virus (RSV), as well as by influenza vaccination [14–23]. These studies have shown that blood gene expression signatures are distinctive for individuals with infection-induced ARI. ARI gene expression signatures show highly significant enrichment for transcripts encoding proteins involved in interferon signaling and pattern recognition induced innate immunity responses [14, 16].
Transcriptome analysis in influenza-infected mouse lungs has revealed distinct phases of the host response extending over at least a two month period after infection [20]. In previously reported studies in humans with ARI, transcriptional profiling was only performed on RNA samples collected either at a single timepoint of peak symptoms, or within the initial 2 to 3 days of hospitalization. The dynamic changes over the entire time course of naturally acquired infection and illness in humans are less clear. Experimentally induced influenza infection has been used to obtain information about changes in temporal gene expression [14, 21]. Huang et al made important observations about the differences in response between asymptomatic and symptomatic individuals. These studies, however, were limited by sample size and could not contrast other common respiratory virus agents with influenza. Menachery et al reported a contrasting gene signature between pH1N1 and coronavirus infected airway epithelial cells [22]. The genes they investigated were limited to interferon-stimulated genes. Studies to characterize the temporal dynamics of the systemic transcriptional response to ARI in humans are necessary to better understand the biology of infection, the host response and occurrence of disease. Furthermore, serum antibody responses to influenza virus infection have large inter-individual variation [5, 6]. Several influenza vaccine studies showed genes that play a role in antigen presentation and T cell recognition are associated with influenza vaccine-induced antibodies [19, 23–25]. Whether the same sets of genes contribute to the variation in antibody response to naturally acquired influenza infection is not known.
Approaches to uncover the modular organization and function of transcriptional systems have shown promise in facilitating functional interpretation and discovering biological networks. These models have been successfully applied in several biological contexts [26–28]. Weighted Gene Co-expression Network Analysis (WGCNA) group sets of genes with similar transcriptional patterns together to form a transcriptional module. Since the probability for multiple transcripts to follow a complex pattern of expression across all the samples by chance is low, such sets of genes should constitute coherent and biologically meaningful transcriptional units [29, 30]. Recently developed differential co-expression analysis goes beyond identification of differentially expressed genes (DEGs) or pathways to identify differential co-expression pattern [31–33]. Under the premise that pairwise correlations between gene expression levels result from regulatory relationships among the genes, major changes in co-expression patterns between two conditions may indicate dysfunctional regulatory systems in disease.
The clinical, virological and immunological results of our prospective study of ARI in a young adult population that included influenza and other known pathogenic viruses have been reported [5, 6]. Using genome wide transcript profiling we provide evidence in this report for three distinct phases of response among those persons with ARI: a) acute systemic activation of the innate response; b) recovery with extensive cell repair and proliferation; and 3) restoration of baseline gene expression patterns. These results provide new transcriptional correlates for the evolution of ARI. The results indicate a central role for interferon and innate immunity in the acute phase of the illness. The recovery phase has not been well characterized previously and suggests new avenues for understanding the restoration of biological system equilibrium after infection and illness.
1610 healthy adults were prospectively enrolled before the influenza seasons of 2009–10 and 2010–11. Of these, 142 (8.8%) who subsequently developed a moderate influenza-like illness were enrolled for follow up; none met the criteria for severe respiratory disease. Of the 142 enrolled ill subjects, 133 reported for all scheduled study visits and had technically adequate gene expression data (vide infra). Table 1 summarizes the infection and demographic data for these 133 subjects. Viral culture and RT-PCR for respiratory viruses indicated that 64 were infected with influenza A virus, and 9 were infected with influenza B virus. Infection with a rhinovirus, respiratory syncytial virus (RSVA/RSVB), coronavirus (OC43, 229E, NL63, HKU1), or enterovirus (Entero) was also detected in a number of the subjects with influenza-like symptoms. There were 24 individuals with an influenza-like illness for whom no virus was identified. The subjects were predominantly European-Americans (80.5%), consistent with the study area population.
We analyzed the global gene expression profiles of peripheral whole blood in the 133 adults with an ARI at up to seven time points before, during, and after the occurrence of illness (Fig 1A and 1B). Because the subjects were enrolled prospectively we had control samples taken from the same subject before occurrence of illness (baseline samples). A total of 890 microarray analyses were completed. Samples which failed QC were excluded from the analyses (N = 10), leaving 880 high quality arrays from which the subsequent analysis was conducted. Differential expression analyses for each day, compared to the baseline were then stratified by viral agent. We first analyzed the gene expression profiles in 49 subjects from whom an influenza virus infection was identified. The 24 subjects with influenza A virus infection in the 2009 cohort were used as a discovery group, and the consistency of differential expressed genes was assessed in another 21 influenza A virus infected subjects and 4 influenza B virus infected subjects in the 2010 cohort as a validation group. After performing significance testing with corrections for multiple tests, we detected highly significant expression differences in thousands of transcripts during the period of influenza illness (days 0, 2, 4, 6) in both discovery and validation groups. In contrast, once the subject had clinically recovered there were no significant expression differences detected (day 21, and spring samples).
Since blood is a complex tissue, changes in transcript abundance can be attributed to either transcriptional regulation or changes in the composition of leukocyte populations. To “deconvolute” these two phenomena, we computed a cell score derived from the expression profile of each sample using a composite of lymphocyte, neutrophil or monocyte specific transcripts. We found that lymphocyte lineagespecific transcripts were depressed in the acute phase of influenza virus infection, increased above baseline in the recovery phase, and then returned to baseline on day 21 (Fig 2A). An opposite change in neutrophil score was observed (Fig 2B). Expression levels of monocyte markers were increased in the acute phase and returned to baseline on day 6 (Fig 2C). These changes in established lineage markers of the broad cell populations probably track the changes in cell composition in the peripheral blood. The changes in lymphocyte and neutrophil proportions we predicted “in silico” are consistent with the changes described in experimental human challenges with influenza virus [34, 35]. Changes in lineage composition could explain part of the differential gene expression observed during the infection. We therefore recomputed the differential expression analysis using the lymphocyte and neutrophil scores as covariates in a series of contrasts focusing on days 0 to day 6 compared to baseline. Although the p-values were slightly increased, the rank ordering of genes showing highly specific differential expression was nearly identical (Tables 2–5). This indicates that while cell composition does affect estimates of total transcript abundance, the most important component of the differential expression arises from changes in transcript abundance within those populations. On a global scale, changes in the host transcriptomes were observed from the first day of illness through day 6 evaluations. A total of 4,706 differentially expressed genes (DEGs) (BH-corrected P values <0.05 in both cohorts) were identified over the course of 6 days of influenza virus illness (S1A Fig). 2119 transcripts, corresponding to 1421 genes, were responsive to the infectious stimulus on day 0 (day 1 or 2 of illness). The number of DEGs peaked at day 4. On day 6, only a small number (N = 46) of DEGs were newly detected (i.e. DEGs that first appeared on day 6 and were not detected at any time before). 738 out of the 1140 DEGs with |log2 Fold-Change| > 0.3 were first detected on day 0 (S1B Fig).
We plotted a heatmap of the 202 transcripts (S1 Table) showing the most significant pattern of differential expression compared to baseline (Fig 2D and 2E), and determined the transcript order by hierarchical clustering. These genes fall into two clusters: 1) genes that were regulated in the acute phase of influenza virus infection, and 2) genes that became differentially expressed at a later time-point (recovery phase). All individuals showed complete recovery to the baseline transcript profile by day 21 after onset of illness. In the acute phase, there was a very large increase in components of the interferon pathway and innate immunity (e.g. IFI44L, IFIT1, MX1, IFITM3, OAS2, IFI27 and IFIT3, see Table 2), as well as decreased expression of genes involved in translational elongation and protein biosynthesis (e.g. RPS4X, RPS18, RPS6, RPS8 and RPL5, see Table 3). This was most intense on the first day of illness and continued for 2–4 days. This phase was followed by a characteristic recovery phase in which there was a transition to genes involved in antigen binding and antibody secretion (IGJ, LOC652694, IGLL1 and MZB1, see Table 4) and genes regulating cell morphogenesis (STRADB, DPYSL5, EPB42, LST1 and MAP1S, see Table 5). Inter-individual variations in the magnitude of transcriptional response at each phase were observed (S2 Fig), and greater variations were seen at the times when the transcriptional responses were strong. The expression profiles for individuals infected with Influenza A and B virus were indistinguishable. Likewise individuals infected with both influenza and rhinovirus were not different from those infected with influenza virus alone. No statistically significant differences in expression of any transcripts were identified that marked the mixed infection group.
Although there were significant differences in gene expression between the non-influenza virus infection group (e.g. HRV, RSV, coronavirus and enterovirus) and the influenza group (Fig 2D and 2E, S3 Fig), the patterns of the three-phase transcriptional responses were nearly identical, and the differential expression was largely explained by differences in the magnitude of effect. This indicates that the host response to acute respiratory viral infection, despite the distinctive biology of these diverse viruses, is largely conserved. By performing differential expression analysis comparing influenza virus and rhinovirus infection group (Criteria for DEGs were BH-corrected P values < 0.0001), we found that comparing to rhinovirus, influenza virus infection and illnesses induced a larger magnitude and longer duration of activation of interferon signaling pathway, and a greater depression in translation and protein biosynthesis (S4A and S4B Fig, S2 Table). Some of the DEGs encode kinase or kinase inhibitor (e.g. MAPK1, PAK2, CDKN1A and CDKN1B, S4C Fig). Several protein phosphatase encoding genes were also differentially expressed, such as PPM1M, PPP2R4, PPP3CA, etc. However, the magnitudes of differential expression in these genes were small and there were only two transcripts showed |log2 fold-change| >1.5: IFI27 was consistently upregulated in the influenza virus group on days 0–6 but not upregulated in the rhinovirus or other infection groups (Fig 2F); in addition, PI3 was consistently downregulated in the influenza-infected individuals but not in the other groups (Fig 2G). The fold changes of IFI27 and PI3 transcript levels comparing the first day of illness with baseline were also measured by RT-qPCR, and were consistent with the microarray result (S5 Fig).
We also examined the pattern of gene expression in the group of individuals reporting symptoms of acute viral respiratory illness but who were negative in PCR or culture tests for the tested viral pathogens. These individuals had gene expression profiles nearly identical to those observed in the known virus groups, including the acute and recovery phases of gene expression. The transcript levels of IFI27 and PI3 in these subjects were more similar to the non-influenza infection cases. Within this group there was some variability in the magnitude of the transcriptional responses, including large variation in IFI27 and PI3, perhaps suggesting either additional etiologic heterogeneity or incomplete sensitivity of the culture and PCR assays (S3 Fig).
Three subjects had a systemic expression profile consistent with activation of interferon signaling on the day of enrollment (Fig 2D). One of these individuals had persistent elevations of these transcripts throughout the study. One subject reporting illness symptoms did not have the signature of acute systemic response on day 0 but had the typical signature by day 4. The remaining 4 subjects showed ‘off-cycle’ activation profiles consistent with additional intercurrent infections with or without severe symptoms.
We used previously published lineage and activation state marker sets to compute cell type scores for each sample (S3 Table). Lineage specific transcripts lists were obtained and then mapped on to the Illumina array probe identifications. The expression levels of the lineage-specific markers on each day were computed by taking the average of all the influenza-infected individuals. The changes in the expression levels are likely influenced by both the proportion of the cells in the peripheral blood as well as the transcriptional state of those cells. In the acute phase of infection there was a slight depression of lineage markers for NK cells, followed by up-regulation of the marker gene GPR56 on day 2 that became stable by day 21 and thereafter (Fig 3A). The same procedure was used to compute a score for the activation status. Notably, NK cells showed very dramatic increases in activation state during the acute infection but then the activation signature rapidly resolves in the convalescent phase as the infection subsided (Fig 3B). The changes observed in the 2009 cohort were replicated in the 2010 cohort. Our findings of the dramatic activation of NK cells during the early phase are consistent with the observations from influenza-infected mouse lungs [20]. These data show an intense activation of NK cells during the acute phase of infection.
After the gene expression status of the peripheral blood cells of influenza-infected individuals were profiled over the time course of illness, we then searched for molecular, cellular and biological processes that best correspond to the host gene expression responses. For this, we analyzed the functional annotation of differentially expressed genes using DAVID gene ontology. Analyses of significant differentially expressed genes on day 0 and day 2 (BH-corrected P value <0.05 in both discovery and validation cohorts) revealed that the upregulated genes were mostly enriched in defense response, response to other organism, response to virus, innate immune response, positive regulation of cytokine production, and positive regulation of tumor necrosis factor production (Fig 4A), while the downregulated genes were involved in translational elongation, translation, cellular protein metabolic process, rRNA binding, and cellular macromolecule biosynthetic process (Fig 4B). The functional interpretation of differentially expressed genes in the recovery phase (day 4 and 6) is much less clear: there are a range of protein metabolic process and regulation of ubiquitin-protein ligase activity functions represented in the upregulated genes (Fig 4A), and actin cytoskeleton organization functions associated with the downregulated genes (Fig 4B), but how these changes may be integrated will require further investigation. We observed a higher level of gene enrichment in the interferon signaling and the tumor necrosis factor production pathway in the acute phase of influenza virus infection compared to rhinovirus infection.
We next examined how biological pathways might be altered during the course of influenza with respect to baseline, by performing separate content analyses of DEGs on day 0 or day 6. We found that upregulated genes in acute influenza virus infection were enriched for canonical pathways specific to interferon signaling (S6A Fig), role of pattern recognition receptors of virus, TREM1 signaling, antigen presentation pathway, activation of IRF by cytosolic pattern recognition receptors. Cellular processes such as dendritic cell maturation and crosstalk between dendritic cells and NK cells were also enriched, indicating the activation of these pathways in acute influenza. On the other hand, the downregulated genes were significantly enriched for pathways related to gene translation and cell proliferation, such as EIF2 signaling, regulation of eIF4 and p70S6K signaling and mTOR signaling. A group of pathways distinct from those seen in the acute phase were enriched in DEGs in the recovery phase (S6B Fig). A large number of upregulated genes in the recovery phase were functioning in the protein ubiquitination pathway. Stress response pathways (e.g. hypoxia signaling in the cardiovascular system and NRF2-mediated oxidative stress response) were also enriched. In addition, significant enrichment in multiple growth factor signaling pathways (e.g. GM-CSF signaling, HGF signaling and PDGF signaling) and cell cycle regulation (e.g. mitotic roles of polo-like kinase) were observed.
Highly co-expressed genes usually share common regulatory mechanisms or participate in the same biological process. To reveal distinct patterns on how host genes are co-expressed in different stages of influenza virus infection, the WGCNA method was applied to the gene expression profiles of samples from the first or second day of illness (day 0), day 2, day 4 and day 6, thereby the network organization is approached through inference of variable gene co-expression patterns and dynamic pathway activity rather than a fixed predefined gene annotations.
We examined the differentially expressed transcripts (BH-corrected P values <0.05 and |log2 Fold-Change| > 0.3 in both cohorts) and detected 6 co-expression modules on day 0 (designated Day 0_1 to Day 0_6) (Table 6). Module Day 0_1 contains genes that are highly upregulated (Fig 5A), and many of them are interferon signaling pathway genes (e.g. IFI6, IFI44L, IFIT1, IFIT3, IRF7 and STAT1). Genes in module Day 0_2 are enriched for translational elongation, the majority of the transcripts in this module are downregulated on day 0 but the transcript levels then gradually increased and became above baseline on day 4 (Fig 5B), suggesting the host translation system was attenuated in the acute phase of influenza but then recovered during the later phase.
We applied the same method to DEGs identified on day 2, day 4 and day 6 (Table 6). Almost all the GO terms over-represented in day 2 modules were observed on day 0, with the exception of the GO terms enriched in module Day 2_5: Hemoglobin complex / oxygen transport. We found many hemoglobin genes in this module, and their transcript levels were decreased on day 4, including HBD, HBE1, HBG1, HBG2, HBA1, etc (Fig 5C). DEGs on day 4 were grouped into 9 modules and these modules became more diverse in GO term enrichment. A small module Day 4_4 contains some of the top upregulated genes (Fig 5D), such as IGLL1, IGJ, LOC652694, MZB1, which are involved in antibody secretion.
To elucidate transcriptional regulatory networks within each module, we performed transcriptional factor (TF) enrichment analysis using Pscan. The results for module Day 0_1, Day 0_2, Day 2_5 and Day 4_4 are shown in the networks (Fig 5E–5H) with the red nodes representing enriched TFs. Module Day 0_1 is regulated by the TFs in the interferon signaling pathway. Module Day 0_2 involves genes targeted by ETS transcription factor family. Meanwhile, the binding motifs of ARNT and MYCN are enriched in both Module Day 2_5 and Day 4_4. Many genes in these networks contain binding motifs of multiple TFs, implying the TFs are highly coordinated in regulating downstream targets.
Differential co-expression patterns, wherein the level of co-expression of gene groups differs between illness and pre-illness, can arise from an influenza infection-related change in the regulatory mechanism governing that set of genes. We performed differential co-expression analysis on all the genes that are differentially expressed on day 0 in influenza-infected individuals. We found that the correlations between gene expression levels of all the gene pairs are higher on day 0 compared to baseline (Fig 6A). Particularly, the modules 1 in the lower left corner have significant correlation differences between day 0 and baseline, suggesting that the module genes are in the same regulatory network. For example (Fig 6B), the correlation coefficient between the expression levels of OAS2 and RNASEL gene are nearly 0 at baseline, yet they became highly correlated (r = 0.72) on day 0. We computed the correlation coefficients among all gene pairs and plotted the results as a function of the magnitude of correlation or connectivity (Fig 6C). The co-expression patterns among these genes peaked on the first day of illness, gradually weakened thereafter, and become indistinguishable with baseline by day 21. The gene expression correlation also increased, though to a lesser extent, in HRV infection group (S7 Fig).
We measured serum antibody to the pH1N1 viruses in all subjects at enrollment and after surveillance for illness [5, 6], and thus were able to record the magnitude of antibody response (delta H1N1 titers). We wished to identify the genes whose transcripts levels are correlated with antibody response. The Day 0 transcript levels of 2119 DEGs were tested for their correlation with the delta antibody titers among 58 ill subjects with H1N1 infection. We found that a total of 229 genes showed evidence of significant correlation between gene expression on the first day of illness and the antibody response (Fig 7, S4 Table). Of these, 168 showed evidence of positive correlation and 61 of negative correlation. LILRB4 (Leukocyte immunoglobulin-like receptor subfamily B member 4) showed most significant positive correlation, and a member of the forkhead transcription factors, FOXO3 exhibited most significant negative correlation. Content analysis revealed that immune response (GO:0006955) were most enriched in the genes showed positive correlation with antibody response (e.g. OAS1, CD14, APOBEC3G, IFITM3 and LILRB4). B-cell proliferation genes (e.g. CD40, SASH3, CDKN1A and TICAM1) were strongly correlated with high antibody response. Genes that showed negative correlation (e.g. FOXO3, DAPK2, SGK1, and TP53INP1) were enriched for apoptosis and programmed cell death pathways (GO:0012501).
This prospective study of 2009 pandemic influenza A virus infections and illnesses in healthy adults in a university community detected clear gene expression patterns correlating with moderate influenza. Transcriptional profiles for ill subjects were examined at 7 time points, including baseline, the first day of illness and up to 21 days after as well as after the influenza season. The aim of this study was to use an unbiased genome-wide approach to identify genes whose expression is regulated by occurrence of an ARI and networks that are activated in the host response to infectious stimuli during different stages of an acute respiratory viral illness.
The results show a gene expression signature that strongly corresponds to influenza virus infection. Components of interferon pathway and innate immunity (IFI44L, IFITM3, MX1, IRF7, OAS2, STAT2, etc.) are significantly upregulated in the acute phase of infection, while the expression levels of genes involved in translational elongation and protein biosynthesis are decreased. Other researchers have identified host gene expression patterns that are associated with viral infection in human airway epithelial cells [17] and bronchial epithelial cells [28]. DEGs in type I interferon or STAT1 signaling were similar to those found in our current study. In particular, IFI27, an antiviral molecule that regulates interferon-mediated apoptosis, was the most highly up-regulated gene in these studies [14,17]. Our study identified IFI27 and PI3 as the most differentially expressed genes comparing influenza virus and rhinovirus infections. IFI27 up-regulation was observed in hospitalized infants with RSV bronchiolitis [36], yet it was not seen in the three healthy adults infected with RSV in our study. PI3 (peptidase inhibitor 3) encodes elafin, a potent neutrophil elastase inhibitor, localized to the injury sites in the lung [37]. PI3 protein was shown to possess antimicrobial and anti-inflammatory activities [38]. Its mechanism of action is, however, poorly understood and down-regulation of PI3 has been reported previously in patients with acute respiratory distress syndrome (ARDS) [39], but not in studies of viral infections.
A key finding in this study was a recovery phase that involves differential expression of a set of genes distinct from those observed in the acute phase of infection. Molecular characterization of a recovery phase has not been previously reported and the functions in the immune response of most of the differentially expressed genes are much less clear. Protein ubiquitination pathways and protein metabolic process are associated with the genes upregulated in the recovery phase. Symptomatic influenza cases exhibited extensive regulation in multiple growth factor signaling and cell proliferation pathways during the recovery phase. How these results may be integrated will require further investigation.
Comparison of influenza and rhinovirus illnesses indicated that the intensity of the increase in activation of the interferon and innate immunity pathways is less in rhinovirus infections. In addition to IFI27 and PI3 that have the largest expression difference between influenza and HRV, CDKN1A and CDKN1C, which are essential genes involved in cell cycle control, appeared to be differentially expressed specifically in influenza virus infection. It has been reported that influenza virus infection induced cell cycle arrest in G1/S phase [40], and the transcriptional reprogramming of cell cycle correlated with the severity of influenza illness [41]. The differences in host transcriptional response to influenza virus and rhinovirus infection might be explained by the fact that influenza virus replicates in the nucleus of host cells while HRV replicates in cytoplasm [42], or explained by the distinct viral mechanisms in histone modification [22]. Individuals who did not have an identified pathogen associated with illness had conserved systemic expression signatures that were indistinguishable from the influenza and rhinovirus groups, with a large variation in the intensity of transcriptional response. This suggests that they were actually infected with one of the respiratory viruses for which we tested but which was not detected, or that they had an infection with another infectious agent that induces a similar transcriptional response. This question will be the subject of future investigation using highly sensitive next generation sequencing methods.
The finding of ‘out of cycle’ individuals suggests that there are many subclinical infections or other subclinical disorders in healthy adults. Given that acute viral infection stimulates gene pathways known to be involved in adult onset autoimmune disorders raises the possibility that the number and intensity of infections may alter risk in genetically susceptible individuals. One individual exhibited activation at all time points. This may represent a systemic disorder, a possibility that is now being examined in that subject. This suggests the possibility that gene expression profiles may be used in the detection of such disorders as an adjunct to standard immunological testing.
We used weighted gene co-expression network analysis (WGCNA) to cluster the DEGs detected on day 0 –day 6 into 26 modules. This module construction strategy takes advantage of the biological variability inherent in the prospective cohort study in order to uncover the modular organization and function of transcriptional systems. The time-course transcriptional profiles make it possible to study the transcriptional regulation of these gene co-expression networks during different phases of influenza illness. While GO terms enriched in the acute phase modules are “response to virus” and “translational elongation”, recovery phase modules are over-represented in a new set of GO terms, such as “endoplasmic reticulum part”, and “programmed cell death”. We also found the hemoglobin genes in module Day 2_5 (e.g. HBD, HBE1, HBG1) are downregulated in response to influenza virus infection, however, whether this is due to true transcriptional regulation, or a decrease in their percentage in comparison to the white cells will need further investigation. Furthermore, the TF regulatory networks in 4 modules were uncovered, which provides better insights to the underlying mechanisms of host response to ARIs and will facilitate drug and vaccine development.
Our study went beyond gene co-expression and investigated the differential co-expression patterns in influenza virus and rhinovirus infections. The idea behind this is that the identification of changes in gene co-expression patterns between illness and baseline samples could provide information about infection-affected regulatory networks. Our result demonstrates that the gene expression correlations are enhanced on a global scale in the response to ARI; a small module containing 273 transcripts has the largest increase in network connectivity strength. This suggests qualitative change in the gene network upon an infectious stimulus.
We know there are several cell-types in whole blood sample and the proportion of these cell-types varies across samples, so it is possible that the co-expression modules were driven by variation in markers for various cell-types. However, the differences in gene expression correlation between baseline and the first day of illness are so large that it cannot be fully explained by the variation in the expression values of the cell lineage markers, which does not change much between baseline and illness. Thus, the changes in regulatory mechanisms are the major contributor to the differential co-expression patterns.
The pandemic influenza A/H1N1 virus emerged in April 2009 and was the dominant influenza virus circulating in humans in our study periods. By measuring the H1N1 antibody titers on the same individual before and after the influenza season, we were able to record the magnitude of antibody response (delta H1N1 titers) and account for individual variation in a way that would not have been possible otherwise. We identified 229 genes whose transcriptional levels were correlated with the antibody response. Although the sample size we had for the correlation analysis is relatively small, over 1/3 of the genes identified in this study have previously been shown to be correlated with antibody response to influenza vaccination [24]. B cell proliferation genes, which predict influenza vaccine-induced antibody response [23], were also correlated with the antibody response to naturally occurring influenza infection. These findings provide more insight into the molecular mechanisms of antibody production and secretion, and may also contribute to influenza vaccine development.
Several limitations of this study are noteworthy. First, we studied two cohorts of healthy young adults. Those subjects who subsequently developed influenza-like illness had moderate symptoms [5]. Children, the elderly, were not included and, fortunately, none of the research subjects developed severe illnesses. Second, this study did not allow analysis of the subjects who had influenza infection without symptoms. Based on seroconversion rates, 38% of the subjects were probably infected with influenza A H1N1 but did not have symptoms sufficient to trigger a follow up study visit. Third, all the subjects with influenza-illness enrolled in year 2009–2010 were infected with influenza A H1N1. And in year 2010–2011, only 9 were infected with influenza B and 3 were infected with influenza A H3N2, all the others were infected with H1N1 infection. Thus the sample size was not sufficient for comparing host transcriptional response to influenza A H1N1, H3N2 and influenza B. Fourth, the transcriptional responses to infection of cells residing in the secondary immune tissues, like lymph nodes or spleen, might be different from that of peripheral blood. Future research may investigate the correlations of gene expression between cells residing in different tissues. Finally, while antibody titers have been used to assess humoral immune responses, it is clear that they do not capture the complexity of the host response to ARIs. Additional studies would be necessary to establish the causal relationship between the genes identified and the antibody response, and whether they also regulate cytokine or chemokine levels.
Despite these limitations, the findings in this study demonstrate the power of serial measurements of gene expression, within the context of a prospective clinical trial, to identify candidate genetic mechanisms that determine responses to infection. We have genotyped all these research subjects and have begun analyzing the impact of common genetic variation on the gene expression patterns. Because we have made repeated measurements on the same individual over time, we should be able to account for the effect of person in a way that would not have been possible using cross-sectional methodologies. The dynamic nature of the measurements should also allow the identification of genetic effects that are either enhanced by or only evident after the strong perturbation of acute infection.
The study was conducted at Texas A&M University, College Station, TX. The protocol and informed consent were approved by the Baylor College of Medicine and Texas A&M University institutional review boards before the study began. Healthy adults age 18 to 49 at the college and in the community were invited to enroll to be followed for acute respiratory illness (ARI) through two consecutive influenza seasons 2009–2010 and 2010–2011. All adult subjects provided written informed consent.
After subjects provided consent, a medical history was taken to ensure good health, and baseline specimens were obtained. Surveillance for influenza began during the September 2009 enrollment period because pH1N1 as a cause of influenza was identified in the population during enrollment. Subjects were given thermometers and instructions to call and report for evaluation within 48 hours of onset for any ARI (Fig 1A). Except for the Thanksgiving holiday period and 4 weeks of the Christmas holiday period, a coordinator and physician enrolled persons presenting within 48 hours of onset with a new ARI with fever or that caused them to miss school, work, or social activities. Specimens were obtained and medical care was provided, including the antiviral zanamivir if indicated. Enrolled persons were seen 2, 4, and 6 days later for repeat evaluation, specimen collections, and medical care and 21 days later for collection of convalescent specimens. These subjects are those included in the present report. Surveillance for influenza was terminated after 5.5 months; all subjects were asked to return for specimen collection and to provide a medical and ARI history. The study was repeated 2010–2011 with surveillance for influenza limited to January to April as community surveillance did not detect influenza before the Christmas break.
A study physician obtained an oral temperature, completed a symptom survey, and performed a respiratory system examination at each illness visit. All cases were classified as clinically moderate using standard criteria.
Serum specimens obtained at enrollment, acute and convalescent visits for illnesses, and the terminal visit were tested simultaneously using hemagglutination-inhibition (HAI) antibody tests following previously described methods. Virus antigens were a locally obtained pH1N1 virus (A/Baylor/09) and the most recently prevalent seasonal A/H1N1 virus (A/Brisbane/59/07), A/H3N2 virus (A/Perth/16/09), and B virus (B/Brisbane/60/08).
A combined 8-mL nasal wash and throat swab specimen was collected at each illness visit. Specimens from the day 0 and 2 visits were tested for all respiratory viruses in tissue cultures. All specimens were also tested by reverse-transcriptase polymerase chain reaction (RT-PCR) for respiratory viruses including influenza A, pH1N1 influenza, influenza B, picornavirus/rhinovirus, respiratory syncytial virus, human metapneumovirus, parainfluenza viruses, coronaviruses, and adenoviruses.
We collected peripheral whole blood samples (2.5 mL) in PAXgene RNA stabilization tubes (QIAGEN Inc., Valencia, CA, U.S.A.) at each visit of those enrolled for illness and froze the samples at—80°C until RNA purification to minimize gene expression changes induced by handling and processing. RNA purification was performed using the PAXgene Blood RNA system (QIAGEN Inc., Valencia, CA) according to manufacturer’s instructions. Quality control of RNA samples was performed using spectrophotometry (NanoDrop-1000 Spectrophotometer, Thermo Fisher Scientific, Waltham, MA, U.S.A.) and microfluidic electrophoresis (Experion Automated Electrophoresis System, Bio-Rad Laboratories, Hercules, CA).
All statistical analyses on the gene expression data were performed in R Statistical Software [44], version 2.14.1. Differential gene expression analyses with cell composition covariates contrasting the individual day-specific data with the baseline sample obtained at the time of enrollment were performed using function for linear model fitting in the limma R package [45]. The significance of differences in gene expression was tested using a Bayes moderated t-test [46]. Correction for multiple testing was addressed by controlling the false discovery rate (FDR) using the Benjamini and Hochberg (B.H.) method. A transcript probe was considered significantly differentially expressed if the B.H. corrected P value was < 0.05. The heatmap function in R Statistical Software was used to generate a heatmap of mean-centered normalized expression values.
Gene expression profiles were investigated for correlation with cell composition in the whole blood. Cell lineage and activation state markers were used as described in Abbas et al [47]. A full list of marker genes we used to compute the cell scores is provided in S1 Table. Cell lineage scores for all individuals were obtained by taking the first Principal Component (PC1) of average-normalized expression values for each of the lineage-specific gene sets. [When this method was used to compute the cell scores for the 121 subjects whose whole blood transcriptional profile and Complete Blood Count (CBC) are publicly available at Gene Expression Omnibus (GEO accession: GSE30119), the resulting expression-based lymphocyte and neutrophil scores showed a high correlation (r2 = 0.64 and 0.65, respectively) with actual measurements of percent lymphocyte and neutrophil in the CBC (S8 Fig).] Neutrophil and lymphocyte scores were then introduced as quantitative covariates in the linear models of the differential expression analyses to account for the differences in cellular composition between individuals.
TaqMan (Applied Biosystems, Foster City, CA) quantitative real-time reverse transcriptase polymerase chain reaction (RT-PCR) was performed on baseline and Day 0 paired RNA samples from 18 randomly selected subjects with influenza infection only and 11 subjects with rhinovirus infection only. cDNA was first synthesized from approximately 2 ug of total RNA in a 20-ul reaction volume using the High Capacity RNA to cDNA kit (Applied Biosystems). TaqMan probes, available as “Assay on Demand”, were used in the analyses of the expression levels of 2 target genes, IFI27 (Hs01086373_g1) and PI3 (Hs00160066_m1), as well as endogenous control gene GAPDH (Hs03929097_g1). Quantitative RT-PCR was performed on 1ul of cDNA in triplicates with the CFX96 Touch Real-Time PCR Detection System (Bio-Rad, Hercules, CA). The fold increase in mRNA expression was determined using the ΔΔCT method with the baseline sample of each pair as calibrators.
Gene lists were analyzed using Ingenuity Pathway Analysis (IPA) software and DAVID Ontology (http://www.david.abcc.ncifcrf.gov) to identify significantly enriched pathways. Expressed genes represented in the full dataset were used as the background. The Biological Process, Molecular Function and Cellular Component subsets of the Gene Ontology (GO) were used for enrichment analysis. DAVID Ontology uses t-test to derive P values and applies the Benjamini-Hochberg method to correct for multiple testing. IPA uses a right-tailed Fisher’s exact test to derive P values for identifying significantly overrepresented pathways. A smaller P value indicates the overrepresentation of a pathway or a GO term by the DEGs is less likely due to random chance.
To identify groups of host transcripts that showed coordinated regulation in response to acute illness, we applied the weighted gene co-expression network analysis (WGCNA) [29]. The WGCNA method constructs networks or modules consisting of groups of genes that are highly correlated across a set of samples. Briefly, the absolute value of the Pearson correlation coefficient is calculated for all pairwise comparisons of gene-expression values. The Pearson correlation matrix is then weighted and transformed into an adjacency matrix. WGCNA uses the topological overlap matrix based dissimilarity measure as input of hierarchical clustering. A dendrogram (cluster tree) of the network is then obtained from hierarchical clustering. Finally, modules are defined by cutting branches off the dendrogram. WGCNA was performed using the WGCNA package provided in R software.
JASPAR is an open-access database (http://jaspar.cgb.ki.se) derived exclusively from sets of nucleotide sequences experimentally demonstrated to bind transcription factors. Transcription factor binding specificity is represented by position-specific scoring matrices (PSSM) in JASPAR. Employing the profiles available in JASPAR, Pscan (http://www.beaconlab.it/pscan) scans a set of sequences (promoters positions -450 to +50 with respect to the transcription start site) from co-regulated or co-expressed genes and identifies the enriched transcription factor binding site motifs by comparing the average matching value of the matrix on the sequences analyzed and that on the whole promoter set (same set of regions with respect to the transcription start site) of the same organism. Z-test was used to derive P values and Bonferroni method was applied to correct for multiple testing.
All the transcripts showed differential expression on Day 0 were used in the differential co-expression analysis (DiffCoExpr) [31]. To detect changes in correlations between gene pairs within module and also between pairs of modules during ARIs, DiffCoExpr, an untargeted approach in which gene modules are not pre-defined, was carried out for each day contrasted with the baseline. Briefly, an adjacency matrix as the Spearman correlation coefficients for all pairs of genes was built for each day and baseline. Then, the correlation changes on each day compared to baseline were quantified by the difference between signed squared correlation coefficients. Finally, the Topological Overlap based dissimilarity matrix was derived from the adjacency change matrix, and was used as input for gene clustering and module detection.
|
10.1371/journal.pbio.1001207 | Electrostatically Biased Binding of Kinesin to Microtubules | The minimum motor domain of kinesin-1 is a single head. Recent evidence suggests that such minimal motor domains generate force by a biased binding mechanism, in which they preferentially select binding sites on the microtubule that lie ahead in the progress direction of the motor. A specific molecular mechanism for biased binding has, however, so far been lacking. Here we use atomistic Brownian dynamics simulations combined with experimental mutagenesis to show that incoming kinesin heads undergo electrostatically guided diffusion-to-capture by microtubules, and that this produces directionally biased binding. Kinesin-1 heads are initially rotated by the electrostatic field so that their tubulin-binding sites face inwards, and then steered towards a plus-endwards binding site. In tethered kinesin dimers, this bias is amplified. A 3-residue sequence (RAK) in kinesin helix alpha-6 is predicted to be important for electrostatic guidance. Real-world mutagenesis of this sequence powerfully influences kinesin-driven microtubule sliding, with one mutant producing a 5-fold acceleration over wild type. We conclude that electrostatic interactions play an important role in the kinesin stepping mechanism, by biasing the diffusional association of kinesin with microtubules.
| Animal and plant cells contain a molecular-scale “railway” network, in which the tracks, called microtubules, radiate out from the cell centre and locomotive proteins, called kinesins, haul their molecular cargoes along the microtubule tracks. This railway system transports many different cargoes to where they are needed, so it is crucial for the cell's organization and function. Breakdowns in this transport system can cause diseases like Alzheimer's, and drugs that temporarily halt transport make powerful anti-cancer agents. Precisely how kinesin motor proteins move along their microtubule tracks is an important question in biology. We know that some kinesins have twin “heads” that alternately bind to and step along microtubules in a coordinated walking action. But more usually, kinesins have only one head. How single-headed kinesins produce force and movement is poorly understood. In this study, we address this question and show that electrical attraction between single kinesin heads and microtubules is a critical factor deciding the direction of movement: each time the head approaches a microtubule, it slides forwards by the electrical attraction between the engine and the track.
| Kinesins form a large family of ATP dependent microtubule-based motor proteins. At least 14 sub-families have been identified [1]–[3], the members of which play a wide variety of roles in intracellular transport, including vesicle and organelle transport, cytoskeletal reorganization, and chromosome segregation [4]. Underpinning these diverse activities is a coupling of ATP turnover, microtubule bind-release cycles, and unidirectional mechanical motion. Several features of the mechanisms by which kinesins generate force and movement are known, but many uncertainties remain. Kinesin-1, the best studied kinesin, has twin heads and moves towards microtubule plus ends using a head-over-head walking action that can do work against loads of up to ∼7 pN [5],[6]. Importantly however, the minimal motor domain of kinesin-1 is a single head [7]. Teams of single kinesin-1 heads can drive directional microtubule sliding, with each head in the team contributing intermittent impulses of force and motion. Less is known about this mechanism, by which individual kinesin heads generate directional force.
Broadly, two different classes of model have been proposed for the mechanical cycle by which kinesin heads generate force and movement—biased binding models and unbiased binding models. In biased binding models, the motor domain diffuses back and forth on a spring-like tether, using thermal energy from the bath to stretch out the tether, locking on to the track at a moment when the spring is stretched out in the progress direction, and then maintaining its grip on the track whilst the spring relaxes. Biased binding models like this (Figure 1, left) are sometimes referred to as thermal ratchets [8]. The classic example of this type of model is the Huxley 1957 [9] model for the myosin crossbridge. In biased binding models, most of the ground gained is due to directionally biased diffusion-to-capture. The directionally biased capture event is envisaged to involve or trigger a directional conformational change and one or more coupled chemical steps, but the conformational change is negligibly small compared to the stepping distance. By contrast, models with unbiased binding (Figure 1, right) envisage that the probability of binding of kinesin heads to microtubules is the same in both directions and that directional stepping is entirely due to one or more directional conformational changes that occur after the motor has engaged with its binding site. Current controversies over the role of neck linker docking in the kinesin cycle relate to this same dichotomy. Neck linker docking is a conformational change that is clearly important in the kinesin mechanism [10], but whether neck linker docking can do appreciable work remains uncertain. The results of molecular dynamics simulations argue that substantial work could be done [11]. On the other hand, measurements of the free energy difference between the docked and undocked neck linker indicate ∼5 pN nm [12], suggesting that neck linker docking could not do the work necessary to account for kinesin's ability to step ∼8 nm against a ∼7 pN load.
Since conformational changes, including neck linker docking, undoubtedly do occur once the kinesin head is attached to the microtubule [13], the key problem is to find out whether a biased binding mechanism contributes appreciably to the kinesin mechanical cycle or whether instead binding is unbiased and the generation of directional force is entirely due to one or more conformational changes that follow microtubule binding.
There is clear evidence that tethered single kinesin heads can develop impulses of directional force and displacement. These step-displacements have been estimated using single molecule optical trapping to be 3–4 nm, and attributed to biased binding [14],[15]. Many theoretical models [16],[17] posit that biased binding occurs and that it is driven by one or more directional sawtooth binding potentials. As yet, however, a specific molecular mechanism is lacking. This is the problem we address in the current work.
It is known for a number of non-motor systems that electrostatic interactions can effectively maneuver associating proteins into a suitable binding configuration, a phenomenon known as electrostatic steering [18],[19]. Formation of the final tightly bound complex from the encounter complex may require internal structural rearrangements as well as more local effects, including dehydration of the binding interface. Electrostatics is known to play a role in the binding of kinesin to microtubules, with roles established for the negatively charged E-hook of tubulin, and for the positively charged K-loop of kinesin, in both the Kif1a (kinesin-3) [20] and kinesin-1 [21] mechanisms, and for charged residues and ionic strength in general [22]. In the present work we have sought to test whether long-range electrostatic guidance might govern not only the rate, but also the approach trajectory, of kinesin-microtubule encounters.
To approach this question, we performed electrostatic calculations and atomistic Brownian dynamics simulations in parallel with in vitro motility assays of electrostatically engineered mutant kinesin motors. Our results demonstrate a strong tendency for long-range electrostatic guidance to enhance kinesin-tubulin association and encounter complex formation. Expanded simulations of kinesin dimers on short sections of microtubule indicate that conserved electrostatic interactions not only enhance association but also enable kinesin heads to bind preferentially to sites lying ahead in the progress direction. We further find that the tethering of two heads in a dimer reduces the search space for binding sites on the microtubule lattice, effectively enhancing directional bias and providing a mechanism to track single microtubule protofilaments. Simulations with a range of subfamily representatives and selected charge neutralizing mutations suggest that different kinesin subfamilies have tailored their electrostatic properties to modulate association rates and the directional bias of the association reaction along the microtubule. We conclude that electrostatic interactions play an important role in kinesin stepping by guiding the biased diffusional association of kinesin with microtubules.
Electrostatic calculations of available motor domain crystal structures spanning 11 kinesin sub-families reveal considerable diversity in patterns of surface charge distribution (Figure 2A and Movie S1). Nevertheless, all structures analyzed possess an invariant region of positive potential (blue) in the nucleotide-binding site and over the back face, particularly loop8, loop7, loop12, and alpha5 (including residues R284, K281, R278, K141, K237, R161, and K166). Also apparent are regions of consistent negative potential (red) located near the loop preceding α3 (residues D144 and E170), giving rise to a common underlying asymmetric charge distribution in the kinesin family (Figure 2B and Movie S2).
The conserved positive potential at the nucleotide-binding site reflects the role of this region in coordinating the negatively charged phosphates of ATP. The other conserved region of positive potential spreads across a considerable part of the microtubule-binding surface of the head (Figure 2B), reflecting the established role for this surface in binding to the negatively charged surface of the microtubule. Alanine scanning mutagenesis [22] and limited proteolysis [23] support this view and more recent high-resolution cryoelectron microscopy studies [24],[25] confirm that following microtubule binding this region becomes buried in the microtubule-kinesin interface. Our analysis identifies several further regions of more subtle conservation of positive charge, such as those in the neighborhood of α3 and α6 (including residues R326, K328, D177, E178, and D123). Such regions are not identified with conventional sequence analysis methods [26].
Further comparison and clustering of the calculated electrostatic potentials identified groupings with similar charge distributions (Figure 2C). These results indicated that electrostatic properties are more similar within known sub-families than between sub-families. We selected two representative motor domain structures from four of the largest clusters (representing kinesin-1, 3, 5, and 13 sub-families) as the inputs for our Brownian dynamics simulations. Brownian dynamics simulations were employed to characterize the association process, determine association rates, and investigate the role of long-range electrostatic forces in the association mechanism. Comparison of simulations with and without charges on the motor-domain shows that electrostatic interactions enhance the association rates for all sub-families studied (Figure 3 and Movie S3). As the different motor domains have a range of net charges (+5 to −3), it is unlikely that rate-enhancement arises from nonspecific attraction due to monopole interactions; rather, enhancement of association rates is directly related to the non-uniform charge distribution on kinesin and tubulin. Inspection of BD trajectories clearly shows the steering of the conserved motor domain's positive surface patch toward the negatively charged surface of tubulin (Figure 3B,C), leading to a preferred binding site between tubulin subunits.
Examining successfully associated trajectories indicates that the preferred motor domain approach path lies along a directional trajectory leading from the inter-subunit interface (the alpha-beta junction) toward a single preferred association site located at the beta-alpha intra-heterodimer interface (Figure 3B). The Brownian motion during the approach to binding becomes biased, generating a plus end-directed shearing movement during diffusion-to-capture. Along the preferred approach path, the motor domain's positive patch is predominantly oriented toward the tubulin surface (Figure 2C). This indicates that the motor domain rotates into a specific orientation at an early stage (at a center-to-center distance of ∼60 Å, corresponding to a maximal surface-to-surface separation of ∼15 Å), so that during approach, rotation is constrained such that subsequent motion consists largely of steered translations along the approach trajectory (see also Movie S3). Studies by others on the barnase-barstar system have also characterized significant electrostatic interactions at similar surface-to-surface separation distances [27],[28]. Even at two Debye lengths (∼15 Å at 150 mM ionic strength), interactions will be reduced by about 1/7 compared to contact, which can still yield significant steering effects for highly charged proteins [28]. Simulations with kinesin and tubulin show that at higher ionic strength, electrostatic steering is partially quenched (Figure S4).
BD mimics the physical process of diffusional association under the influence of electrostatic interactions. Our simulations indicate that the distinct charge distributions of different kinesin sub-families lead to a range of sub-family-specific association rates (Figure 3A). Kinesin-3 is predicted to have the highest relative association rate (2.6×108 M−1 s−1) followed by kinesin-1 (8.27×107 M−1 s−1), kinesin-13 (2.75×107 M−1 s−1), and kinesin-5 (1.2×107 M−1 s−1).
Different structures from the same subfamily were found to have very similar association rates reflecting their common charge distributions. Simulations performed under varying salt concentrations showed a similar sub-family trend resulting in decreased association rates at higher ionic strength for all sub-families (see Figure S1).
Simulations of monomeric kinesin-1 motor domains interacting with a microtubule fragment consisting of 7 protofilaments, each with 5 tubulin heterodimer subunits (see Figure 4A), indicated that freely diffusing kinesin-1 motor domains have an intrinsic preference for sites at the plus-end of microtubules (Figure 4B). A similar trend was found for other subfamily members, including minus-end directed kinesin-14 (see Figure S2). These simulations indicate that single motor domains have an equal propensity for each tubulin dimer internal to the microtubule lattice. Intriguingly, simulations performed on charge neutralized microtubule lattices have an overall reduced association rate to all sites and do not display a noticeable plus-end preference (see Figure S2). Together these results indicate that electrostatic features present at the plus-end tip of microtubules favor kinesin association. Minoura and colleagues [29] recently showed that charged nanoparticles diffuse one-dimensionally on microtubules and that the amplitude of the diffusional excursions reduces exponentially as the charge increases. It is possible that the provision of extra charge density at microtubule ends represents a general mechanism for targeting the plus-ends of microtubules.
Additional simulations were performed on kinesin-1 and kinesin-14 (Ncd) dimers with one freely diffusing head tethered by a spring to a microtubule-bound partner head. Results from these simulations indicate dramatically different binding preferences (Figure 4D–F). Kinesin-1 tethered heads clearly favor the forward plus-end binding site, whilst Ncd tethered heads favor the rearward minus-site. This result indicates an intrinsic or underlying dimer-enhanced directional bias that exists independent of neck-linker [30] or stalk [31],[32] docking and undocking. The majority of binding events occur on the protofilament to which the partner head is attached. Tethering appears to enhance biased binding by reducing the search space for binding sites (Figure 4C–E). Note that surprisingly the same electrostatic interactions and tether geometry that favor the plus-end-biased binding of dimeric kinesin-1 favor the minus-end-biased binding of dimeric kinesin-14. Control simulations without charges returned no apparent directional preference (Figure 4F). Hence, different kinesin subfamilies appear to have tailored their electrostatic properties to not only enhance and modulate association rates but also to influence directionality.
The core result from our simulation is that conserved electrostatic features on the kinesin head facilitate its electrostatic guidance during the diffusional approach to microtubule binding, leading to a consistent plus-end-directed diffusional motion of the kinesin head in the moments before binding.
The simulations allow us to examine the roles of particular residues (on both tubulin and kinesin) in forming the field responsible for this directionally biased diffusion-to-capture. We analyzed the effects of charge-neutralizing mutations on the rate constants of association using BD simulations and the recently developed transient complex approach (see Materials and Methods). By definition the transient complex includes the final bound conformations from successful BD trajectories. We use the ensemble of transient complex configurations to calculate the average electrostatic interaction energy (ΔGelec) and the electrostatic interaction energy compared to wild-type (ΔΔGelec) (Table 1 and Figure 5).
The specific predictions made by our simulations about the effects of mutations allow us to test the reliability of our simulations by mutating these residues in the real-world proteins. Computationally, each surface exposed charged residue on kinesin-1 was mutated to alanine and the effect on predicted relative association rates monitored (Table 1). Figure 5A displays these results in relation to the crystallographic structure of kinesin-1 (PDB code: 1bg2). Note the prominent effect of mutations on the rear face of the motor domain. In contrast, mutation of residues on the front face of the motor domain was found to have little impact. Rear positions with a significant influence include those residues contributing to the conserved positive potential patch (i.e., residues R284, K281, R278, K141, K237, R161, and K166, all of which are ranked highly in Table 1). Additional positions in α6 (such as K313, R421, E309, and E311) and β1c (K44) along with the loop before α3 (D144 and E170) were also found to have a significant influence. Also shown in Figure 5 are the published results of experimental alanine scanning mutagenesis by Woehlke and colleagues [22]. Note the excellent correspondence to the results of the Woehlke study, which measured the effects of alanine substitutions on the ATPase and motor activity of kinesin, with the sites highlighted in the current study as influencing association rates and electrostatically guided diffusion-to-capture. Both our calculations and these earlier experiments indicate that substitution of positive residues on the microtubule binding face of kinesin decreases, whilst substitution of negative residues increases association rates.
Mutations that decrease the association rate do so by neutralizing the conserved electrostatic features essential for electrostatic steering. We obtained the largest decreases in the binding rate (of ∼2.3×107 M−1 s−1) for sites including R284A, K281A, and other contributors to the invariant rear positive potential patch. Association rates could be enhanced (up to a value 7.15×107 M−1 s−1 for N263R and E170A/D144A) by substituting residues from subfamilies that have an enhanced association rate. A number of control mutants (including D177A/E178A) were also examined and found to yield similar rates to the wild-type complex (8.19×107 M−1 s−1). Note that D177A and E178A were selected as controls as these residues have a similar proximity to the putative tubulin-binding site as E170A and D144A but were not highlighted by electrostatic conservation analysis.
In tandem with our simulations, we performed in vitro experiments to test the effects of electrostatic mutations on kinesin function. Our computational analysis (Figure 5 and Table 1) identified charged residues predicted to have a profound effect on the on-rate of kinesin-1 to microtubules. Simulations also indicate that the distinct charge distribution of different kinesin sub-families can lead to a range of sub-family specific association rates (Figure 3A). To further probe the origin of these differences we focused on a three-residue segment at the C-terminus of helix α6. This region was observed to have a distinct sub-family-specific charge distribution in different kinesin sub-families (with a consensus sequence of RAK in subfamilies-1, -3, -5, and -13; SVN in kinesin-14; MTQ in kinesin-6 and RAR in kinesin-4). This region was previously shown to be essential for ATPase and motility [33] and was highlighted by both our electrostatic analysis and in another coarse-grained modeling study (Zheng et al., in prep). We made a series of experimental point mutants in NKin, a fast kinesin-1 from Neurospora Crassa, and assayed the effects of the mutations on microtubule sliding velocity, microtubule-activated ATPase, and tubulin-activated ATPase. A single-headed NKin construct was used, so as to mimic the conditions of the simulation. Tables 2 and S1 and Figure 6 summarize the results.
All the mutants retained microtubule-activated and tubulin-activated ATPase activity. Both R321A (AAK) and the potentially more disruptive charge-reversal R321D (DAK) mutation are predicted by our simulations to have little effect, and the experiments confirm this. K323R (RAR) and K323A (RAA) are predicted to accelerate binding somewhat, and indeed increased microtubule sliding velocity 2-fold, compared to wild-type single head NKin. Replacing the RAK sequence with AAA resulted in a ∼3-fold velocity increase and R321K (KAK) produced a ∼5-fold increase in the velocity of kinesin-driven sliding microtubules. R321K does not affect the net charge on the molecule but does profoundly enhance the association of the motor to its microtubule track. Using purified pig brain tubulin (both as unpolymerized heterodimers and as microtubules, polymerized in the presence of Mg-GTP and taxol-stabilised), we measured the rates of microtubule-activated and tubulin-activated ATP hydrolysis and ADP release for wild-type and for RAK mutants (see Table S1). All constructs, wild type and mutant, were activated by free tubulin heterodimers, but to a lesser extent than by microtubules. For microtubule activation, the KAK mutant, which is 5-fold faster in motility assays, has a slightly reduced Vmax in solution compared to wild type (∼54 s−1 compared to 97 s−1) and a ∼5-fold weaker apparent affinity for microtubules (Km ∼28 µM compared to 6 µM), The AAK mutant, which has wild type velocity in motility assays, also has a weaker apparent affinity for microtubules (Km ∼19 µM) but shows an increased Vmax (222 s−1). These results support a conventional model in which kinesin binds to microtubules in two steps, at first forming a “weak” state that attaches to microtubules but is not activated by them, and then shifting into a “strong” state that does show microtubule-stimulated product release [34]. Our simulations deal with the binding reaction that populates the initial, weakly bound state. We expect mutations that stabilize electrostatic interactions to accelerate the formation of this initial, weak binding state, and potentially also to accelerate exit from the strong state back into the weak state (Figure 7).
These dual effects over-populate the weak binding state, and this can account for the properties of our mutants in ATPase assays and motility assays. Microtubule sliding assays are accelerated because internal system drag, due to slowly detaching heads, is reduced. Microtubule-activated ATPase, averaged across the entire kinesin population, is little affected. We hypothesise that this is because the influence of faster initial formation of the weak-binding state is balanced by depopulation of the strong binding states (Figure 7).
Relating to Figure 7, we note that in order to explore electrostatic effects, we have treated the kinesin head as a rigid-body and focused exclusively on the diffusion-to-capture process. In future work we will aim to explore the role of electrostatics in the weak-to-strong conformational change and in subsequent steps in the mechanism.
In summary, we find using atomistic Brownian dynamics simulations and in vitro mutational analysis that conserved electrostatic interactions enhance association and enable kinesin heads to preferentially bind tubulin heterodimers lying ahead in the progress direction. Furthermore, we find that the tethering of two heads in a dimer reduces the search space for binding sites on the microtubule lattice and further biases binding to a single microtubule protofilament. Simulations with different subfamily representatives and selected charge neutralizing mutations suggest that different kinesin subfamilies have tailored their electrostatic properties to modulate both their association rates and their directional bias along the microtubule. Taniguchi and colleagues [35] recently suggested that directional bias in walking kinesin dimers is predominantly entropic. It will be interesting to test this concept in relation to our proposal that directional electrostatically biased diffusional association is an intrinsic feature of the force-generating mechanism of kinesin minimal motor domains.
Available kinesin crystal structures were obtained from the RCSB protein data bank and processed with the Bio3D package [36]. Processing involved initial extraction of motor domain coordinates corresponding to residues 9 to 325 in conventional kinesin-1. Subsequent alignment and superposition steps were as described in Grant et al. [26]. Missing regions of the various structures underwent standard molecular mechanics modeling and refinement protocols with the AMBER9 package [37]. Microtubule models were constructed by fitting multiple tubulin dimers to the 8 Å electron density map of Downing and coworkers [38].
Electrostatic calculations were performed with APBS (version 0.10.1) [39], using AMBER charges and radii at 310 K. Due to the high charge densities of the systems under consideration, the full, nonlinear Poisson–Boltzmann (PB) equation was solved in a multi-level fashion. Atomic charges were mapped to grid points via cubic B-spline discretization (chgm: spl2). The dielectric boundary between solute (with a dielectric constant of 4) and solvent (with a dielectric constant of 74) was specified as the van der Waals surface (srfm: mol and srad: 0).
Electrostatic potentials for available kinesin motor domain structures were analyzed with SurfaceDiver (version 1.0) [40]. Surface Diver employs spherical harmonic decomposition and a finite set of rotation-invariant descriptors to compare surface electrostatic properties. Based on these descriptors, molecules can be compared and clustered according to their electrostatic features without prior structural alignment. Operational parameters included a zero atom inflation radius (irad 0) and a maximal decomposition radius of 40 Å (rmax 40). Decomposition was performed on a total of 40 spherical surfaces (nsph 40) with a spherical harmonic decomposition order of 64 (spho 64). Complete-linkage hierarchical cluster analysis was performed with R and the Bio3D package.
The BrownDye simulation package (version 1.0) [41] was employed for sub-family and mutant Brownian dynamics (BD) simulations. All atom models were used for both kinesin and tubulin. Because of uncertainties over the conformational dynamics of the neck linker, simulations used the head only (corresponding to residues 9–325 of kinesin-1, as for the electrostatics calculations above). Effective charges were used to reproduce pre-computed electrostatic potentials (see above). The influence of these potentials on the diffusional motion of both kinesin and tubulin was determined from the standard Ermak and McCammon algorithm [42]. Association rates were computed at 150 mM ionic strength with a modified version of the Luty, McCammon, and Zhou algorithm [43]. An adaptive time step with a minimum value of 1.0 ps was employed. Trajectories were propagated until the transient complex was obtained (see below) or until a center-to-center distance c (beyond b) was reached. Upon reaching c a pretabulated solution to the diffusion equation was used to determine whether the molecules would “escape” to infinity or return to some location with a center-to-center distance b. To obtain adequate statistics, 200,000–500,000 trajectories were simulated for each kinesin-tubulin pair.
The current version of BrownDye treats proteins as rigid bodies and does not take into account short-range interactions (van der Waals and hydrogen bonds). However, these interactions become important for short distances. Hence, the transition from encounter or transient complex to the subsequent bound states is beyond the realm of the current BD simulations and requires the application of more detailed models with explicit treatment of flexibility and short-range interactions.
As previously introduced, binding partners can be considered to pass through a transient intermediate state (A*B), in which the two proteins have near native separations and orientations. From this transient complex (also referred to as the encounter complex), non-diffusional rearrangements lead to the tightly bound native complex (AB).(1)Hence, the overall binding rate (ka) is given by:(2)The current BD simulations probe the diffusion-controlled rate (kD) for reaching the transient complex. In the transient complex, kinesin and tubulin must satisfy particular translational and rotational constraints. Defining these constraints provided a robust set of criteria for assessing successfully associated BD trajectories.
Initial atomic models for each kinesin-tubulin complex were built by fitting different kinesin crystal structures to a kinesin-tubulin complex obtained from a 9 Å CryoEM model of Moores and coworkers [25]. These complexes underwent molecular mechanics refinement with the AMBER9 package and corresponding all-atom potential function ff99SB (see Text S1). The resulting lowest energy models were used as the starting configurations for probing the bound state and the transition to the unbound state via the transient complex method of Zhou and coworkers [44].
The algorithm for identifying the transient complex boundary has been described in detail elsewhere. Briefly, to sample bound and unbound configurations, both kinesin and tubulin were treated as rigid. The kinesin motor domain was systematically translated and rotated with respect to the larger, fixed-in-space tubulin dimer. Steric clashes were monitored along with the number of inter subunit contacts (defined as heavy atoms having interfacial contacts less than 5 Å). For clash-free configurations, the number contacts (Nc) together with interface separation (r) and rotation angle (χ) were recorded (see Figure S3). The value of Nc (denoted as Nc*) at the onset of a sharp increase in χ was used to define the transient complex. These configurations (with Nc = Nc*) effectively separate the bound state, with numerous short-range interactions (high Nc) but restricted translational and rotational freedom (low r and χ), from the unbound state, with at most a small number of interactions (low Nc) but expanded configurational freedom (large r and χ).
Measuring the effects of mutations on the rate constants of association is a powerful tool to decipher the mechanism of association. Mutated residues were given a modeled conformation based on the most probable rotameric state and subsequent side-chain energy minimization with the AMBER9 package. BD simulations and the transient complex approach were used to examine the effect of a mutation on the association rates and binding affinities. As in previous studies, 100 configurations were randomly selected from the transient complex ensemble to calculate the average electrostatic interaction energy (ΔGelec) and the electrostatic interaction energy compared to wild-type (ΔΔGelec):(3)(4)where the two terms on the right side of equation 4 denote ΔGelec after and before the mutation, respectively. For each transient complex configuration, ΔGelec was calculated as described in equation 3. These results were then averaged to yield ΔGelec*. See Alsallaq et al. [44] for further details.
Experiments used a 6xHistidine-tagged single-head NKin (6xHis-NKin343) as a starting construct, in which point mutations were created using PCR mutagenesis. Successful clones were verified by restriction site digestion and sequencing (Cogenics). The Histidine-tagged proteins were expressed in BL21/DE3 E. coli cells and purified using HisTrap Ni columns (GE Healthcare) using an AKTA Purifier system. Microtubule and tubulin-activated kinesin ATPase activities were measured using an enzyme-linked fluorescence assay [45], in a buffer (50 mM Pipes pH 6.9, 0.2 mM MgCl2, and 0.1 mM EGTA, 0.1 mg/ml BSA), at 25°C. For experiments involving microtubules Taxol was added to this buffer to a final concentration of 20 µM. Kd and Vmax were determined by fitting the data to a hyperbola using Prism 4 for Macintosh. Motility assays were performed following the method described by Kaseda et al. [46]. Nitrocellulose-treated coverslips (0.1% nitrocellulose in isoamyl-acetate) were coated in penta-His antibody (Qiagen cat. No. 34660, diluted 1∶10 in PBS), incubated in a moisture chamber for 1 h, and then extensively washed with 1 mg/ml BSA in PBS to remove any unbound antibody. Histidine-tagged kinesin at 0.3–3 µM in assay buffer (50 mM Pipes pH 6.9, 0.2 mM MgCl2, 0.1 mM EGTA, 5 mM DTT, 20 µM Taxol, 0.2 mg/ml Casein, 1 mM ATP) was then flowed into the chamber and allowed to bind to the surface for 10 min. Unbound kinesin was washed away using assay buffer, taxol-stabilised microtubules introduced and allowed to bind for 10 min. Unbound microtubules were washed off with assay buffer containing the oxygen scavenger system [47] at 25°C. Control coverslips lacking antibody did not recruit microtubules from the overlying solution. Microtubule motility was recorded by video-enhanced DIC microscopy and quantified using the freeware RETRAC software (http://mechanochemistry.org/software). Motility assays were made in the same buffer conditions as the ATPase assays with the addition of 1 mM DTT and 0.1% casein.
|
10.1371/journal.pntd.0003139 | Sensitivity and Specificity of Multiple Kato-Katz Thick Smears and a Circulating Cathodic Antigen Test for Schistosoma mansoni Diagnosis Pre- and Post-repeated-Praziquantel Treatment | Two Kato-Katz thick smears (Kato-Katzs) from a single stool are currently recommended for diagnosing Schistosoma mansoni infections to map areas for intervention. This ‘gold standard’ has low sensitivity at low infection intensities. The urine point-of-care circulating cathodic antigen test (POC-CCA) is potentially more sensitive but how accurately they detect S. mansoni after repeated praziquantel treatments, their suitability for measuring drug efficacy and their correlation with egg counts remain to be fully understood. We compared the accuracies of one to six Kato-Katzs and one POC-CCA for the diagnosis of S. mansoni in primary-school children who have received zero to ten praziquantel treatments. We determined the impact each diagnostic approach may have on monitoring and evaluation (M&E) and drug-efficacy findings.
In a high S. mansoni endemic area of Uganda, three days of consecutive stool samples were collected from primary school-aged children (six - 12 years) at five time-points in year one: baseline, one-week-post-, four-weeks-post-, six-months-post-, and six-months-one-week-post-praziquantel and three time-points in years two and three: pre-, one-week-post- and four-weeks-post-praziquantel-treatment/retreatment (n = 1065). Two Kato-Katzs were performed on each stool. In parallel, one urine sample was collected and a single POC-CCA evaluated per child at each time-point in year one (n = 367). At baseline, diagnosis by two Kato-Katzs (sensitivity = 98.6%) or one POC-CCA (sensitivity = 91.7%, specificity = 75.0%) accurately predicted S. mansoni infections. However, one year later, a minimum of three Kato-Katzs, and two years later, five Kato-Katzs were required for accurate diagnosis (sensitivity >90%) and drug-efficacy evaluation. The POC-CCA was as sensitive as six Kato-Katzs four-weeks-post and six-months-post-treatment, if trace readings were classified as positive.
Six Kato-Katzs (two/stool from three stools) and/or one POC-CCA are required for M&E or drug-efficacy studies. Although unable to measure egg reduction rates, one POC-CCA appears to be more sensitive than six Kato-Katzs at four-weeks-post-praziquantel (drug efficacy) and six-months-post-praziquantel (M&E).
| Schistosomiasis is a parasitic disease infecting over 200 million people. It remains a major public health concern despite treatment of over 120 million people in sub-Saharan Africa alone. Accurate diagnostic methods are essential for monitoring drug efficacy and long-term control program success. The World Health Organization recommends two Kato-Katz thick smears (Kato-Katzs) from a single stool for Schistosoma mansoni diagnosis to map prevalence and areas for control interventions. Although highly specific, Kato-Katzs are thought to be insensitive at low egg counts. The recently refined urine point-of-care circulating cathodic antigen test (POC-CCA) has been proposed as a diagnostic alternative for mapping areas for interventions, and potentially for assessing drug efficacy. Over three years we assessed the accuracy of six Kato-Katzs and a single POC-CCA in detecting infections in Ugandan primary-school children at 11 time points with repeated praziquantel treatments. Our results demonstrate that two Kato-Katzs accurately detect S. mansoni infection pre-treatment, but at least three days of two Kato-Katzs per stool or one POC-CCA are required for annual monitoring and treatment evaluation and/or drug-efficacy studies. One POC-CCA may be more sensitive in measuring S. mansoni prevalence than six Kato-Katzs, but its accuracies for rigorous intensity measures are still to be proven.
| Schistosomiasis remains a major public health concern despite praziquantel reaching over 30 million people in endemic areas in 2013 [1]. Goals to eliminate schistosomiasis by 2020 have been articulated by the World Health Organization's (WHO) ‘Roadmap for Neglected Tropical Disease (NTD) Implementation’ [2], and the London Declaration of the NTD Coalition [3]. Accurate diagnostic techniques, recently highlighted by Gomes and colleagues [4], are essential for monitoring and evaluation (M&E) of mass drug administration (MDA) programs at all stages [5]–[9], and particularly when considering elimination [2], [10], [11] and/or drug-resistance pharmacovigilance [12], [13].
The WHO recommends two Kato-Katz thick smears (Kato-Katzs) from a single stool [14] for Schistosoma mansoni diagnosis to determine prevalence to map areas for control interventions [15]. Kato-Katzs have assumed 100% specificity, but large inter- and intra-specimen variation [16]–[18] and low sensitivity for the detection of low intensity infections have been reported [19]–[21]. In Brazil, where M&E programs use only one Kato-Katz, S. mansoni prevalence has been significantly underestimated in low intensity regions [22], [23]. This may be associated with overestimated cure rates (CRs) and praziquantel efficacy [24], potentially missing drug resistance. Conversely, using one Kato-Katz can overestimate infection intensities [22]. These differences in sensitivity for prevalence versus infection intensity highlight complex interactions with egg reduction rates (ERRs) as true intensities decrease. Inter-day variation of excreted egg numbers post-treatment remains poorly understood. Two Kato-Katzs are commonly used for annual M&E of program impact and/or drug-efficacy studies, without being rigorously tested in these scenarios. Detailed analyses of the sensitivity and specificity of single and multiple Kato-Katzs following praziquantel treatment is urgently required.
It is not possible to directly measure S. mansoni adult worm numbers, due to their location in the mesenteric system, with eggs counted from Kato-Katzs used as a proxy for infection intensity. Immunodiagnostics for adult worm circulating cathodic antigens (CCA) or circulating anodic antigens (CAA) detect current infections and are potentially more sensitive for diagnosis of cases in low transmission areas [25]–[27]. Recent developments using innovative immunomagnetic separation with several target CCAs [28] and novel monoclonal antibody diagnostics for serum CCA [27] show high sensitivity even in low endemic areas. Serum CAA tests are more accurate than urine CAA tests [29], with advances made on a, not yet commercially available, point-of-care-CAA (POC-CAA) [30]. In contrast, urine CCA tests were more accurate than serum CCA tests [29]. Being easier to collect and more socially acceptable than stool or blood, urine POC-CCAs were developed for rapid, non-invasive diagnostics and proposed as alternatives to Kato-Katzs for S. mansoni prevalence mapping [31]–[37]. One urine POC-CCA is as sensitive as two Kato-Katzs for S. mansoni diagnosis for prevalence mapping [35] or two years post-treatment [36] and has now been used to map S. mansoni prevalence in low intensity areas in Uganda [37]. Major POC-CCA limitations are their ability to only semi-quantitatively measure infection intensity [32], [34], [38], inaccuracy for S. haematobium infection detection [39], [40], and not measuring soil-transmitted helminth (STH) infections. Intensity of infection measures are vital for control program M&E and drug-efficacy evaluation. Therefore knowledge on the ability of POC-CCAs to detect S. mansoni intensity reductions is essential [34]–[36], [41].
We compared the accuracy of the currently available POC-CCA and one to six Kato-Katzs (two smears per day from three consecutive stool samples) for S. mansoni diagnosis, in primary-school children in Mayuge District, Uganda, at baseline and after up to ten praziquantel treatments per child over three years (STARD Checklist S1 and Figure 1). We evaluated the epidemiological implications of diagnostic methods for control program M&E and praziquantel-efficacy studies. We inform on the number of Kato-Katzs required to accurately detect S. mansoni infections pre- and post-praziquantel treatment and whether POC-CCAs are suitable alternatives. Our detailed longitudinal design enabled novel investigations into individuals' recent versus total praziquantel treatments, aiding biological understanding of differences between Kato-Katz and POC-CCA results post-praziquantel treatment. We predicted that the accuracy of M&E and drug-efficacy findings are limited by Kato-Katz sensitivity at low infection intensities post-treatment. We also predicted that a single POC-CCA would have a higher sensitivity than multiple Kato-Katzs, and be more informative for prevalence monitoring as control programs progress.
Approvals were granted by the Uganda National Council of Science and Technology (Memorandum of Understanding: sections 1.4, 1.5, 1.6) and the Imperial College Research Ethics Committee (EC NO: 03.36. R&D No: 03/SB/033E). Verbal assent was given by every child before inclusion into this study and at school committee meetings comprising of parents, teachers, and community leaders before the onset of the study. Written consent for the children to participate in the study was attained from each head teacher. Participation was voluntary and children could withdraw or be withdrawn from the study at any time without obligation. Children were treated with 40 mg/kg praziquantel and 400 mg albendazole (active against STH infections) as detailed below.
Samples were collected, between 2004 to 2006, from primary-school children, in a high S. mansoni-endemic area, in Mayuge district, Uganda from three schools on the shores of Lake Victoria: Bugoto Lake View, Bwondha, and Musubi Church of God. Children at Musubi were, to the authors' knowledge, praziquantel-naïve. Children at Bugoto and Bwondha had received 40 mg/kg praziquantel one year previously in 2003 [42]. Inclusion criteria were to have lived in the area since birth and to attend the schools sampled.
In 2004, samples were collected at five time-points: baseline, one-week-post-, four-weeks-post-, six-months-post- and six-months-one-week-post-praziquantel treatment (Figure 1). In 2005 and 2006, samples were collected pre-, one-week-post-, and four-weeks-post-praziquantel re/treatment. On the third day of sampling, at baseline, six-months, one-year, and two-years all children were treated with 40 mg/kg praziquantel and 400 mg albendazole (active against STH infections). At one-week post-treatment, children with infections of >100 S. mansoni eggs per gram of stool (EPG) were retreated with 40 mg/kg praziquantel. At all other time-points all children with positive diagnoses for S. mansoni or STHs were retreated with 40 mg/kg praziquantel and 400 mg albendazole respectively.
Cohort and sample collection are described elsewhere [43]. In brief, 110 children from Bugoto, 110 from Bwondha and 68 from Musubi were recruited in 2004 with an equal sex ratio, aged six to 12 years, without prior knowledge of infection status and/or symptoms of S. mansoni infection. In addition, at one- and two-years, 30 praziquantel-naïve six year old children were recruited at each school and followed up with the original cohorts at the time points described above. This enabled monitoring of the impact of MDA on untreated children entering the school system, assessing diagnostic accuracies for Kato-Katzs and POC-CCA, in praziquantel-naïve and praziquantel-exposed children, as control programs progress.
Diagnostic accuracy increases with the number of Kato-Katzs, however, in Brazilian low intensity regions, the additional benefit of more than six Kato-Katzs from repeated stools was negligible [21], supporting our six Kato-Katzs ‘gold standard’. Stool samples, marked with unique child IDs, were collected on three consecutive days, between 10:00 and 12:00 hours. Two 41.7 mg Kato-Katzs were prepared per stool and read onsite using a compound microscope with natural light source, by highly trained personnel from the Ugandan Vector Control Division, Ministry of Health. S. mansoni, hookworm, Ascaris lumbricoides, and Trichuris trichiura egg counts were recorded. Five percent of slides were reread after the study for S. mansoni, A. lumbricoides, and T. trichiura egg counts for quality control, but no significant differences were observed. One urine sample per child was collected between 10:00 and 12:00 hours on the first day. In the first year, at all five time-points, POC-CCAs (European Veterinary Laboratory, The Netherlands) were performed, according to the producer's protocols, by the first author, blind of other test results. Microhematuria was tested for using Hemastix (Bayer, United Kingdom).
SPSS version 19 (SPSS, Inc., Chicago, IL, United States of America) was used for all statistical analyses. The double entered data were not normally distributed and could not be normalized by transformation, therefore non-parametric tests were used. Individuals without the full six Kato-Katzs were excluded from the study (Figure 1). Arithmetic mean infection intensities were categorized as by the WHO (S. mansoni: light = 1–99 EPG, moderate = 100–399 EPG and high ≥400 EPG; A. lumbricoides: light = 1–4999 EPG; Hookworm: light = 1–1,999 EPG; T. trichiura: light = 1–999 EPG) [15]. Exact confidence intervals (CIs) were calculated for prevalence measures and standard errors for EPGs.
There were 1065 samples with six Kato-Katzs and 367 samples with six Kato-Katzs and a POC-CCA result (Figure 1). Baseline S. mansoni prevalence (χ2 = 0.38, d.f. = 2, p = 0.83) and EPG intensity (Kruskal-Wallis: H = 3.416, d.f. = 2, p = 0.18) were not significantly different between schools and all statistics were performed on the combined data. Baseline prevalence in the main (six Kato-Katzs) dataset was 94.8% with an arithmetic mean infection intensity of 249.8 EPG, similar to the POC-CCA dataset prevalence (94.7%) and intensity (259.0 EPG) (Table 1). Hookworm prevalence was 51.0%, whilst A. lumbricoides and T. trichiura infections were low at 1.0% and 9.4%, respectively (Table 1). S. mansoni prevalence at baseline, as recorded by POC-CCA-t+ and POC-CCA-t− was 88.2% and 78.9%, respectively (Table 1).
Observed cure (measured at four-weeks-post-praziquantel treatment) and reinfection (measured at six-months) rates depended on sampling method and effort (Figures 4 and 5). Two Kato-Katzs underestimated S. mansoni reinfection whilst overestimating CRs (Figure 4A) (two Kato-Katzs CR = 81.5%; six Kato-Katzs CR = 70.4%). Cure rates determined with POC-CCA-t− were 47.8% and 26.1% for POC-CCA-t+. One-week-post-recent-praziquantel (from data at both one-week and six-months-one-week), prevalence was significantly lower when measured by POC-CCAs than by six Kato-Katzs (Figure 4B) (OR 0.33 (95% CI: 0.19, 0.59).
Pre-re/treatment and at four-weeks-post-praziquantel-re/treatment in years zero, one, and two, the number of days of Kato-Katzs did not significantly affect the infection intensities (Figure 5) (all p>0.05). However, at one-week-post-re/treatment (one-week, six-months-one-week, one-year-one-week and two-years-one-week), two Kato-Katzs and to a lesser extent four Kato-Katzs systematically, and significantly, overestimated the mean infection intensity (Friedman χ2 = 33.08, d.f. = 2, p<0.001).
Infection intensity measured by the strength of the POC-CCA bands (graded from negative, + (inc. trace), ++ and +++) did not accurately predict the six Kato-Katzs infection intensity categories (negative, light, moderate, and heavy) (Table S1). However, strong positive correlations were seen between the ordinal POC-CCA band strengths and Kato-Katzs infection intensity categories (baseline: r = 0.402, p = 0.003; one-week: r = 0.647, p<0.001; four-weeks: r = 0.389, p = 0.001; six-months: r = 0.413, p<0.001; six-months-one-week: r = 0.424, p<0.001) as well as between POC-CCA band strengths and individual arithmetic mean EPG (Figures 6A to E).
Pre-treatment, one POC-CCA-t+, and two Kato-Katzs had sensitivities above 90%, however the POC-CCA-t− and t+ NPVs were extremely low, similar to one Kato-Katz (Table 3). At one-week-post-praziquantel, one POC-CCA was less sensitive and had lower NPVs than one Kato-Katz, whilst two Kato-Katzs had an NPV of only 55.6%. At four-weeks-post-praziquantel POC-CCA-t+ had high sensitivity and NPV (>80%), whilst POC-CCA-t− and one to six Kato-Katzs all had sensitivities and NPVs of <60%. Indeed, two Kato-Katzs only detected a quarter of the infections. At six-months-post-praziquantel POC-CCA-t+ was more sensitive than six Kato-Katzs, however all diagnostics showed NPV values below 15%.
Some Bugoto and Bwondha children had been treated once before the study and new praziquantel-naïve cohorts were recruited each year. We therefore also analyzed our data along timelines specific for each individual child's praziquantel exposure (Tables S2 and S3). Key results were not affected by re-analyzing the data in this manner. Pre-treatment, six-months, one-year and one-year-six-months POC-CCA-t+s showed high sensitivities but low NPVs throughout (Table S2). One-week-post-recent-praziquantel, three Kato-Katzs were required for >90% sensitivity in general, and four-weeks-post-recent-praziquantel four or five Kato-Katzs were required for >90% sensitivities (Table S3). An increased Kato-Katz sampling effort was required year on year to achieve sensitivities of >90% (Table S3), which was not clearly seen in the original M&E timeline (Figure 2, Table 2). Praziquantel-naïve children and children one-year-post-praziquantel required three Kato-Katzs for accurate S. mansoni diagnosis, whilst four Kato-Katzs were required at two-years, and five Kato-Katzs at three-years.
We evaluated one to six Kato-Katzs and one POC-CCA for S. mansoni diagnosis before and after multiple rounds of praziquantel treatment, and how test choice affects M&E and drug-efficacy interpretations. Our data support using one POC-CCA-t+ or two Kato-Katzs for pre-treatment mapping in high endemicity areas [35]–[37]. However, as MDA continues, five Kato-Katzs were required for diagnosis of children after three to ten praziquantel treatments, particularly pertinent with the Ugandan control program already in its 11th year. The sensitivity of two Kato-Katzs was low four-weeks-post-praziquantel treatment indicating that, if using this method, half the individuals would erroneously indicate a cured/negative infection. The high sensitivity of two Kato-Katzs and their agreement with the ‘gold standard’ observed at baseline was not maintained over time and with recurring praziquantel treatments. Indeed, such high sensitivity and agreement was not observed again throughout this study.
POC-CCAs are shown to be more sensitive but less specific than two Kato-Katzs [33], [37], [45], [46]. Our data show that one POC-CCA-t+ at four-weeks-post-praziquantel for praziquantel-efficacy studies and six-months-post-praziquantel for M&E, was more sensitive than six Kato-Katzs at the same time periods. Our POC-CCA-t+ baseline sensitivity (91.7%), from a 94.8% S. mansoni prevalence population, was comparable with that previously published from Côte d'Ivoire (sensitivity = 86.9%, prevalence = 91.8%) [33]. In contrast, our 73% sensitivity at four-weeks-post-praziquantel (prevalence = 34.2%) was greater than in the low prevalence Côte d'Ivoire region (sensitivity = 56.3%, prevalence = 32.9%) [33]. This may be explained by that study's rigorous nine Kato-Katzs ‘gold standard’, with our six Kato-Katzs possibly still missing infections. In addition, in Côte d'Ivoire, three POC-CCAs were performed, increasing sensitivity, in comparison with our single POC-CCA [33]. The lack of POC-CCA reproducibility data, even from single urine samples [35], are a key limitation of our study. Though utilizing matching components, our accuracies from European Veterinary Laboratory POC-CCAs, may vary from the Rapid Medical Diagnostics' POC-CCAs used in Côte d'Ivoire, however differences were not observed at higher prevalence.
At four-weeks-post-praziquantel, prevalence levels as indicated by six Kato-Katzs and one POC-CCA was nearly double (61.8%) than for just six Kato-Katzs (34.2%). Cure rates using two Kato-Katzs were >80% versus 70% with six Kato-Katzs, and only ∼25% with POC-CCA-t+. Similar results have been seen for S. haematobium [47]. Further discordance between Kato-Katzs and POC-CCA at six-months (specificity of 5.3%) may be explained by high numbers of infections missed by Kato-Katzs. It is unlikely that POC-CCA false positives are the full explanation due to high specificity at four-weeks-post-praziquantel treatment, with potentially more ‘true’ negatives and only 1% of POC-CCA giving false positives in non-endemic areas [35]. We believe that the low POC-CCA specificities are, in part, due to low sensitivities of Kato-Katzs.
When six Kato-Katzs and one POC-CCA were the combined ‘gold standard’, baseline and one-week accuracies were relatively unaffected. Four-weeks-post-praziquantel Kato-Katzs sensitivities were substantially lower, having profound implications on what is a suitable ‘gold standard’ when communities have received multiple praziquantel treatments. Latent class modeling [35], [36], with additional diagnostics [39] overcome theoretical difficulties, but is not fully applicable in praziquantel-efficacy studies, providing weighted prevalence rather than individual infection and clearance data.
Studies from the same region in Uganda demonstrating reduced S. mansoni infection prevalence and intensity levels in response to MDA [42], [48], used only two Kato-Katzs and may have overestimated annual reductions [37]. We strongly recommend, as treatment campaigns continue, increased sampling efforts and/or alternative tools to accurately record program success and CRs, to detect early drug-resistance indicators, as for STHs [49]. The need for a higher number of Kato-Katzs for accurate diagnosis as the number of previous praziquantel treatments increases is not unexpected considering the small amount of stool used in each Kato-Katz and the progressively lower egg counts. Our baseline data indicated that two Kato-Katzs had a sensitivity of >90%, whereas in contrast, praziquantel-naïve individuals (Table S3) required four Kato-Katzs for accurate predictions. This apparent conflicting result may be explained by the high number of praziquantel-naïve recruits each year, sampled after several school-based MDA rounds, lowering infection intensities through reduced force-of-infections [50], supported by Figure 3, where Kato-Katz sensitivities decrease with EPG.
S. mansoni infection intensities were not expected to vary with Kato-Katz sampling effort. However, one-week-post annual or biannual treatment, mean intensities decreased from day one to day three, likely due to continued daily reductions in egg excretion post treatment. This, and our discordant POC-CCA and Kato-Katzs results post treatment, raise interesting questions regarding parasite antigen and egg clearance, such as residual-egg clearance (with individuals with intensities of >100 EPG at one-week, not retreated, but negative at four-weeks), praziquantel-induced fecundity compensation and/or increased egg expulsion (with higher EPGs at one-week-day-one, than at baseline, as also observed in S. haematobium [51]).
In contrast, at four-weeks-post-praziquantel, positive POC-CCA results in egg negative individuals may have occurred due to juveniles unaffected by treatment, newly acquired infections, and/or worms which survived treatment, but with reduced or cessated egg production (embryostasis). Drug-induced embryostasis, has been demonstrated in Onchocerca volvulus [52] and Ascaris suum [53]. Embryostasis could explain our lower sensitivities (73%) at four-weeks-post-praziquantel than those observed in a stable, low transmission Western Kenyan region (prevalence 38.8%, sensitivity = 96%) [34]. In this Kenyan region a large proportion of individuals may be truly negative with no egg or antigen excretions. Embryostasis could significantly affect drug-resistance selection, with worms repeatedly exposed to praziquantel, without dying or being detected by standard parasitological techniques. Being impossible to sample adult worms directly, studies on worm antigens, egg production and molecular studies incorporating sibship analyses informing adult breeding numbers [54], [55] post-praziquantel treatment may elucidate this.
As intensities decrease, costs of accurate diagnoses by Kato-Katzs will rise due to greater sampling requirements. Diagnoses using urine, rather than stool, remain quicker, cheaper on labor costs, more convenient, socially acceptable and may improve compliance [35]. In low endemicity areas, pooled urine samples for POC-CCAs could reduce costs further. However, key POC-CCA limitations are their inability to detect STHs, and inaccuracy measuring infection intensities and treatment resolutions. Multiple smears from one stool (versus multiple stools) and FLOTAC [49] may be viable diagnostic alternatives. If one sampling day can accurately detect schistosomiasis and STH infection intensities and ERRs, it may be highly cost-effective, warranting further research. As the geographical distribution of STH infections are more homogeneous than schistosomiasis, WHO recommends surveys of smaller subsets of schools for mapping and M&E [15]. We therefore recommend widespread POC-CCA use, with Kato-Katzs performed in a subsection of schools. For drug-efficacy studies we recommend at least six Kato-Katzs or one POC-CCA, with further research on clearance dynamics of eggs and antigens post treatment needed.
At least four Kato-Katzs (two smears per stool from two stools) are required for M&E, in the early years of a MDA program in a highly endemic area, increasing to six Kato-Katzs (two smears per stool from three stools) by year three. One POC-CCA is a suitable alternative to current prevalence M&E protocols, but they provide no information on STHs and limited intensity data post treatment, therefore we recommend their use for S. mansoni M&E with Kato-Katzs performed in a subset of schools. For drug-efficacy studies, at least six Kato-Katzs (two smears per stool from three stools) are required for accurate prevalence assessment four-weeks-post-praziquantel treatment. POC-CCAs may be a promising alternative with low specificity findings potentially due to low Kato-Katzs sensitivity, however further work is required to elucidate POC-CCA's full potential for drug-efficacy studies. Further work on improved ‘gold standards’ is required to elucidate discordant POC-CCA and Kato-Katzs results. Data on multiple Kato-Katzs from a single stool post-treatment would ascertain if accuracies of multiple days of Kato-Katzs or POC-CCAs could be matched, minimizing logistical costs without overestimating M&E success and drug efficacy, whilst retaining vitally important intensity data.
|
10.1371/journal.pcbi.1003984 | Dissecting Dynamic Genetic Variation That Controls Temporal Gene Response in Yeast | Inter-individual variation in regulatory circuits controlling gene expression is a powerful source of functional information. The study of associations among genetic variants and gene expression provides important insights about cell circuitry but cannot specify whether and when potential variants dynamically alter their genetic effect during the course of response. Here we develop a computational procedure that captures temporal changes in genetic effects, and apply it to analyze transcription during inhibition of the TOR signaling pathway in segregating yeast cells. We found a high-order coordination of gene modules: sets of genes co-associated with the same genetic variant and sharing a common temporal genetic effect pattern. The temporal genetic effects of some modules represented a single state-transitioning pattern; for example, at 10–30 minutes following stimulation, genetic effects in the phosphate utilization module attained a characteristic transition to a new steady state. In contrast, another module showed an impulse pattern of genetic effects; for example, in the poor nitrogen sources utilization module, a spike up of a genetic effect at 10–20 minutes following stimulation reflected inter-individual variation in the timing (rather than magnitude) of response. Our analysis suggests that the same mechanism typically leads to both inter-individual variation and the temporal genetic effect pattern in a module. Our methodology provides a quantitative genetic approach to studying the molecular mechanisms that shape dynamic changes in transcriptional responses.
| Genetic variation is postulated to play a major role in transcriptional responses to stimulation. Such process involves two inter-related dynamic processes: first, the time-dependent changes in gene expression, and second, the time-dependent changes in genetic effects. Although the dynamics of gene expression has been extensively investigated, the dynamics of genetic effects yet remain poorly understood. Here we develop DyVER, a method that combines genotyping with time-series gene expression data to uncover the timing of transitions in the magnitude of genetic effects. We examine gene expression in yeast segregants during rapamycin response, finding several distinct ways of change in the magnitude of genetic effects over time. These include impulse-like and sustained transitions in genetic effects, acting both in cis and trans. Our findings suggest that associations of genes with the same genetic variant often occur via the same timing of state transition in genetic effects. Furthermore, the results uncover a previously unknown variant whose impulse-like temporal genetic effect suggests a novel molecular function for determining the timing rather than the magnitude of response. Our results show that steady-state association studies miss important genetic information, and demonstrate the power of DyVER to render a comprehensive map of dynamic changes in genetic effects.
| Inherited variation in gene expression is likely to have a major effect on cellular and disease phenotypes, and may allow the underlying DNA polymorphisms (genetic variants) to be identified [1]. The genetic effect of a particular variant on a certain RNA is the quantitative change in gene expression that is associated with changing the variant's genotype (allele). Two recent studies have demonstrated that genetic effects on longitudinal gene expression data might be either stable – where the genetic effect is similar at all time points (a non-dynamic effect pattern; Fig. 1A) – or flexible, changing the magnitude of effect during time points (a dynamic effect pattern; Fig. 1B,C) [2], [3].
Dynamic effect patterns may be described in terms of the shape of changes in genetic effects over time. A linear-like genetic effect pattern (Fig. 1B) reflects a gradual change in the magnitude of genetic effects, whereas in a non-linear genetic effect pattern (Fig. 1C), the level of genetic effect is sustained in some time periods and spikes up or down in others (Fig. 1C). In most studies, transcription responses across individuals have been monitored only in two time points (before and after stimulation) and therefore the dynamics of changes in genetic effects over time could not be characterized [4]–[9].
Understanding non-linear genetic effects can, in principle, allow the timing of influence of certain regulatory mechanisms to be revealed. For example, a single state-transitioning in genetic effects may uncover the timing of alteration in a regulatory mechanism interacting with a genetic variant (e.g., transition to a new steady state at t3, Fig. 1C, left). Such a mechanism can be revealed even when additional mechanisms are acting in parallel (e.g., up-regulation during the entire time course; Fig. 1C, left). The linear genetic effect pattern, in contrast, lacks sharp alterations and therefore does not specify finely-timed information about regulatory mechanisms (Fig. 1B).
This study is focused on mapping temporal patterns of non-linear genetic effects and using this information to address major questions about dynamic transcription responses. Which dynamic genetic effect patterns are prevalent in global gene responses? Are there any general principles - either functional or mechanistic - shared among genes carrying the same temporal genetic effect patterns? Can we derive insights about the mechanisms underlying such dynamic genetic effect patterns?
Here we developed DyVER (Dynamic Variant Effect on Response), a statistical framework to predict genetic variants and study their dynamic changes in genetic effect sizes. DyVER was mainly designed to achieve an accurate detection of non-linear genetic effects (Fig. 1C) during time points. The methodology is based on the notion of a two-state digital model that pinpoints the particular time point at which a rapid change in genetic effects occurs; it is therefore suitable for revealing the timing of state transitions in genetic effects. DyVER takes as input synchronous data in several time points and across a population, and is tailored for recombinant inbred strains that are commonly utilized in genetic studies [2], [10]–[14].
DyVER differs from extant genetic approaches in several aspects. First, some existing methods construct a full model of the response curve across individuals. Their number of parameters is therefore increasing with the number of time points (e.g., [15]). DyVER, in contrast, is primarily designed for the specific task of identifying the time points of alterations in effect sizes. This partial modeling allows the use of only a small number of parameters regardless the number of time points and the shape of the temporal pattern. Secondly, DyVER is focused on modeling the dynamics in genetic effects while eliminating the confounding gene expression variables. This is unlike extant approaches, which commonly fit both gene expression and genetic effects to a certain function over time [11], [15]–[20]. Finally, if desired, DyVER can exploit the order in the input time course data, unlike several approaches that are based on unordered correlated traits (e.g., multivariate methods [21], [22] or dimension reduction methods [23]). Notably, DyVER is a practical translation of differential expression approaches (with or without time-series data [24]–[26]) for the case of statistical genetic studies.
Here we report on the use of DyVER to investigate temporal gene responses at six time points after stimulation with the TOR inhibitor rapamycin and across genotyped yeast segregants [27]. The results depict a complex map of non-linear changes in genetic effects. We identify a causal variant that affects the timing of spike up in transcript levels. Importantly, our findings suggest a previously unknown high-order temporal coordination of genetic effects: modules of genes influenced by a common dynamic genetic variant not only participated in the same biological pathway, but also shared orchestrated dynamics of genetic effects. Based on this modularity, we hypothesize that in some cases dynamic effect patterns are a property of the regulatory mechanism within which a genetic variant resides (rather than a property of the target responding transcript). We demonstrate that using this notion it is possible to enhance the identification of underlying causal genes based on their characteristic temporal effect pattern. Our results indicate the utility of studying dynamic genetic effects acting on global gene transcription.
We devised a new method, DyVER, to identify genetic variants that underlie the expression of genes and their particular dynamic effect patterns. DyVER takes as input the measured transcription response of a gene over several consecutive time points following stimulation and across a cohort, as well as a set of potential genetic variants and their genotyping (Fig. 2A). Given a candidate genetic variant with two alternative alleles, DyVER proceeds in three steps (Methods): (1) It first calculates the observed effect of the variant, namely the difference in gene response between strains carrying the two distinct alleles (Fig. 2B). The observed genetic effects are used as data in the subsequent steps. (2) To identify non-linear dynamic shapes of genetic effects, DyVER assumes a ‘digital’ regulatory model that distinguishes two possible states of genetic effects: first, a strong effect of genetic variant on the gene response (denoted the high-effect state); and second, a lower (such as zero) effect, or possibly an opposite effect (denoted the low-effect state). Several previous methods have employed a two-state model, although not in a dynamic or a genetic effect context [28]. Based on a maximum likelihood approach, DyVER seeks a genetic variant and a sequence of states that best describe the dynamic changes in the size of the genetic effect. For example, if a gene is affected mainly by a variant v during a late time interval, DyVER successfully infers the correct effect pattern low→low→high→high for the correct variant v as it attains the highest likelihood score (Fig. 2B and C, right panel). For incorrect variants, the likelihood scores are typically lower (Fig. 2B and C, left panel). DyVER's predicted sequence of states is referred to as the temporal two-state model. Finally, (3) DyVER calculates the statistical significance of association for each genetic variant based on a likelihood ratio score that takes as input the inferred temporal two-state model (Fig. 2D). We refer to this score as the DyVER score. Notably, although DyVER requires synchronous observations in particular time points, it is still possible to apply DyVER on partial observations in each of the time points (Methods).
Overall, step 1 allows DyVER to focus on dynamics in genetic effects regardless of the magnitude of transcription response, whereas the discrete modeling in step 2 allows detecting any sequence of spikes up or down in genetic effects. The two-state model from step 2 enhances the performance of the DyVER score (step 3) by allowing a separate parameterization for each of the states. Specifically, to infer an optimal temporal two-state model, DyVER uses a two-state hidden Markov process where the observed effects are treated as the outcome of a sequence of hidden high-effect and low-effect states (step 2; Fig. 2C). The corresponding likelihood function consists of two components: (i) the probability of observed effects given a certain temporal two-state model; and (ii) the probability of a temporal two-state model, which may use a penalty factor to prioritize two-state models with a lower number of transitions between states, assuming dependencies among consecutive time points. In the absence of penalty, the order of time points is irrelevant and therefore the predicted two-state model can be viewed as a partition of an unordered group of time points into two sub-groups. The DyVER score exploits this partition for a different parameterization of the (unordered) time points in each of the two states. The addition of the penalty factor makes it possible to avoid an overfitted two-state model that is then given as input to the next step, hence further improving the DyVER score's performance.
We compared DyVER's performance to that of five alternative methods. In the first method, the most naïve approach, an ANOVA test is applied at each time point independently and the predicted genetic variant is the one with the most significant (minimal) ANOVA P value score. The second method builds on dimension reduction using principal component analysis (PCA): Given T time points for each strain as input, it first reduces the T-dimensionality of the data into a single dimension by projecting each strain onto the first principal component. Next, it applies an ANOVA test on this one-dimensional data [23]. The third method models dynamics in gene expression as well as dynamics in genetic effect sizes [15]. For comparison, in the fourth method, a linear time term is included as a covariate in the ANOVA test to model dynamic changes in gene expression (without direct modeling of dynamics in genetic effects). Finally, we compared DyVER to a random prediction of association relationships. We called these approaches ‘naïve’, ‘PCA’, ‘detailed dynamics’, ‘expression dynamics’ and ‘random’, respectively. In both DyVER and all compared methods, for each simulated gene, the resulting P values were Bonferroni-corrected for the testing of multiple genetic variants. The quality of predicted variants were evaluated using the accuracy metric, defined as the tradeoff between the sensitivity and specificity of revealing genetic variants across different significance cutoffs. The accuracy metric ranges between 0 and 1 for poor and excellent performance, respectively (Methods).
To characterize DyVER's ability to reveal dynamic genetic variants and distinguish their effect patterns, we generated synthetic collections of genes that are associated with genetic variants over time. A single synthetic ‘collection’ consisted of 500 genes, 300 of them associated with a genetic variant over time, with two characteristic parameters: (i) the number of time points, and (ii) the effect size (in all cases we used 50 strains and 100 genetic variants). In a complete synthetic ‘dataset’ we generated 72 collections for various numbers of time points and effect size values. Overall, four synthetic datasets were generated in this study, each consisting of a different key class of dynamic effect patterns (see Methods): a linear-like pattern (Fig. 1B), a single state-transitioning based on a sigmoid function (Fig. 1C, left), and impulse and multiple-pulse (complex) patterns based on the product of two sigmoid functions (Fig. 1C, middle and right, respectively) [24]. In the following, we first analyze the performance of the DyVER's predicted associations (based on the DyVER score) in the absence of penalty and then present the contribution of the penalty factor.
DyVER showed good accuracy in all non-linear dynamic effect patterns (0.5 penalty; Fig. 3). Fig. 3A presents the accuracy metric for synthetic datasets of varying numbers of time points. Accuracy values are averaged across the eight collections of distinct effect size. In all non-linear dynamic effect patterns, DyVER displayed the best accuracy in all tested time points ranging between 3 and 27, with improved accuracy for a larger number of time points. Importantly, although DyVER was not designed for linear-like effect patterns, it still attains the second-best performance for this case. The ‘expression dynamics’ approach yielded the most accurate predictions for the linear case, but attained poor results in the non-linear case. The tradeoff between sensitivity and specificity in the accuracy measure across the different methods is further demonstrated in Figure S1A and B. Results were similar for varying effect sizes (Fig. 3B and Figure S1C and S1D) and for an additional synthetic dataset that is based on prototypical effects in C. elegans (Methods; Figure S2). Furthermore, although DyVER's accuracy is reduced in the case of missing data, it is still notably high in comparison to alternative methods (Figure S3). Taken together, our results indicated that DyVER performs well on a broad range of genetic effect patterns.
We next aimed to characterize DyVER's applicability to short-term steady state of high genetic effects. To tackle this goal we compared two synthetic impulse datasets, both consisting of 27 time points across various effect sizes. For all genes, the short-impulse dataset consisted of a high-effect steady state of short duration (five time points), whereas the long-impulse dataset consisted of a high-effect steady state of long duration (fifteen time points). Figure S4A records the performance of DyVER compared to the five alternative methods on the short-impulse and the long-impulse datasets, and clearly shows that DyVER outperformed the alternative procedures when genetic influences were acting in short impulses, even with low-effect sizes.
The performance of both DyVER and the alternative methods declined when applied on a short impulse compared to a long impulse of genetic effects, but notably, the performance reduction was lowest with DyVER (Figure S4B). For example, for high-effect sizes (0.625), the sensitivity of DyVER is 0.7 and 1 with short and long impulses, respectively. The sensitivity of PCA, in contrast, is respectively 0.47 and 1 with short and long impulses for the same effect size. Thus, even when genetic variants acted during short time intervals, DyVER still performed relatively well. This was unlike the alternative methods, whose performances were drastically reduced even for relatively high-effect sizes.
DyVER predicts a temporal two-state model, which may provide insights concerning the timing of changes in genetic effects (Fig. 2C). To evaluate the quality of this prediction, we compared the ‘ground truth’ (simulated) models against the inferred two-state models. We chose to work with the established error rate statistics, defined as the number of erroneous two-state models expressed as a fraction of the total number of significant correctly predicted variants. We called this metric a two-state pattern error rate (in short, error rate), and calculated it both for the case of stringent (exact) matching or flexible (non-exact) matching between the true and inferred models (Methods). In both cases, we found that DyVER performs well in predicting two-state models, where the flexible case outperforms the stringent case, as expected. For example, using single state-transitioning patterns with nine time points, effect size 0.75, significance cutoff 0.001 and the absence of penalty (probability of transition 0.5), the stringent and flexible error rates are 0.41 and 0.33, respectively (Figure S5). The error rate increased with decreasing penalty (e.g., for transition probabilities of 0.01 (high penalty) and 0.5 (no penalty), stringent error rates are 0.32 and 0.41, respectively). As expected, error rates rose when a higher statistical significance cutoff (0.05) was used, whereas the gap between the error rates for different significance cutoffs remained relatively constant when the penalty increased. Results obtained for other effect sizes were similar.
Collectively, our results indicated that DyVER outperforms extant methods even in the absence of penalty and the presence of missing data (Fig. 3, Figures S1–S4), and that these performance can be even enhanced by the addition of a penalty component (Figure S5). These results hold when the complexity of dynamic effect patterns is relatively low, as in the case of genetic effects in biological data (e.g., Figure S6).
We applied DyVER in an unbiased manner (without penalty) to the available dataset of 95 yeast segregants that were stimulated by rapamycin and profiled at six time points (Methods) [27]. DyVER predicted 351 associations to 145 distinct variants (false discovery rate [FDR] 6%). Of these 351 associations, 145 had highly significant dynamic associations (15% FDR, Table S1, Methods) and 105 of them showed non-linear genetic effect patterns (Fig. 4). In agreement with previous findings [2], [11], our results suggest that non-linear associations are prevalent: of the eight previously known causal genes, six were found to have an association with at least one target gene exhibiting a non-linear genetic effect pattern (Table S2). Correlations among genetic effects of consecutive time points were much larger than correlations between non-consecutive time points [P value <10−15 (Wilcoxon test)], justifying our ‘memoryless’ Markov assumption that the next time point is mainly dependent on the current time point (Figure S7). The 105 genes carrying non-linear effect patterns were partitioned into groups based on their predicted two-state pattern (Table S1); seven two-state pattern groups (C1–C7) were created, each including at least two genes (Fig. 4A and B).
The partition revealed three prototypical non-linear genetic effect patterns (Fig. 4A), including (i) a single upward spike followed by a sustained high level of genetic effect (70 genes in C1–C4). These different groups were characterized by distinct timing of a state-transitioning, including an abrupt change in early time points (0–10 min, C1), as well as an intermediate-early (0–20 min, C2) and intermediate-late (20–30 min, C3) single state-transitioning. For example, SFA1 and ESF1 (in groups C1, C2) demonstrate a sustained genetic effect with a state transition at 0–10 and 0–20 minutes after rapamycin stimulation, respectively (Fig. 4B). In the case of the four genes exhibiting a late state-transitioning (at 30–50 min, C4), a sustained new level of genetic effects might occur at later time points that were not measured in the current dataset [27]. (ii) A single downward spike of genetic effect (C5–C6, 22 genes). In group C5, we observe an abrupt downward spike in 10–20 minutes followed by a sustained low level of genetic effect (for example, PHM6, Fig. 4B). Group C6 represents a delayed gradual single state-transitioning during 20–50 minutes. (iii) An impulse of high genetic effect at 10–30 minutes after treatment (9 genes in C7, e.g., UGA4, Fig. 4B). Overall, the single state-transitioning patterns were over-represented, whereas complex patterns of genetic effects were rare (1 gene, YER053C-A) and were under-represented [cis: P value <10−19, trans: P value <10−50 (t-test), (Figure S6A)]. Our findings of rare complex patterns in yeast parallel similar observations in the mouse (Methods, Figure S6B); Yet, the particular shape of effect patterns may differ between biological systems (Figure S8).
We next explored the pleiotropic trans-acting variants that arise from this analysis. Using DyVER's predictions we organized the genes into six co-association modules, each containing a group of (at least two) genes with the same trans-associated variant (Fig. 5A and B). Functional enrichment strongly related all six modules with specific biochemical pathways. For example, the entire module no. 3 consists of genes that play a role in uptake of phosphate (Pi) from extracellular sources and its accumulation in vacuoles (5 of 5 genes; Fig. 5A and B, Figure S9A). The module's validated causal gene is PHO84, a high-affinity phosphate transporter that carries a missense mutation in one of the parental strains (Figure S9B) [29], [30]. The nine genes in module no. 5 carry two distinct functionalities and are therefore treated as two distinct sub-modules, no. 5-I and no. 5-II (three daughter cell-specific genes and six poor nitrogen source degradation genes, respectively, Fig. 5A).
Next we examined whether module genes show characteristic temporal effect patterns. On analyzing the modules we found that modules nos. 1, 3, 4, 5-I and 5-II relate to a specific prototypic temporal genetic effect pattern, whereas the remaining two modules (nos. 2 and 6) are more general and show several distinct patterns (Fig. 5A). For example, module no. 1 contains 34 genes, 32 of which have an upward spike (a single state transition) of genetic effect at 10–30 minutes after rapamycin stimulation [FDR 0.01 (hyper-geometric test)]. As another example, module no. 3 contains five genes, all showing a downward spike of genetic effects at 10–30 minutes after stimulation. Specifically the downward spike occurs either 20–30 minutes after stimulation [4 genes, FDR 0.01 (hyper-geometric test)] or 10–20 minutes after stimulation (1 gene, Fig. 5A–C, Figure S9C). Overall, we found four modules with over-represented patterns of single state-transitioning at specific time points (nos. 1, 3, 4 and 5-I) and one sub-module of an impulse effect pattern (no. 5-II). The observed coordination of temporal genetic effects does not necessarily reflect a coordination of transcription responses (Figure S10). In previous reports, baseline expression levels were used to identify eight genetic variants underlying similar modules (Table S2), but the coordinated temporal genetic effects and the timing of upward or downward spikes of genetic effects were not characterized.
A plausible explanation for the ‘shared variant, shared temporal genetic effect pattern’ hypothesis is that the same molecular mechanism underlies both inter-individual variation and the dynamics of genetic effects. In such cases, the dynamic pattern of effect is an attribute of the underlying regulatory mechanism (rather than of the target genes), probably owing to temporal changes in the influence or activity of the regulatory mechanism. This hypothesis is further supported by the consistency in the timing of state transitions in module genes and their underlying (known) causal genes (Figs. 5D versus 5E): The trans-associated causal gene of module no. 1 (IRA2) attains a sustained-like pattern of gene expression that resembles the temporal genetic effect pattern of its target genes (Fig. 5D and E, left). The cis-associated causal genes in modules nos. 3 and 4 (PHO84 and GPA1) exhibit drastic changes in their transcription response at the same time point at which there is a (downward or upward) spike in the genetic effect of their target genes (20–30 and 30–40 min; Fig. 5D and E, middle and right, respectively).
The poor nitrogen source degradation system (module no. 5-II) demonstrates the ability of our method to reveal novel associations acting on the timing of response and affecting an entire cellular pathway (Figs. 5, 6). During growth on relatively poor nitrogen sources (allantoate, allantoin, and GABA), yeast cells activate premeases responsible for uptake of nitrogen sources and further increase the expression of enzymes that participate in degradation of poor nitrogen sources for the generation of ammonia. Exposure to the TOR inhibitor rapamycin also leads to the same nitrogen-regulated response [31]. Module no. 5-II consists of six of the twelve genes in the allantoin, allantoate and GABA degradation pathways, with all six genes having a significant impulse effect pattern (DAL1, 2, 4, 7, 80 and UGA4; Fig. 6A–C). An additional gene in these pathways, DAL5, is weakly associated using the same impulse pattern at the same genomic position (Fig. 6A–C).
The impulse pattern reflects a difference in the timing of initiation of response among the strains carrying the RM and BY alleles in Chr2: 533–562 kb. For example, strains carrying the BY allele showed early up-regulation of DAL80 in response to rapamycin, which was already detected at 10 minutes after stimulation. The RM-carrying strains, in contrast, showed a clear delay in response to rapamycin, but all strains reached a similar expression level by 30 minutes after stimulation (Fig. 6D). The underlying genetic variant acting on the timing rather than on the magnitude of response has not been previously documented. In the genomic interval (Chr2: 533–562 kb), two genes (RPB5, CNS1) have temporal transcription profiles that match the expected early impulse of high genetic effect, the promoter of five genes (RPB5, CNS1, ADH5, RTC2, YBR144C) is bound by nitrogen-related transcription factors [32], and four genes (RPB5, CNS1, ADH5, RTC2) were previously reported in nitrogen-related cellular processes (Figure S11). These criteria therefore suggest that RPB5 or CNS1 are two leading candidates in module 5-II.
In this work we present the DyVER computational algorithm for identifying genetic variants that lead to dynamic changes in genetic effects. DyVER was tailored to identify abrupt changes in the levels of genetic effects, which may provide valuable information about the timing of alterations in the particular regulatory mechanisms interacting with the underlying genetic variant. In comparison with other approaches, DyVER attained the most accurate identification of non-linear genetic effect patterns, even in the absence of penalty (Fig. 3, Figures S1–S4), likely due to (i) a focus on genetic effects rather than on modeling the original phenotype values, and (ii) the prior knowledge about the separation of the time points into two distinct groups that differ in their observed effects (encoded in the temporal two-state model), thus allowing a different parameterization for each of these groups.
DyVER is using an HMM-based model for revealing genetic variants acting on time-series gene expression data. HMM modeling has been applied in various contexts, but not for the case of direct identification of underlying genetic variants. For example, HMM has been utilized for the identification of CNVs or haplotypes [33], [34]. Alternatively, an existing method was mainly focused on revealing differential expression between conditions using an HMM approach [26]. DyVER extends this method by providing a statistical genetics P value score and by allowing a number of parameters that is not increasing with the number of strains.
Our method opens multiple directions for future investigations. First, it is important to extend DyVER for the case of outbred heterozygous population, including human. In the current study, DyVER was designed for the case of a inbred (homozygous) strains that are common in genetic studies (e.g., in yeast, nematode, fly, mouse and rat) due to several major advantages: first, inbred strain enable controlled stimulations, and second, they avoid major challenges that are common in human genetics, including haplotype analysis, rare variants and uncontrolled variables. Future extensions may generalize the method for the heterozygous case, possibly by calculating genetic effects between each pair of genotypes (rather than between the only two possible genotypes as in the homozygous case), requiring to add additional one or two Gaussians within each of the model states. Second, the usage of a our probabilistic model leads to several limitations: the number of states should be specified in advance; we only capture correlations between sequential time points but cannot capture higher-order correlations among time points; and we generally assume that the probability of a time point is independent of the probabilities of its neighboring time points. Future improvements that handle more than two states and a more sophisticated probabilistic graphical model [35] may therefore enhance DyVER's performance. Third, DyVER relies on at least a few synchronized strains in each of the time points. Although DyVER allows missing data and possibly different strains in different time points (Figure S3), it still cannot be applied on non-synchronous data (as in [11]). Data imputation methods can potentially enhance the DyVER analysis beyond this synchronization requirement.
Building on the DyVER approach, we analyzed temporal gene expression patterns following rapamycin treatment in yeast segregants. Our analysis identified 105 genes exhibiting significant non-linear genetic effects over time, 56 of them are well-established associations (in modules 1,2,3,4,5-I and 6), and the remaining genes are new candidates for future experimental investigations (e.g., Fig. 4B). For example, our study suggests a novel genetic variant residing in chr2: 533–562 kb as the underlying regulator of the timing of upward spikes in gene expression after rapamycin treatment. Reassuringly, this regulator acts primarily on genes that play a role in poor nitrogen source degradation (6 of 6 genes, module 5-II, Fig. 6).
The application of DyVER in yeast provided several novel insights that were mainly attained due to the unique capability of DyVER to classify associations based on their optimized temporal effect patterns. First, we use the temporal effect pattern to automatically organize the genes into clusters based on their predicted patterns (Fig. 4A and Figure S12). This organization is substantially different from previous studies [2], [11] that have grouped time-series associations only manually. Based on this clustering, we found that abrupt single state-transitioning and impulse patterns occur in certain prototypical time points. In particular, DyVER identified an upward spike of genetic effect at 0–10, 0–20, 20–30 and 30–50 minutes (22, 34, 10 and 4 genes, groups C1, C2, C3 and C4, respectively); a downward spike followed by a new sustained low level of genetic effect (6 and 16 genes at 10–20 and 20–50 minutes, groups C5 and C6, respectively), and a single pulse of high genetic effects (9 genes, group C7, Fig. 4).
Second, many studies have shown that groups of co-associated genes also share similar functionalities. Interestingly, our results indicate that such co-associated genes typically share not only a similar functionality, but also a similar predicted pattern of temporal genetic effect (Fig. 5). One plausible explanation is that a causal regulator typically alters its functionality during its response to stimulation; therefore, a genetic variant interacting with such a regulator is likely to affect its target only during those time intervals in which the regulator is functional. Based on this rationale, the temporal effect patterns in target genes may uncover the temporal dynamics of their causal regulatory mechanisms. Thus, DyVER's characterization of temporal effect patterns, which are probably a property of the causal regulatory mechanisms, may provide a starting point for improved identification of causal genes. For example, it might be possible to pinpoint a causal gene in a genomic interval based on its predicted dynamics over time (as demonstrated in Fig. 5D and E and Figure S11C). Furthermore, it may be possible to discriminate between two genetic variants differing in their dynamic over time, even when these variants are co-localized at a nearby genomic position (as in module nos. 5-I and 5-II, Fig. 5A).
Taken together, our results highlight the utility of studying temporal genetic effect patterns to discover and characterize dynamic causal regulators. The next step is to extend and apply our approach to map genetic effects in transcriptome of a wide range of mammalian cell types.
To generate synthetic data we first generated 50 strains carrying 100 genetic variants, sampling one of the two alleles with equal probabilities. A single synthetic collection consists of 500 genes, of which 300 are associated with a certain variant over T time points. Overall, for a single dataset we generated 72 collections, constructed for all combinations of eight possible ‘effect sizes’ (defined below, ranging between 0.125 and 1) and nine different numbers of time points (ranging between 3 and 27). In all cases, the low-effect state represents the absence of effect () and the high-effect state represents the presence of an effect (), where is the effect size. is the simulated observed effect, which is generated by sampling from a Gaussian distribution . A dataset was constructed for each class of temporal effect patterns. For a single state transition effect pattern (here, sustained) we used a sigmoid function:
Where , q = v = 0.5 and . For an impulse effect pattern we used the product of two sigmoid function with five parameters [24], where , is the length of the impulse effect (here, ):
To generate the complex pattern for T time points we concatenated two impulse patterns, each for T/2 time points. For the dataset of linear effect patterns, observed effects are sampled from a linear function:
For the purpose of comparing predicted to gold-standard temporal two-state models (Figure S5) we generated a different collection of synthetic sustained dataset as follows: we first generated the temporal two-state model by sampling from the corresponding distribution (from equation 3) with . The observed effects were then generated by sampling from the corresponding Gaussian distribution , , where and are the mean of the high- and low-effect state, and . To generate an input with a percentage of k% missing data, in each time point, we omitted the information for k% randomly selected strains (thus, each time point consists of a different list of strains).
An additional synthetic dataset was created similarly to the above datasets, but using previously published functions in C. elegans [11]. For each of the 300 associated genes in this synthetic dataset, we first randomly chose a function out of the 18 functions that were published in C. elegans; the observed effects were then sampled from this selected function.
The compared methods were implemented as follows. In the ‘naïve’ method we assumed a simple fixed effect model on each time point independently, , where is the observed expression level for strain j carrying genotype i; is fixed effect of genotype i and . The most significant (minimal) ANOVA P value score is taken as the resulting P value. In the ‘PCA’ method, we project the T-dimensionality of each strain into the first principal component and then applies an ANOVA test assuming a fixed effect model where is the first principal component for strain j carrying genotype i; is the fixed effect of genotype i and (the first principle component was chosen since it performs better than the consecutive components, see Figure S15). For the ‘expression dynamics’ method, we used the model where is the observed expression level in time point t for strain j carrying genotype i, and are two fixed effects for genotype i and . The formulation was implemented using the lme4 R package. In all cases above, an F-test was used to test the model. For the more sophisticated ‘detailed dynamics’ method, we use the longGWAS R package that is part of its original publication [15].
For each synthetic dataset, DyVER was applied to predict a genetic variant using the DyVER score (P values were Bonferroni-corrected for multiple variants). To quantify the ability to correctly predict such genetic variants, we define the accuracy measure. Genes are split into two groups: one contains genes that are associated with a genetic variant, and the other contains the remaining, non-associated genes. A mapping method may provide a negative prediction (i.e., a non-significant P value for all candidate variants), or alternatively, a positive prediction of either the correct variant or an incorrect variant. We define true positives as associated genes whose correct genetic variant is predicted with a significant P value. True negatives are non-associated genes that were not significantly associated with any variant. False negatives are associated genes that were not significantly associated with any variant. Finally, false positives are defined as erroneous significant predictions as a result of two possible scenarios, either a non-associated gene that is wrongly predicted to be associated with a certain variant, or alternatively, an associated gene whose predicted variant is incorrect. We adopt the standard formulations for sensitivity (number of true positives out of the total number of positives) and specificity (number of true negatives out of the total number of negatives).
Similarly to a standard ‘Receiver Operating Characteristic’ (ROC) analysis, we can plot the sensitivity against the 1-speificity across different P values, providing an overall view of the performance of the method: the higher the curve, the better the accuracy (defined as the area under the curve). Notably, using a standard sensitivity definition, sensitivity should increase with higher P value thresholds. In contrast, using our definition of sensitivity, it is dependent on the particular predicted variant. Thus, even with a very high P value threshold and many affected genes, the sensitivity of a random algorithm might remain close to zero. The accuracy therefore ranges between 0 (for a random prediction) and 1 (for a perfect prediction).
Finally, to quantify the ability of DyVER to correctly predict the temporal two-state model, we define the two-state pattern error rate (shortened to error rate) as the number of wrongly predicted temporal two-state models expressed as a proportion of the total number of (significant) correctly identified variants. We test two different rules for matching between the simulated and predicted model. In the stringent case, we require a fully correct two-state model, and in the flexible case, we require correct transitions between states but allow incorrect timing of transition.
We applied DyVER to genotyping data and gene expression data that were monitored during six time points following exposure to rapamycin in 95 yeast segregants and their two parental yeast strains: BY4716 (BY) and RM11-1a (RM) [27]. DyVER was applied to the log expression of 2700 genes with the highest difference between the BY and RM parental strains. To ensure that the biological results are unbiased, DyVER was applied with penalty 0.5. Multiple testing was controlled as follows: DyVER score P values were first Bonferroni-corrected for multiple variants; the corrected DyVER score P values were then controlled for multiple testing of genes (FDR 6%). We then further filtered the genes based on the dynamic association score (FDR 15%). In total out of 2700 genes, we obtained 351 (13%) predicted associations (based on the corrected DyVER score P value) and 145 (5.3%) predicted dynamic associations (based on the dynamic association score). Next, we further removed 40 genes carrying linear-like patterns, based on strong correlation with a linear model (r>0.95) and more than 5% change in genetic effect in any two consecutive time points (Table S1). The partition into groups was generated automatically according to DyVER's predicated two-state model (Figure S12).
In addition, we applied DyVER to genotyping data and log gene expression data of 403 genes that were monitored using a meso-scaled technology during three time points following exposure to lipopolysaccaride in 45 mouse BXD strains [2]. Of the 403 genes, 14 genes (3.4%) were identified as significant dynamic associations (FDR 10%; Figure S6B).
|
10.1371/journal.pgen.1006005 | The Genomic Basis of Evolutionary Innovation in Pseudomonas aeruginosa | Novel traits play a key role in evolution, but their origins remain poorly understood. Here we address this problem by using experimental evolution to study bacterial innovation in real time. We allowed 380 populations of Pseudomonas aeruginosa to adapt to 95 different carbon sources that challenged bacteria with either evolving novel metabolic traits or optimizing existing traits. Whole genome sequencing of more than 80 clones revealed profound differences in the genetic basis of innovation and optimization. Innovation was associated with the rapid acquisition of mutations in genes involved in transcription and metabolism. Mutations in pre-existing duplicate genes in the P. aeruginosa genome were common during innovation, but not optimization. These duplicate genes may have been acquired by P. aeruginosa due to either spontaneous gene amplification or horizontal gene transfer. High throughput phenotype assays revealed that novelty was associated with increased pleiotropic costs that are likely to constrain innovation. However, mutations in duplicate genes with close homologs in the P. aeruginosa genome were associated with low pleiotropic costs compared to mutations in duplicate genes with distant homologs in the P. aeruginosa genome, suggesting that functional redundancy between duplicates facilitates innovation by buffering pleiotropic costs.
| Novel traits play a key role in evolution by providing organisms with access to new ecological niches. Novelty is often conspicuous at a phenotypic level, but it is difficult to determine its underlying genetic basis. To address this problem, we have studied how the bacterium P. aeruginosa evolves novel metabolic traits, such as the ability to degrade new sugars, in real-time. After 30 days of evolution we sequenced the genomes of bacteria that have evolved novel metabolic traits. We found that mutations mainly affected genes involved in transcription and metabolism. Our main finding is that novelty tends to evolve by mutations in pre-existing duplicated genes in the P. aeruginosa genome. Duplication drives novelty because genetic redundancy provided by duplication allows bacteria to evolve new metabolic functions without compromising existing functions. These findings suggest that past duplication events might be important for future innovations.
| An evolutionary innovation is a new trait that allows organisms to exploit new ecological opportunities. Some popular examples of innovations include flight, flowers or tetrapod limbs [1,2]. Innovation has been proposed to arise through a wide variety of genetic mechanisms, including: domain shuffling [3], changes in regulation of gene expression [4], gene duplication and subsequent neofunctionalization [5,6], horizontal gene transfer [7,8] or gene fusion [9]. Although innovation is usually phenotypically conspicuous, the underlying genetic basis of innovation is often difficult to discern, because the genetic signature of evolutionary innovation erodes as populations and species diverge through time.
One way to circumvent this difficulty is to directly study the evolution of innovation in real time using microbial model systems [10,11]. The large population size and short generation time of microbes allows for rapid evolution under conditions that can be easily replicated. Samples from evolving populations can be cryogenically preserved in a non-evolving state so that evolved genotypes can be directly compared with their ancestors. Also, bacteria have compact genomes, making it possible to characterize the functional and genetic basis of adaptation [12,13]. Recent experiments using this approach have provided detailed examples of the evolution of a number of innovations [14–19], such as novel metabolic traits [15] and ecological specialization [19]. However, there is a difference between evolving a new trait (innovation) and improving an already exiting one (optimization) [17] and it remains unclear if evolutionary adaptations that require qualitatively new traits (innovations) generally have a different genetic basis than adaptations that require mere fine tuning (optimization) of an existing trait.
The objective of this study is to determine the genomic mechanisms underpinning evolutionary innovation and optimization using bacterial metabolism as a model system. To achieve this goal, we allowed populations of P. aeruginosa founded by a single clone to evolve in Biolog microtiter plates containing culture medium supplemented with 95 unique carbon sources. Crucially, the ancestral clone produces a clear bimodal pattern of growth on these carbon sources: in some of the carbon sources it grows poorly while in others it grows well. Carbon sources that support little or no growth above the carbon-free control challenge bacteria to evolve novel metabolic traits. These carbon sources can therefore be used to study evolutionary innovation. In contrast, carbon sources that allow the ancestral clone to grow to at least a moderate population density challenge bacteria to improve existing traits. These carbon sources can be used to study the genetic basis of evolutionary optimization. Following 140 generations of evolution we identified carbon sources that populations consistently adapted to. We then isolated clones from populations that evolved in these carbon sources and used whole genome sequencing of more than 80 evolved clones to determine the genetic basis of evolutionary innovation and optimization. To understand the pleiotropic consequences of innovation we used high-throughput phenotypic assays to measure the fitness of the clones evolved in a single carbon source in the 94 remaining substrates of the Biolog plate. This experimental strategy has two main benefits. First, by comparing the mutations and phenotypes observed in clones adapted through innovation and optimization it is possible to test for distinct genomic signatures associated with innovation. Second, by studying the evolution of multiple novel traits, it is possible to make general conclusions about the genetic basis of innovation.
We first assessed the growth of P. aeruginosa PAO1 in the 95 unique carbon sources provided by Biolog microtiter plates. Each well on a Biolog plate contains a common inorganic growth medium that is either supplemented with a unique carbon source (95 wells), or not supplemented and acts as a negative control (1 well). The parental PAO1 strain (ancestral clone hereafter) showed a clear bimodal pattern of growth in these 95 carbon sources, both in terms of viable cell titre and optical density (Fig 1A, see Materials and Methods). Some carbon sources supported very low levels of growth that were comparable to the growth observed in the negative control well; selection on these substrates challenges P. aeruginosa to evolve new metabolic traits. In contrast, other carbon sources supported good levels of growth; selection on these substrates challenges P. aeruginosa to optimize existing metabolic traits. Although this distinction is intuitive, it is necessary to formally define a threshold between innovation and optimization. To do so we fitted a mixture distribution to the viable cell titre for the 95 carbon sources. We used the point where the two distributions intersected to classify the carbon sources in two groups: innovation (carbon sources that supported poor growth, similar to the carbon-free control) and optimization (carbon sources that supported growth to high population density). This classification was also supported by optical density data (see Materials and Methods).
We evolved 4 replicate populations founded by the ancestral clone in each of the 95 carbon sources present in the Biolog microtiter plates by serially propagating cultures on 4 replicate Biolog plates for 30 daily serial transfers, which corresponds to approximately 140 generations of bacterial growth. At the end of the evolution experiment, we tested for adaptation on each of the 95 carbon sources by comparing the growth rate of the 4 replicate populations that had evolved on each carbon source to the growth rate of the ancestral clone on the same carbon source. We used growth rates to assess adaptation because they provide a higher resolution than viable cell titre, which can allow for the detection of small differences in the rate of adaptation across substrates. We note, however, that growth rate and viable cell titre measures strongly correlate (r = 0.887, P < 0.001).
Given that evolutionary innovation involves the origin of novel phenotypes, whereas optimization involves the refinement of existing phenotypes, optimization should evolve more readily than innovation. Consistent with this expectation, the proportion of populations that evolved increased growth rate was significantly lower on carbon sources that challenged bacteria to innovate as opposed to optimize existing traits (51.50% vs. 63.89%, P = 0.01, One-tailed Fisher's exact test). Moreover, the fraction of carbon sources where all 4 replicate populations evolved increased growth rate was almost 50% lower on carbon sources that challenged bacteria with evolutionary innovation as opposed to optimization (Fig 1B; P = 0.042, One-tailed Fisher's exact test).
To understand the genetic basis of adaptation, we sequenced the genome of 4 independently evolved clones from carbon sources where all 4 replicate populations evolved increased growth rates. Our rationale for this sequencing strategy is as follows. Parallel increases in growth rate suggest that selection was very effective on these substrates, increasing the probability that clones from these substrates carry potential beneficial mutations. Second, by sequencing multiple clones that evolved on the same substrate it is possible to identify genes that show parallel molecular evolution. Parallelism is common in bacterial populations, and it provides a simple way to identify genes that contribute to adaptation [19–22]. Specifically, we sequenced the genomes of 84 clones from carbon sources that challenged bacteria to both innovate (8 carbon sources, 32 clones) and optimize existing traits (13 carbon source, 52 clones). The ancestral clone produces a clearly bimodal distribution of growth on these 21 carbon sources, with an approximately 10-fold difference in mean viable cell density between carbon sources where innovation as opposed to just optimization occurred (S1 Fig). We identified 143 unique mutations in the genomes of the 84 sequenced clones, amounting to a mere 1.70 mutations per clone on average (S1 Data). These were all mutations that accumulated in the course of the experiment and that were not present in the ancestral clone. Most of the mutations that we identified were SNPs (74%), but we also detected short indels (8%), large deletions (12%), and duplications (4%). The proportions of these types of mutations did not differ between clones that had adapted through innovation and optimization (S2 Fig and S1 Table; P = 0.213, Pearson’s X2 test). Although populations of P. aeruginosa sometimes evolve elevated mutation rates during cystic fibrosis infections [19] and during long-term selection experiments [23,24], we did not find any hypermutator strains with mutations in genes involved in DNA replication and repair, such as the methyl-directed mismatch repair pathway.
Several lines of evidence suggest that most of the SNPs that we detected were beneficial mutations. First, the vast majority (97/106) of point mutations we detected were non-synonymous (S1 Table). We only detected three synonymous mutations and two affected a gene where parallel synonymous evolution occurred, suggesting that these were beneficial synonymous mutations [25]. Thus, our estimate of the rate of substitution of non-synonymous mutations to putatively neutral synonymous mutations is 97/1. Second, the number of mutations per clone was approximately 40% higher in clones that had to adapt through innovation (2.1 mutations per clone) as opposed to clones that adapted through optimization (1.53 mutations per clone) (S3 Fig; P = 0.034, two-sample one-tailed Kolmogorov-Smirnov test). Given that the number of generations of evolution was highly similar across carbon sources (see Materials and Methods), this difference in the number of genetic changes is consistent with the idea that populations that had to adapt through innovation were exposed to stronger selection. This difference is particularly striking, given that populations that had to adapt through innovation were associated with a small population size, which should reduce the rate of fixation of beneficial mutations. Finally, parallel molecular evolution was very common: 65.73% of the mutations occurred in genes that were mutated in more than one clone (S1 Data), which is significantly greater than the amount of parallel evolution expected due to chance alone (permutation test, P < 0.001).
Gene-level parallel evolution tended to occur between replicate clones that evolved on the same carbon source, and genes that were only mutated in clones from an individual carbon source accounted for 75% of the parallelism that we observed. Interestingly we found parallel evolution in all 4 replicate clones that evolved on 5 carbon sources (L-alanyl-glycine, glycyl-l-glutamic acid, L-serine, D,L- α glycerol phosphate, and glycerol), involving 24 mutations. In every case, parallel evolution on these substrates involved transcriptional regulators. Recent work in the closely related bacterium P. fluorescens suggests that parallel evolution by mutations in transcriptional regulators is common because it provides an efficient mechanism to translate genetic variation into phenotypic variation [26]. This may explain why parallel regulatory evolution was very common on some substrates. Rigorous test of this idea is outside the scope of this paper and it would require further experimental work, as in [26]. We also observed higher-order parallel evolution involving different genes that act in the same operon. Parallelism by definition becomes more common as the scale at which it is measured increases; for example, parallelism is necessarily more common when it is measured at the level of genes than at the level of nucleotides. However, it is difficult to objectively measure parallelism above the level of the gene at a genome-wide scale, especially given the large number of genes of unknown function in the P. aeruginosa genome, and we therefore, focused our analysis of parallelism at the level of genes.
Like many free-living bacteria, the genome of P. aeruginosa is made up almost entirely of protein coding sequences (89.4% coding DNA). Because innovation involves the origin of novel phenotypes, it is reasonable to expect that innovation should be associated with more radical changes to proteins. The vast majority of mutations that we observed were non-synonymous substitutions in protein coding regions (S1 Table), but the relative frequency of radical amino acid substitutions did not differ significantly (Z-test; P = 0.84) between evolutionary innovation (n = 21; 53.8%) and optimization (n = 28; 56.8%). Short insertions and deletions (indels) that introduce frameshifts can also produce radical changes to proteins. However, we only found 6 indels that introduced frameshifts, making it impossible to test for a difference in the frequency of indels observed under innovation (n = 4) and optimization (n = 2). In summary, innovation and optimization did not leave distinct signatures on the structure of proteins in our experiments.
To gain further insights into the mechanistic basis of adaptation, we compared the functional roles of genes carrying mutations in clones evolved through innovation and optimization. Changes in the regulation of gene expression have been proposed to play an important role in evolutionary innovation [16,27,28]. Mutations in regulatory genes were common, and in many cases these mutations could be clearly linked to metabolic traits that were under selection (S2 Table). For example, adaptation to L-serine repeatedly evolved by non-synonymous mutations in a transcription factor (PA2449) that regulates the expression of genes involved in serine metabolism [29]; similarly, acquiring the ability to metabolize L-Alanyl-Glycine repeatedly evolved by mutations in pdsR, a repressor of a di-peptide and amino acid transport operon. We found that the proportion of mutations in genes involved in transcription was greater in clones from populations that had to adapt through innovation as opposed to optimization, supporting the idea that altered gene expression is an important feature of innovation (Fig 2, S3 Table; P = 0.036, One-tailed Fisher’s Exact Test).
Intergenic mutations also have the potential to change gene expression, for example by altering transcription factor binding sites [16]. For example, evolution in both L-aspartic acid and L-glutamic acid resulted in parallel substitutions in the promoter region of a P. aeruginosa homolog (PA5479) of a Bacillus subtilis L-aspartate and L-glutamate transporter [30]. Similarly, one clone evolved in D-Serine and one clone evolved in Glycerol have, respectively, a SNP upstream D-Amino acid dehydrogenase (PA5304) and a SNP upstream a glyceraldehyde-3-phosphate dehydrogenase (PA2323). However, intergenic sequences make up only 10.6% of the P. aeruginosa genome, suggesting limited potential for adaptation by regulatory mutations in non-coding sequences. Consistent with this idea, we detected only a very small number of intergenic mutations in clones evolved through innovation (n = 4) and optimization (n = 6), making it impossible to rigorously test the role of intergenic mutations in innovation.
It is difficult to make a priori predictions regarding associations between other functional categories of genes and innovation, but we found that innovation was also preferentially associated with mutations in metabolic genes (Fig 2, S3 Table; P<0.01, One-tailed Fisher’s exact test), whereas optimization was associated with mutations in genes involved in cell processes and signalling (Fig 2; P<0.01, One-tailed Fisher’s exact test).
Recent work in experimental evolution has focused on understanding the detailed molecular mechanisms by which individual beneficial mutations increase fitness (e.g.:[15,18,25,31,32]), and this work has made an important contribution to a broader functional synthesis in evolutionary biology [33]. We found that 46% of all mutations occurred in genes that were only mutated on a single carbon source and 83.6% of the mutated genes were only mutated in one carbon source, suggesting that substrate-specific adaptation was a key driver of evolution in this experiment. In many cases, mutations in these genes can be putatively linked to the metabolism of the carbon source that populations evolved on, and this was particularly the case for genes involved in transcription and metabolism and among clones that had to adapt through innovation (S2 Table). At the same time, we also found mutations in a small fraction of genes (16.4%) across multiple carbon sources. As an extreme example we found a gene (PA1561) involved in aerotaxis, mutated 16 times across 10 substrates, suggesting that mutations in this gene represent a general adaptation to Biolog plates. Unfortunately, it is impossible to precisely measure the substrate specificity of the mutations that we detected without carrying allelic replacement experiments to generate strains carrying single mutations. In S2 Table we provide a list of the mutations that occurred on each carbon source and their putative role. However, rigorously determining the biochemical basis of the fitness advantages conferred by individual mutations is outside the scope of this article, as our goal is to understand the genetic mechanisms of evolutionary innovation, and not the biochemical basis of novel metabolic pathways. Moreover, achieving a detailed functional understanding of adaptation in this system would be incredibly challenging given the diversity of selective pressures that we imposed and the diversity of mutations that we observed.
Gene duplication is a major source of evolutionary innovation [6,34], and some elegant studies show that it can facilitate adaptation in bacterial populations [35,36]. We detected six cases of de novo gene duplication. Every case involved parallel duplications, suggesting that duplication was adaptive. Strikingly, all four clones that adapted through innovation on glycyl-L-glutamic acid evolved independent duplications of a 5.6 Kb region that contains an operon (PA4496-PA4500) involved in di-peptide and amino acid transport [37] (S4A Fig). Using information on the frequency of SNPs in the sequenced clones, we were able to re-construct the evolutionary history of these duplications. Adaptation to glycyl-L-glutamic acid evolved via a repeatable two-step process. The first is a missense or nonsense mutation in the repressor of the operon, psdR (PA4499). The second is a tandem duplication of the operon, most likely as a result of homologous recombination between the flanking sequences of the operon (S4B Fig). The inactivation of the repressor plus the duplication of the operon probably results in increased expression of this operon. We were able to infer the chronology of this adaptation because all reads supported the novel mutations in the pdsR gene, as we would expect if duplication followed mutation. This multi-step process of potentiating mutations that alter the regulation of an operon, followed by adaptive gene amplification, is very similar to a previously described mechanism for the evolution of citrate utilization in Escherichia coli [15]. We also found large (> 300 genes, >5% of genome) duplications in two of the four clones that adapted through optimization on hydroxyl-L-proline. These duplications overlapped in a large region comprising most of their genes (262 genes, ≈288 Kb). This overlap suggests that the duplications were adaptive, but their large scale makes it difficult to infer exactly why. Overall, the limited incidence of duplication in clones that adapted through either innovation (n = 4) or optimization (n = 2) suggests that de novo duplication is not frequently involved in metabolic innovation. This result is consistent with recent work showing that de novo duplication makes only a minor contribution to adaptation to gene loss in E. coli [16] and yeast [38].
In addition to the origin of novel duplicate genes, the divergence of already existing duplicate genes in the P. aeruginosa genome can also play a key role in evolutionary innovation [39]. To test its importance in our experiment, we classified P. aeruginosa genes into duplicates and singletons using a clustering method based on Blast similarity searches. Sequence similarity in bacterial genomes can arise as a consequence of gene duplication of existing genes in the genome, which produces paralogs that are similar as a result of shared ancestry. Alternatively, bacteria can acquire new genes that are similar to existing genes in the genome by horizontal gene transfer. In practice it is very difficult to distinguish between these two mechanisms for the origin of novel genes, and Lerat and colleagues have proposed the term synologs to describe homologous genes in bacterial genomes [40].
We found that clones that adapted through innovation acquired more mutations in existing duplicate genes than expected due to chance alone based on the frequency of duplicate genes in the P. aeruginosa genome (Fig 3, P<0.01, Pearson’s X2 test). In contrast, the frequency of mutations in duplicate genes in clones that adapted through optimization was indistinguishable from the frequency of duplicates in the P. aeruginosa genome (S4 Table; P>0.05, Pearson’s X2 test). We repeated this analysis using a broad range of similarity cut-offs to identify duplicate genes (see Materials and Methods). Our results remained robust, we consistently detected an enrichment of mutations in duplicate genes in clones that adapted through innovation, irrespective of the cut-offs used to identify duplicates (S5 Fig and S4 Table). This result suggests that the divergence of existing duplicates plays an important role in the ability to evolve novel metabolic phenotypes. We re-emphasize, however, that this analysis does not distinguish between duplicate genes that arose due to horizontal gene transfer and spontaneous duplication.
What constrains the evolution of metabolic innovations that could allow P. aeruginosa to expand its ecological niche? One possible answer is that fitness costs associated with novel metabolic traits may impose a trade-off that limits metabolic innovation [41–43]. To test this hypothesis, we measured the growth of 2 of the sequenced clones from each carbon source across the 94 alternative carbon sources present on a Biolog plate, and we compared it to the growth of the ancestral clone on each carbon source. We did a total of 4750 growth assays (S6 Fig) and we established conservative criteria to infer positive or negative pleiotropy. Because our evolved clones carried only a small number of beneficial mutations (1.70 mutations per clone on average), we can be confident that altered growth on alternative carbon sources reflects the pleiotropic side-effects of beneficial mutations. However, we cannot entirely rule out the possibility that some clones carried conditionally neutral mutations that spread by hitch-hiking with beneficial mutations, but the scarce number of synonymous mutations suggests that conditionally neutral mutations are infrequent.
Adaptation was associated with pleiotropic costs, because evolved clones showed reduced growth on an average of 14.76 carbon sources that could be used by the ancestral clone. However, the pleiotropic cost of innovation was 70% greater than the pleiotropic cost of optimization, which is consistent with the idea that pleiotropy constrains innovation (Fig 4, S5 Table; P<0.01, Pearson’s X2 test). The precise mechanistic causes of negative pleiotropy are difficult to determine [44] without measuring the effects of the individual mutations that contributed to adaptation in our system. However, the association between evolutionary innovation and mutations in regulatory and metabolic genes suggests that mutations in both of these categories of genes are likely candidates to explain negative pleiotropy. While these observations show that both innovation and optimization have costs, not all pleiotropy may be negative. Surprisingly, we found that positive pleiotropy—where an evolved population shows increased growth on one or more alternative carbon sources—was just as common as negative pleiotropy. The frequency of positive pleiotropy did not differ between clones that adapted through innovation and optimization (Fig 4, S5 Table; P = 0.61, Pearson's X2 test).
Clones that adapted through innovation were enriched in mutations in duplicated genes and paid higher pleiotropic effects than clones that adapted through optimization. This observation is counter-intuitive, because we would expect that mutations in existing gene duplicates should be associated with low pleiotropic costs, given that the other copy of the duplicate may provide functional backup for the mutated copy. To explore this counter-intuitive observation further, we compared the pleiotropic costs expressed by clones carrying mutations in duplicates genes that have close or distant homologs in the P. aeruginosa genome. This analysis is motivated by the assumption that functional redundancy between duplicate genes decays as they diverge from each other. Interestingly, we found that clones carrying mutations in genes that have close homologs have a lower pleiotropic cost than clones without mutations in genes with close homologs (Table 1, S6 Table; P<0.01, Pearson's X2 test). In contrast, we see the opposite pattern in distant homologs: The pleiotropic cost of clones with mutations in genes with distant homologs is higher than that of clones without mutations in distant homologs (Table 1, S6 Table; P<0.01, Pearson's X2 test). Collectively, these results support the idea that redundancy between duplicates minimizes the cost of innovation.
Microbiologists have known for a long time that bacteria can evolve novel metabolic traits in the laboratory [45], and we have taken advantage of the experimental tractability of microbial metabolism to study evolutionary innovation at a broad scale using high-throughput experimental methods coupled to whole genome re-sequencing. This approach provides the opportunity to study the generality of evolutionary outcomes under a range of selective conditions [16,38]. Using this approach, we have shown that there are significant differences in the genomic basis of metabolic innovation and optimization in P. aeruginosa. Opportunistic pathogens, such as P. aeruginosa, encounter a novel niche when they establish long-term infections in human hosts, and altered metabolism plays a role in evolutionary transition to specialization on a pathogenic lifestyle [46]. Understanding the causes of evolutionary innovation may, therefore, contribute to our ability to predict the evolution of host-specialization in pathogenic bacteria.
At a functional level, we found that both innovation and optimization are predominantly driven by substitutions in proteins, which is hardly surprising given that the genome of P. aeruginosa is made up of 90% coding DNA. Interestingly, innovation and optimization leave similar signatures in proteins, and we did not find any evidence of an excess of radical substitutions associated with innovation. In contrast, we found profound changes in the functional roles of genes that contributed to innovation and optimization. Specifically, we found that innovation is associated with mutations in transcription regulators and metabolic genes. Changes in the expression of existing metabolic pathways that have a basal or underground ability to metabolize novel compounds and changes in the structure of metabolic enzymes that increase their activity towards novel substrates could be involved in the origin of innovations. Importantly, previous studies have provided detailed examples of how both of these mechanisms can lead to evolutionary innovation in bacteria [15,45,47–49].
One of the main results of our study is that mutations in pre-existing duplicate genes in the P. aeruginosa genome play an important role in metabolic innovation, but not optimization. It is important to recall that we identified duplicate genes based on sequence similarity, and not necessarily common ancestry. Importantly, this method does not distinguish between duplicates that arise via spontaneous duplication (paralogs) and horizontal gene transfer, but irrespective of the origins of the duplicates, duplication is expected to result in genetic and functional redundancy [50]. Why are duplicates so important for innovation? Our results show that evolving new metabolic traits is associated with pleiotropic costs. This is not surprising given that innovation is associated with mutations in genes involved in transcription and metabolism. Trade-offs between evolving novel metabolic pathways and maintain existing ones may therefore constrain innovation. How can this obstacle be overcome? Carrying duplicate genes produces redundancy, and this redundancy can potentiate innovation through neo-functionalization [4–6,51]. The presence of an extra gene copy with functional overlap increases mutational robustness and this increase facilitates the exploration of novel gene functions while the other copy maintains its ancestral function [52–54]. The importance of gene duplication for mutational robustness is still debated [54–60]. Our results support its importance. We find that mutations in genes that have a close homolog in the genome tend to be associated with lower costs than mutations in duplicate genes that have distant homologs. Therefore, our experiments provide evidences of a link between duplication, robustness, and evolvability in P. aeruginosa.
In contrast to eukaryotes, most new genes in γ-proteobacteria, including P. aeruginosa, are acquired by horizontal gene transfer, and not by gene duplication [40]. For example, we identified approximately 10% of genes in the P. aeruginosa genome as being pre-existing duplicates, whereas Lerat and colleagues [40] estimated that only about 1% of genes in γ-proteobacteria genomes arise by duplication. This discrepancy suggests that pre-existing horizontally acquired genes are likely to have played a key role in evolutionary innovation in our experiment. Horizontal gene transfer has mainly been viewed as an important source of evolutionary innovation by providing bacteria with access to a very wide pool of genes that confer novel and important phenotypes, such as antibiotic resistance in pathogenic bacteria [8,61,62]. Our results suggest that the horizontal acquisition of functionally redundant genes may also play a key role in evolutionary innovation by providing bacteria with increased genetic robustness to mutations that generate novel phenotypes.
Although redundant duplicate genes provide a genetic substrate for innovation, it is well established that acquiring new genes as a result of horizontal gene transfer or gene duplication, carries a fitness cost in bacteria [31,36,63–66]. Owing to this cost, newly acquired genes tend to be lost from bacterial populations unless gene acquisition, per se, is beneficial [6,36] or because addiction genes, such as toxin-antitoxin systems, select against the loss of acquired genes [67]. Indeed, fitness costs may explain why we observed so few instances of de novo duplication. Selection against newly acquired redundant genes in the short term may therefore limit the long-term ability of bacteria to evolve novel phenotypes. A key goal for future work will be to understand how this tension between fitness and evolvability arising from gene acquisition is resolved.
In this study we used two bacterial clones, P. aeruginosa PAO1 (PAO1-wt) and a P. aeruginosa PAO1 containing the luxCDABE operon (PAO1-lux) for luminescence production inserted in a neutral site in the bacterial chromosome (PAO1::mini-Tn7-pLAC-lux) [68]. The two clones are genetically identical except for the lux operon. We cultured the strains in Biolog GN2 96-well plates (Biolog, USA), in a final volume of 125 μl of M9 broth (Fischer Scientific, USA) per well, at 37°C and without shaking. Biolog GN2 plates contain 95 different carbon sources plus a negative control well.
We first cultured bacteria on LB agar plates (Fischer Scientific, USA) at 37°C to obtain isolated colonies. We initiated the evolution experiment by inoculating a single colony in each well of a 96 well-plate containing LB broth. Alternating wells contained PAO1-wt and PAO1-lux clones to control for possible cross-contamination occurring during the experiment. We build two plates with two alternative patterns: well A1 containing PAO1-wt (plate A) and well A1 containing PAO1-lux (plate B). We incubated the 96 well-plate at 37°C and 225 rpm overnight and we started the experimental evolution with two replicates from each plate (plate A1, plate A2, plate B1, plate B2).
We propagated these four plates independently in Biolog GN2 plates for 30 transfers. Specifically, we transferred 5μl of bacterial culture every day into a fresh Biolog plate with 120 μl M9 broth in each well, allowing approximately 4.6 generations per day (log2 of the dilution factor) and a total of 140 generations during the experiment. Note that this calculation estimates equal bacterial density after incubation in a given well (carbon source) over the entire experiment. An increase in the maximum cell density in a well over the 30 days due to adaptation to the carbon source will slightly increase the total number of generations (in our experiment, the most extreme example went up to 143 generations). These differences are not big enough to significantly affect the mutation supply in the different environments over the experiment. We assessed OD and luminescence every day using a BioTek Synergy H4 plate reader (Biotek Instruments, UK). We incubated the plates at 37°C for 24 hours with no shaking.
To assess how well adapted the ancestral strain was to each of the carbon sources in a Biolog GN2 plate, we estimated the number of viable cells in each well after a 16 hour incubation of the parental PAO1 strain. To this end, we first streaked out freezer stocks of wild type ancestors to isolate single colonies, and cultured a single colony in 3 ml LB tubes overnight at 37°C with continuous shaking (225 rpm). On the next day we diluted 5 μl of the overnight culture into 20 ml of M9 broth, and inoculated a Biolog GN2 plate with 125 μl of the dilute cell suspension per well. We incubated the culture for 16 hours at 37°C without shaking, diluted it ten-fold, and measured cell viability using the BacTiter-Glo Microbial Cell Viability Assay (Promega, USA). We assessed luminescence produced in this assay using a BioTek Synergy H4 plate reader (BioTek Instruments, UK) for a total of 16 technical replicates per carbon source. We calculated the final bacterial density in each well with an in-house R script. The numbers of viable cells in the different wells followed a clear bimodal distribution (Fig 1A). We fitted a mixture distribution to the data using the mixtool package in R [69]. We used the point at which the two distributions intersected (107.84 bacteria/mL) to classify Biolog carbon sources into two groups, one in which the ancestor grew very poorly (and required innovation to adapt during the evolution experiment), and another where the ancestor grew well (and required only optimization) (Fig 1A, S1 Fig). To strengthen our classification we also used OD data obtained before performing the BacTiter-Glo Microbial Cell Viability Assay. We performed a hierarchical clustering analysis using viable cells and OD data (S7 Table) using the hclust package in R [70]. Three clear groups clustered together (S7 Fig): carbon sources classified as innovation using the intersection of the two distributions, carbon sources classified as optimization using the distribution intersection and a last group containing three carbon sources that according to the distribution intersection were classified as optimization (N-acetyl-D Glucosamine, Sebacic Acid and Hydroxy-L-Proline). These three carbon sources were clustered together because they have lower OD values than the other carbon sources classified as optimization. We decided to maintain our classification and categorize them as optimization because the BacTiter-Glo Microbial Cell Viability Assay has a better resolution than OD. However, the main results of the paper did not change if we classify these three carbon sources as innovation.
To identify which populations had adapted after 30 days of evolution we measured growth curves both for ancestor and final populations at the population level. We used growth rates to determine differences in the fitness of these populations because the growth rate provides a higher level of resolution than final cell density in the population. This higher resolution is needed because we are now comparing the changes in fitness in a given carbon source over the 30 days of experiment, which are subtle (compared to the differences observed for the parental strain over the 95 different carbon sources). To this end, we cultured freezer stocks of both populations in LB broth (200 μl/well, 96-well plates) overnight at 37°C, 225 rpm. The next day we diluted the bacteria 1000-fold in M9 broth, and cultured them overnight in Biolog GN2 plates at 37°C (125 μl/well). The following day we performed another passage, diluted the culture 1:1000 in fresh Biolog G2 plates using fresh M9 broth (final volume, 125 μl/well), and measured growth rate for 16 hours using a BioTek Synergy H4 plate reader (BioTek Instruments, UK) at 37°C with no shaking. We computed the maximum growth rate with the software Gen5 2.00 (BioTek Instruments, UK). Then, from the populations in which we observed an increase in growth rate, we isolated a single clone from the freezer stock using LB agar plates (Fischer Scientific, USA) and repeated the same protocol to ensure that we found the same behaviour at clone level. We stored frozen stocks of these clones. Subsequently, we sequenced the genomes of selected clones, i.e., clones evolved in carbon sources that fulfil the following condition: The maximum growth rate for the final population (and clone) was higher than that of the ancestral population for all four replicates of populations evolved in that carbon sources.
We performed DNA extractions from clones cultured in 3 ml LB broth (Fischer Scientific, USA) that had been incubated at 37°C with 225 rpm shaking overnight, using the Qiagen Dneasy Blood and Tissue Kit (Qiagen, Inc., Chatworth, California, USA) and the Promega Wizard Genomic 4 DNA Purification Kit (Promega, UK). We quantified DNA using the QuantiFluor dsDNA system (Promega, Madison, WI, USA) following manufacturers' instructions.
We conducted library preparation and sequencing (using HiSeq2000 and 100-bp-paired end reads) at the Wellcome Trust Centre for Human Genetics, University of Oxford. We sequenced 88 genomes, i.e., 2 PAO1-lux ancestral strains, 2 PAO1-wt ancestral strains, and 84 evolved clones. We analyzed sequencing data using a pipeline developed in-house, as previously described in San Millan et al. 2014 [71], and mapped filtered reads to our reference genome, which is P. aeruginosa PAO1 (NC_002516.2) with the insertion of the phage RGP42 (GQ141978.1). We analyzed only those mutations that had accumulated during the experiment and that were not present in our ancestral strains at the start of the experiment. Note that each evolved clone was compared to the specific ancestral clone from which it was derived (i.e. either PAO1-wt or PAO1-lux).
The reads generated in this work have been deposited in the European Nucleotide Archive database under the accession code PRJEB12874.
To elucidate how frequently adaptation to a specific carbon source affects growth in other carbon sources, we performed an experiment at the level of individual clones, using two evolved clones per carbon source and four wt clones as controls. Specifically, we tested those clones whose growth rate had increased during the experiments (the same clones that we used for the whole genome sequencing). We inoculated each clone that had adapted to a particular carbon source in the 95 carbon sources of a Biolog plate. To this end, we cultured each frozen clone and four wt clones in 3 ml LB tubes overnight at 37°C with continuous shaking (225 rpm), diluted 5 μl of the overnight culture on the next day into 20 ml of M9 broth, and inoculated a Biolog GN2 plate with 125 μl of the dilute cell suspension per well. We incubated the plate for 16 hours at 37°C without shaking. Subsequently, we diluted each culture 10-fold, and measured cell viability in 384-well black plates, using the BacTiter-Glo Microbial Cell Viability Assay (Promega, UK), following manufacturer's instructions. We assessed luminescence produced in the BacTiter-Glo assay using a FLUOstar OPTIMA plate reader (BMG Labtech, UK).
We developed an R script to calculate the number of doublings bacteria experience in each well of the Biolog plate correcting for the number of doublings in the negative control well (number of effective doublings). We used the values obtained from 4 wt replicate controls to calculate the 95% confidence interval of number of effective doublings for each well. Then, to assess the pleiotropic effects that the adaptation to a particular carbon source had in the 94 remaining carbon sources, we checked if the number of doublings in each carbon source for that particular clone fell outside the 95% confidence interval calculated using the wt measurements. If it fell outside the 95% confidence interval we counted it as a positive pleiotropic effect in that specific carbon source if the number of doublings were higher than for the wt or as negative pleiotropic effect if the number of doublings were lower than for the wt.
We classified those mutations that involve amino acid replacements as radical if they were associated with a change of polarity group (polar: C, N, Q, S, T and Y; nonpolar: A, F, G, I, L, M, P, V and M; positively charged: H, K, and R; and negatively charged: D and E) and as nonradical when the replacement did not imply a change of polarity group.
To classify P. aeruginosa PAO1 genes into duplicates and singletons, we used BLASTclust (ftp://ftp.ncbi.nih.gov/blast/documents/blastclust.html). BLASTClust is a stand-alone program used to cluster proteins based on pairwise matches using the BLAST algorithm. We considered as singletons all proteins that formed a cluster whose only member was the protein itself, and as duplicates when the cluster contained more than one protein [72,73]. Note that this method does not distinguish between duplicates originated by gene duplication of existing genes in the genome and horizontal gene transfer. To ensure the robustness of the classification we used 10 different cut-offs of minimum length coverage and percentage of identical residues. The results remained robust to the different cut-offs used (S4 Table). Higher values of length coverage and residue identity are more likely to indicate close homology [74]. Moreover, it has been previously suggested that values of 53% coverage and 31% identity are enough to detect homology in duplicates [73,75]. For these reasons and also to ensure having enough number of genes classified in each category, we used the cut-off 70 coverage and 50 identity as an indicator of close homology and the cut-off 50 coverage and 40 identity as an indicator of distant homology.
We experimentally validated observed duplications after 30 days of evolution in glycyl-L-glutamic acid and hydroxyl-L-proline by PCR-amplifying the edges of the duplication and Sanger sequencing the products to confirm the results (S8 Table and S4A Fig).
We assessed statistically if the number of parallel mutations observed in our dataset is higher than the expected by chance. We randomly selected a total of 143 positions in the coding region of the P. aeruginosa PAO1 genome, corresponding to the total number of identified mutations. We then recorded the proportion of loci located in genes with more than one randomly selected position and repeated the procedure 1,000 times. As a result, we generated the expected distribution by chance (mean = 0.04, sd = 0.02). We obtained an empirical estimation of the P-value as the proportion of permutations yielding a value more extreme than the observed in our dataset.
We performed all statistical analyses and produced all graphics using R [61].
|
10.1371/journal.pbio.1001494 | Hedgehog Signaling Acts with the Temporal Cascade to Promote Neuroblast Cell Cycle Exit | In Drosophila postembryonic neuroblasts, transition in gene expression programs of a cascade of transcription factors (also known as the temporal series) acts together with the asymmetric division machinery to generate diverse neurons with distinct identities and regulate the end of neuroblast proliferation. However, the underlying mechanism of how this “temporal series” acts during development remains unclear. Here, we show that Hh signaling in the postembryonic brain is temporally regulated; excess (earlier onset of) Hh signaling causes premature neuroblast cell cycle exit and under-proliferation, whereas loss of Hh signaling causes delayed cell cycle exit and excess proliferation. Moreover, the Hh pathway functions downstream of Castor but upstream of Grainyhead, two components of the temporal series, to schedule neuroblast cell cycle exit. Interestingly, hh is likely a target of Castor. Hence, Hh signaling provides a link between the temporal series and the asymmetric division machinery in scheduling the end of neurogenesis.
| In almost all metazoans, neurons are produced by a group of neural stem cells/progenitors in a precise temporal manner, which is important for generating a functional nervous system. In Drosophila, this “timing” mechanism is mainly governed by the sequential switching of transcription factors in neural stem cells called neuroblasts, such that neuronal fate is associated with its birth order. These temporal factors also coordinate the termination of neuroblast division towards the end of neurogenesis. In this study, we show that Hedgehog (Hh) signaling also regulates the division rate of neuroblasts during their proliferative phase at larval stage, as well as the cessation of proliferation at early pupal stage. Excessive Hh signaling causes premature neuroblast cell cycle exit and early termination of neurogenesis, while loss of Hh signaling results in prolonged proliferation of neuroblasts beyond its physiological window. We also find that Hh signaling acts in concert with the temporal transcription factors, and is itself regulated by these factors. We hypothesize that this mode of interaction (temporal transcription factors with developmentally regulated signals like Hh) during neurogenesis could be widely conserved in other organisms.
| Both intrinsic and extrinsic mechanisms are deployed during development to generate cellular diversity [1]. Extrinsic mechanisms involve cell-cell communication, while intrinsic mechanisms ensure preferential segregation of cell fate determinants into one of the two daughter cells upon completion of cell division. The latter is well exemplified during Drosophila neurogenesis [2]–[5]. Drosophila embryonic neuroblasts (NBs) delaminate from the neuroepithelium and these neural stem cells undergo repeated self-renewing asymmetric divisions. Each division generates a larger daughter that retains NB identity and a smaller daughter, ganglion mother cell (GMC), that normally divides one more time to produce two neurons/glia depending on lineage specificity. At the end of embryogenesis, most NBs enter a proliferative quiescent stage and subsequently resume mitotic activity during early larval stages. These larval NBs, like their embryonic counterparts, undergo extensively repeated divisions to self-renew and at the same time produce postmitotic neurons/glia to build a functional nervous system [6],[7].
During NB divisions, cell fate determinants including Numb, Prospero (Pros), and Brain tumor (Brat) are asymmetrically localised onto one side of the NB cortex (referred to as the basal cortex) via two coiled-coil adaptor proteins, Partner of Numb (Pon, the adaptor for Numb) and Miranda (Mira, the adaptor for Pros and Brat), and are subsequently segregated into the small GMC daughter at the end of NB divisions [8]–[18]. The basal localization and segregation of these cell fate determinants into GMCs are controlled by two evolutionarily conserved protein complexes: the Bazooka (Baz, the fly Par-3 homolog)/DmPar6/DaPKC (atypical protein kinase C) complex and the Partner of Inscuteable (Pins)/Gαi complex, which localize on the opposite side of the cortex (referred to as the apical side) and are bridged together to form a larger protein complex via Inscuteable (Insc) [19]–[27]. Pros is a homeodomain-containing transcriptional factor and acts as a binary switch between self-renewal and differentiation during neurogenesis [28]. It suppresses genes required for NB self-renewal; but its activity is also required to activate genes necessary for GMC differentiation. Hence mis-expression of Pros in the NBs leads to their loss via precocious differentiation [29],[30], while in the absence of Pros, GMCs fail to differentiate, express NB markers, and exhibit increased proliferation [28]. Thus, an imperative task of NB asymmetric division is to segregate Pros exclusively into GMCs. In embryonic NBs, Pros and Mira are transiently localized onto the apical cortex during late interphase and early prophase prior to their basal localizations. The localization of Pros and Mira is initiated by the DaPKC-mediated direct phosphorylation on Mira, which results in the displacement of Mira from the apical cortex and subsequently, via an unidentified mechanism, localize onto the basal cortex [31]. Recently, the conserved protein phosphatase (protein phosphatase 4 [PP4]) complex was identified as an essential mediator for the localization of Pros and Mira during both interphase and mitosis [32]. In the absence of PP4 activity, Pros and Mira are mislocalized to the nucleus during interphase and cytoplasm during mitosis. Consistent with a role of Pros in suppressing NB self-renewal genes, PP4 mutant NBs exhibit reduced proliferation.
Repeated segregation of the same sets of cell fate determinants does not fully explain how extensive cellular diversity can be generated from NB lineages. The generation of diverse progeny from a single NB is also regulated by another NB intrinsic mechanism such that each NB will undergo a specific number of divisions in a defined temporal and spatial context to generate a lineage with distinct neuronal or glial fates [33],[34]. During embryonic neurogenesis, this “timing” mechanism (or temporal series/mechanism) is controlled by sequential expression of a series of transcription factors in the NBs: Hunchback (Hb)→Krupple (Kr)→POU homeodomain protein 1/2 (Pdm)→Castor (Cas)→Grainyhead (Grh), although some NB lineages only express a subset of this series [35]–[37]. Grh is the last transcription factor expressed in embryonic NBs and its expression persists in the postembryonic NBs throughout the larval stage, presumably to maintain mitotic activity of the NB [38]–[41]. The linearity and robustness of the temporal series involves an intricate network of transcriptional regulation encompassing additional players, such as Seven-up (Svp) [35],[37],[42],[43]. Temporal series continues during the larval stage with the redeployment of embryonic temporal regulators Cas and Svp to achieve two major transitions in NB lineages: (1) the neuronal identity switch from larger Chinmo+Br-C− early-born neurons to smaller Chinmo−Br-C+ late-born neurons at L2 stage, and (2) termination of NB proliferation (cell cycle exit) at pupal stage, which is concomitant with cytoplasmic localization of Mira and a burst of nuclear Pros during early mitosis [44]. Furthermore, an early burst of Pros is sufficient to trigger cell cycle exit in type I NBs and cessation of neurogenesis in larvae. Therefore nuclear Pros may act as the physiological means for promoting NB cell cycle exit and cessation of neurogenesis [44]. However Cas is transiently expressed during early larval development [44]. It is unclear how this transient Cas expression acts to trigger a later burst of nuclear Pros in pupae to promote cell cycle exit. It is also unclear how the temporal mechanism is coupled with the asymmetric division mechanism to generate a functional nervous system.
The Hedgehog (Hh) pathway is repeatedly deployed throughout animal development to mediate diverse functions [45]. The core machinery of Hh signalling is evolutionarily conserved from flies to humans, although there is clear divergence in its mechanistic details [46]. In general, Hh ligands are synthesized as precursor molecules that undergo autocatalytic cleavage to yield an N-terminal fragment with a cholesterol moiety tethered to its C-terminal end before its secretion from producing cells. The receptor for Hh is a 12-pass transmembrane protein, Patched (Ptc), which, in the absence of Hh, inhibits the activity of a second downstream effector molecule, Smoothened (Smo). Smo is a seven-pass transmembrane protein that bears resemblance to the mammalian G-protein coupled receptor (GPCR) [47],[48]. The ultimate effector of Hh signalling pathway is the transcription factor Cubitus interruptus (Ci), which can act as a full length transcriptional activator, Ci-155, or a proteolytically cleaved transcriptional repressor, Ci-75 [49],[50]. In the absence of Hh ligand, Smo activity is suppressed by Ptc and, consequently, Ci is phosphorylated and is processed into the repressor form, Ci-75. The binding of Hh to its receptor inhibits Ptc and alleviates its inhibition of Smo, resulting in the stabilization and phosphorylation of Smo C-terminal tail [51]–[53], and subsequently stabilization of the Ci activator form, Ci-155. In Drosophila embryos, Hh signaling is implicated in the specification of a subset of NBs in a spatial pattern. Furthermore, the Hh pathway also functions to reactivate NBs from their quiescent stage during early larval stages [54]. However, its role, if any, after NB reactivation is unknown.
In this study, we investigate the function of Hh signaling during the development of Drosophila postembryonic brain. We show that hh expression is temporally regulated in the postembryonic larval brain and Hh signaling promotes NB cell cycle exit in the early pupae, possibly by mediating nuclear localization of Pros. Earlier (excess) activation of Hh signaling results in defective Pros localization, and leads to under-proliferation and premature cell cycle exit, whereas loss of Hh signaling causes delayed NB cell cycle exit and excess proliferation. Hh expression in postembryonic larval brain depends on an earlier pulse of expression of a component of the NB temporal series, Cas. Hh signaling in NBs in turn shuts down the expression of Grh, the terminal component of the NB temporal series required for the mitotic activity of larval NBs. Hence the timely exit of NBs from the cell cycle depends on the intricate interplay between Hh signaling, components of the NB temporal series (cas, grh) as well as Pros, a component of the NB asymmetric division machinery.
Using the mosaic analysis of a repressible marker (MARCM) system [55] to screen for potential signaling pathways required for asymmetric division of type I NBs in the central brain, we found that components of the Hh signaling pathway were involved in regulating aspects of asymmetric division, as well as the proliferative capacity of NBs. To examine the effects of compromised Hh signaling in the central brain, we generated homozygous clones of a mutant allele of smo, smoIA3, which has a substitution mutation in the Cys rich domain of the extracellular N-terminal domain and fails to transduce downstream signaling [56]. Conversely, hh gain-of-function clones were produced using a loss-of-function allele, ptcS2, which fails to repress Smo function and results in constitutive Hh signal activation, even in the absence of the Hh ligand [57]. In wild-type (wt) animals, Baz/Par6/aPKC and Insc/Pins/Gαi proteins form a complex and localize on the apical cortex of dividing NBs (Figure S1A–S1C), and these apical proteins exhibited largely normal localization in ptcS2 and smoIA3 mutant NBs (Figure S1E–S1G, S1I–S1K, and unpublished data). Numb and its adaptor Pon were correctly localized on the basal cortex in all wt, ptcS2, and smoIA3 mutant NBs (Figure S1D, S1H, S1L, and unpublished data). While Mira/Pros complex was localized to the basal cortex in both wt and smoIA3 mutant NBs (Figures 1D, 1F, 2D, 2F, S1C, S1K); however, both Mira and Pros were largely delocalized and showed cytoplasmic accumulation in ptcS2 mutant NBs (Figures 1E, 2E, S1G). Similar Mira delocalization defects were seen in another ptc allele, ptc13 (Figure S1M). Thus, removing smo function from NBs does not cause any noticeable defects on asymmetric division, whereas ptcS2 mutant NBs specifically disrupt the basal localization of the Mira/Pros but not Pon/Numb complex during mitosis.
In addition to these asymmetry defects, we also observed that both ptcS2 and smoIA3 mutant NBs exhibited defective proliferation compared to wt NBs. At 96 h after larval hatching (ALH), a typical wt type I NB clone induced at 24 h ALH (soon after NB reactivation) contained 40.1±11.7 cells (n = 19). However, ptcS2 clones were smaller than their wt counterparts and contained 26.9±9.8 cells (n = 29, p = 0.0001), whereas smoIA3 produced noticeably larger clones with 50.4±20.5 cells (n = 24, p = 0.0283) (Figures 1A–1F′ and 2O). Moreover, most of the ptcS2 NBs appeared to be arrested in interphase based on the appearance of diffused DNA with markedly reduced cortical Mira (93.3%, n = 45), whereas the surrounding heterozygous interphase NBs displayed strong cortical Mira (Figure 1B and 1G). To further confirm that the clone size difference in ptcS2 and smoIA3 mutants was a consequence of alteration in NB proliferative capacity, we measured the mitotic index using phospho-histone H3 (PH3) as a mitotic marker. Indeed, ptcS2 NBs were significantly less mitotically active than their wt counterparts, whereas a higher proportion of smoIA3 NBs were engaged in mitosis at all time points examined (Figure 1H). We also noted that the increase in mitotic index in smoIA3 NBs was most likely associated with inactivation of Hh signaling as NB clones mutant for the hh null allele, hhAC, were equally mitotically active (Figure 1H). This phenotype, together with an enlarged hhAC clone size (49.4±19.3 cells; n = 33, p = 0.0177; Figure S4A–S4C), indicated that Hh functions in a lineage confined manner to restrict the proliferation of the NBs in the central brain (also see Discussion).
One drawback of using PH3 as a proliferative marker is that its index does not distinguish between alteration in the cell cycle time and proliferative capability of the marked cells. Hence, we conducted a 5-bromo-2′-deoxyuridine (BrdU) labeling assay to investigate the propensity of the NBs to undergo cell cycle progression at around 90 h ALH. By allowing 4 h of BrdU nucleoside analogue incorporation, we found that all the wt NBs and 9.0±2.7 of progeny (n = 33) had BrdU labeling in the nuclei (Figure 1I). In contrast, about half of the ptcS2 mutant NBs (55.0%, n = 34) were devoid of BrdU labeling while most of the rest that managed to incorporate BrdU showed fairly weak signal (Figure 1J). In addition, significantly fewer progeny of ptcS2 mutant NBs (2.7±2.5, n = 34, p<0.0001) had BrdU labeling compared to their wt counterparts. This observation confirms that ptcS2 mutant NBs were not actively proliferating. Conversely, all smoIA3 clones comprised a single BrdU-labeled NB, along with an increased number of progeny that harbored nuclei BrdU (11.4±2.5, n = 41, p = 0.0001) (Figure 1K), consistent with a higher mitotic index of smoIA3 NBs. Collectively, these data showed that ectopic Hh signal activation resulted in reduction in NB proliferation, while defect in Hh signaling increased the proliferative capability/rate of the NBs.
In order to understand the mechanism that underlies changes in NB proliferation, we sought to examine whether there is any alteration to NB and neuronal fates due to the loss and gain of Hh signaling using molecular markers Deadpan (Dpn), Pros, and Elav. In wt, all the NB clones examined contained a single large NB that expresses nuclear Dpn (Figure 2A). The newborn GMC was transiently Dpn-positive due to perdurance of the protein after division. In contrast, Dpn expression was severely reduced in ptcS2 NBs (Figure 2B), while smoIA3 NBs continued to express Dpn as normal (Figure 2C). smoIA3 NB clones, despite having a larger clone size and higher mitotic index than wt, did not exhibit supernumerary NB-like cells as seen in pros or brat mutants [10]. In wt and smoIA3 NBs, Pros is under our detection threshold during interphase, but accumulates on the basal side of the NBs during mitosis, while in GMCs and neurons, Pros is nuclear (Figure 2D and 2F). Notably, ptcS2 NBs often exhibited nuclear Pros during interphase and cytoplasmic enrichment of Pros during mitosis (Figure 2E and unpublished data). Since it had been shown recently that Pros-dependent NB cell cycle exit in early pupae may depend on nuclear localization of Pros during interphase [44], our observations suggested that the reduced proliferation in ptcS2 clones may be an indication of premature NB cell cycle exit. Indeed, ectopic pros expression during larval stage promoted premature NB cell cycle exit with concomitant down-regulation of NB marker Dpn and delocalization of Mira as observed in ptc mutant NBs (Figure S2A and S2B).
To understand if GMC division and neuronal differentiation were affected, we examined the neuronal marker Elav. In wt clones, all of the cells expressed Elav, except for the large Dpn-positive NB and three to five Dpn-negative GMCs (4.4±0.9 cells per clone, n = 23) adjacent to it (Figure 2G and 2O). In contrast, ptcS2 clones had very few Elav-negative GMC-like cells (1.8±1.2 cells per clone, n = 53) compared to the wt clones (Figure 2H and 2O). This is consistent with the ptcS2 NBs being less mitotically active. However, smoIA3 clones often contained four to 15 GMC-like cells (6.7±2.6 cells per clone, n = 25), which were Elav-negative and in direct contact with, or at close proximity to the NB (Figure 2I and 2O). There were two possible explanations for this GMC-like pool expansion: (1) an increase in NB division rate (which we confirmed with live imaging, see below) leading to accumulation of GMCs or; (2) some GMCs remained mitotically active after division and did not differentiate. Our data favored the first possibility as we failed to detect any GMC clones comprising of more than two cells (n>50), suggesting that the GMCs did not undergo extra divisions. Furthermore, differentiation was not arrested in smoIA3 clones as all the mutant cells expressed Elav when examined at adulthood (Figure S3A). These results showed that Hh signaling plays a key role in controlling NB proliferation but does not block differentiation in type I NB lineages. Expectantly, hhAC clones also contained ectopic GMC-like cells (8.1±1.9 cells, n = 12), again reinforcing the view that the diffusion of Hh ligand is restricted within a single lineage (Figure S4A and S4B, see Discussion).
While type I NBs generate GMCs that undergo terminal divisions to produce two neurons/glia, type II NBs generate trans-amplifying GMCs that undergo multiple rounds of asymmetric divisions to generate many neurons/glia [58]–[60]. Hh signaling appears to play a similar role in type II NBs. In wt, each type II NB clone contains 96.8±31.1 cells (n = 21; Figure 2O). ptcS2 clones were smaller and contained only 61.7±29.9 cells (n = 19; Figure 2O), while smoIA3 mutant exhibited larger clones with 138.6±48.6 cells (n = 16; Figure 2O). Together, these data suggest that Hh signaling is likely utilized as a common mechanism in regulating NB proliferation in the larval brain.
To assess whether Hh signaling exerts its effect on NBs via the canonical pathway, we examined NB proliferation in ci clones. When Ci function was compromised in ci94 mutant [61], NB clones exhibited a GMC-like pool expansion phenotype similar to smoIA3 clones (Figure 2J–2L), albeit at a lower frequency (20.0%, n = 15). Pros localization was largely unaffected in both NBs and GMCs/neurons within ci94 clones, similar to those observed in smoIA3 clones (Figure 2M). In contrast, NB clones expressing a constitutively active form of ci, ciNc5m5m, contained fewer Elav-negative GMC-like cells (2.1±1.1 cells per clone, n = 15) similar to ptcS2 clones (Figure 2N). Moreover, ectopic Hh signaling induced by over-expressing a constitutively active form of ci, ci5m5m, or smo, smoRA1234, could also lead to aberrant nuclear Pros localization in the NB, but at a lower phenotypic penetrance than ptcS2 clones (Figure S5 and unpublished data). Incidentally, removing one copy of smo in ptcS2 background (ptcS2, smo3/+) largely suppressed NB proliferation defects as these clones typically consisted of 4.1±1.1 GMC-like cells per clone (n = 25), a number that was comparable to that in wt clones (4.4±0.9 cells; Figure S3B). Furthermore, these NBs exhibited strong Mira crescent (Figure S3C; compared with Figure 1E). Similarly, Pros localization defect was largely rescued as normal crescent could be observed in all mitotic NBs (Figure S3D, compared with Figure 2E). These data indicate that a canonical Hh signaling cascade acts to control proliferation in NB lineages.
We next address the direct involvement of Hh signaling in regulating NB division rate by live imaging NBs in various genetic backgrounds. The majority of the postembryonic NBs have a rather tight frame of proliferative window from second instar larval stage ALH till early pupal stage [7]. It has been demonstrated that Drosophila postembyonic thoracic NBs exhibit increase in cell cycle time from earlier stage (96 h ALH) to later stage (120 h ALH) [44]. We observed a similar cell cycle trend for the postembryonic central brain NBs as live imaging of wt larval brains showed that, under our imaging conditions, their cell cycle lengthened from 93.5±12.8 min at 48 h ALH (n = 7) to 115.0±42.8 min at 72 h ALH (n = 13), and finally 154.1±42.0 min at 96 h ALH (n = 11) (Figure 3A and 3D; Video S1). smoIA3 NBs averaged a cell cycle time that was indistinguishable from that of wt NBs at 48 h ALH (93.6±14.1 min, n = 7) (Figure 3D). However, the cell cycle lengths of smoIA3 NBs were shorter than their wt counterparts at 72 h and 96 h ALH, clocking 76.4±21.8 min (n = 11) and 118.9±24.0 min (n = 9), respectively (Figure 3C and 3P; Video S2). Conversely, ptcS2 NBs significantly extended their cell cycle length at 48 h and 72 h ALH, averaging 183.3±28.4 min (n = 3), and 191.4±34.5 min (n = 7) respectively (Figure 3B and 3D; Video S3). In agreement with the low mitotic index at 96 h ALH in ptcS2 NBs, no dividing NB was observed despite many attempts. This implied that most of the NBs had either exited cell cycle (as supported by our BrdU feeding experiments), or had a long cell cycle time that exceeded our technical limitation to keep the explanted brains healthy in culture medium.
Studies in mammalian systems have shown that cell cycle lengths of neural precursors increase due to lengthening of the G1 phase as development proceeds [62]. We wondered whether the shorter cycling time in smoIA3 NBs could reflect a “younger state” that has a greater developmental potential. Indeed, 50% of the smoIA3 NBs (n = 14) continued to express NB proliferation markers Mira and PH3 at 48 h after puparium formation (APF) (Figure 4B and 4J), when all the wt NBs had exited cell cycle according to their normal developmental schedule (Figure 4A) [6],[7],[63]. Along with the proliferative defects seen in ptcS2 NBs, these observations led us to speculate that Hh signaling might have a role in promoting timely postembryonic NB cell cycle exit. It is known that the timing for termination of NB proliferation in the central brain is under the control of the temporal series with Cas and Svp as two essential players [44]. As implicated by their roles in scheduling cell cycle exit, many of cas24 and svp1 clones (53.7%, n = 41 and 53.0%, n = 34, respectively) retained a single Mira-positive, PH3-positive NB at 48 h APF, in a strikingly similar manner as smoIA3 clones (Figure 4B–4D and 4J). Incidentally, both cas24 and svp1 clones also showed an expansion of GMC-like cells (8.2±1.5 cells and 6.0±1.3, respectively, compared to 4.4±0.9 per clone in wt) at 96 h ALH, reminiscent of clones with compromised Hh signaling (Figure 4F and 4G, 4K). These data raise the possibility that Hh signaling may control NB cell cycle exit in conjunction with the temporal series.
Supporting the notion that Hh signaling is likely to exert an effect on NB cell cycle progression, we found that ptcS2 NBs down-regulated cyclin E (CycE) prematurely at 96 h ALH compared to wt NBs that retained high level of CycE expression in the nucleus (Figure S7A and S7B). While smoIA3 NBs showed normal CycE expression at 96 h ALH, its expression persisted at 24 h APF, consistent with the fact that most smoIA3 NBs remained mitotically active even at pupal stage (Figure S7C and S7D). Similarly, cas24 clones were also found to harbor CycE expressing NBs at 24 h APF (Figure S7E).
Is Hh signaling temporally regulated during NB development? Using RNA in situ hybridization against an intronic region of hh, which detected nascent nuclear transcripts, we first detected hh transcripts in late L2/early L3 larval stages and its expression level strongly increased in late L3 stage brains (Figure 5A and 5B) (see Methods for staging of larvae). Labeling of a single NB lineage showed that hh transcripts were expressed in the nuclei of the GMCs, particularly the newborn GMCs, based on their close proximity to the NB (Figure 5C and 5D). Interestingly, the dynamics of hh transcripts appeared to be cell-cycle specific in NBs as they were below detection threshold during interphase (Figure 5C), but became detectable during mitosis, suggesting that hh is likely transcribed during the G2 phase (Figure 5D). As the intronic probe detects hh pre-mRNA and RNA splicing is repressed during M-phase [64], we used an exonic probe to detect the presence of mature hh mRNA in interphase NBs. Predictably, mature hh cytoplasmic transcripts were detected in NBs and adjacent GMCs, but not in neurons (Figure 5E). Consistent with this expression pattern, Hh protein was found to show a corresponding age-dependent accumulation. Hh was not detected in L2 brains but became progressively apparent by L3 stage (Figure S6A and S6B). Interestingly, Hh protein built up and clustered around the cell surfaces as well as in intracellular puncta of some NBs. Hence, it is likely that Hh acts directly on the NBs to control their proliferation. Moreover, the pattern of hh reception using a ptc reporter line carrying a 2.8-kb upstream sequence fused to lacZ, as well as a ptcH84 enhancer trap insertion line (Figure S6C and unpublished data), clearly showed that NBs (and possibly the GMCs as well) were the signal receiving cells.
We also investigated the temporal expression profile of cas during development and found that Cas was transiently expressed in NBs during early larval development at about 24 h ALH, which coincided with timing of NB reactivation (Figure 5F) [7],[65], and its expression became quickly restricted to a small population of NBs by 28–30 h ALH (Figure 5G). Subsequently, Cas was detected mainly in GMCs and neurons, as well as a cluster of smaller-sized, Dpn-expressing intermediate neural precursors (INPs) located at the dorsal medial region of the central brain (Figure 5H).
Since Cas acts as a transcription factor [66] and given the temporal expression profiles of hh and cas, we tested the possibility that hh expression might be under the control of cas. In wt, hh transcript was often detected in the GMCs, which could be identified by the presence of cortical Pon. Although hh transcript levels remain high in the majority of the cas24 mutant clones induced at early L2 after the pulse of Cas during NB reactivation (68.4%, n = 19) (Figure 5I), hh expression was abolished or significantly reduced in 82.6% of the cas24 clones (n = 23) induced during embryonic development, presumably affecting the cas expression in early larval stage (Figure 5J); thus placing hh downstream of cas and suggesting that the competency for hh expression was likely dependent on the larval pulse of Cas, which occurred around 24 h ALH (Figure 5F). Intriguingly, constitutive expression of Cas by inducing UAS-cas clones in the central brain at both embryonic and larval stages (Figure S8A and S8B) also affected hh expression in a similar fashion as cas loss-of-function. Together, these data showed that stalling the temporal series either by removing or mis-expressing cas could negatively influence hh expression. Furthermore, UAS-cas expressing clones often harbored a Mira-positive NB at 24 h APF (Figure S8C), signifying that misregulation of temporal progression, similar to loss of hh signaling, can also extend the NB proliferative phase beyond its normal developmental limit.
Cas is a zinc finger protein capable of acting both as a transcriptional activator and repressor by binding to recognition sites with the consensus sequence of G/C C C/T C/T AAAAA A/T N [35]. Does the regulation of hh expression by Cas reflect the direct binding of Cas to the hh promoter? By scanning about 30 kb of genomic sequence flanking hh locus, we identified multiple potential DNA-binding sites that contained the Cas recognition sequence. We performed quantitative chromatin immunoprecipitation (ChIP) assays on S2 cells transiently transfected with Flag-tagged Cas and used the genomic region from pdm-1, a known Cas target, as a positive control. Cas directly binds to several consensus sites in the genomic region flanking pdm-1 ([35] and personal communication with W. Odenwald). Indeed, ChIP results showed that the pdm-1 cis-regulatory region were about three times more enriched with Flag-tagged Cas compared to control actin-5C promoter site (Figure 5K) and it was approximately 18 times more enriched compared to the non-transfected control (unpublished data). Supporting the notion that Cas directly associates with hh genomic region, Cas was highly enriched at a 6-kb genomic region encompassing hh transcription initiation site where there were 19 sites that matched at least eight out of ten bases of the Cas consensus binding sequence (Figure 5K). Specifically, the enrichments for Cas at hh genomic region compared to that at actin-5C promoter site and to the non-transfected control were up to 4.2 and 21.6 times, respectively. Thus, our results suggest that Cas is a direct positive regulator of hh, and its transient pulse of expression during early larval development is necessary for the subsequent Hh expression in postembryonic GMCs.
To further substantiate the relationship between cas and hh expression, we examined whether Hh signaling interacts genetically with cas. As shown earlier, cas24 NBs underwent persistent proliferation till late pupal stage as evidenced by the expression of proliferative markers Mira and PH3 at 48 h APF (Figure 4C and 4J). Ectopic activation of Hh signaling in this background by simultaneous expression of ptcRNAi partially suppressed the extended proliferation phenotypes (persistent Mira expressing NBs; Figure 4C, 4E, and 4J) as well as expanded GMC-like cells (Figure 4F, 4H, and 4K), while NBs expressing ptcRNAi alone did not exhibit any noticeable phenotype compared to wt NBs at that stage (unpublished data). These data indicate a clear genetic interaction between Hh signaling and Cas (thus the temporal series) (Figure 4J and 4K). To further substantiate this interaction, we used an alternative approach to introduce elevated Hh signaling in cas24 clones. It has been shown that the PP4 and PP4R subunits complex functions to down-regulate Hh signaling by dephosphorylating and destabilizing Smo [67]. Hence, compromising PP4 activity leads to elevated Hh signaling activity. Indeed, suppression of the persistent proliferation phenotype associated with cas mutant NBs was also evident when PP4-19C, the catalytic subunit of PP4 complex, was compromised in the cas24 background (Figure 4I–4K). The total GMC-like cells within cas24 clones were brought down from 8.2±1.5 (n = 14) to 5.5±1.3 (n = 23) and 4.5±1.2 (n = 21) cells per clone with the introduction of ptcRNAi and PP4-19CRNAi transgenes, respectively, a level close to that seen in wt clones (Figure 4K). Similarly, unlike cas24, the majority of the NBs in ptcRNAi; cas24 and PP4-19CRNAi; cas24 double-mutant clones were no longer mitotically active at 48 h APF (Figure 4E and 4J). Together, these data indicate that Hh signaling interacts genetically with the temporal series.
Previous studies placed svp downstream of cas in type I NB lineages during the progression of temporal series at larval stage [44]. Incidentally, we found that svp clones exhibited GMC-like cell expansion as well as an extension of the NB proliferation window similar to those seen in cas mutants (Figure 4D, 4G, 4J, and 4K). As Cas is potentially a direct regulator of Hh, we first sought to investigate if Hh signaling could in turn, affect the expression of svp. In wt clones, Svp was detected in the nucleus of the NBs and a minority of neurons at 40 h ALH (Figure S9A). Interestingly, svp expression was unaffected in smoIA3 clones in which Hh signaling was compromised (Figure S9C), indicating that Hh signaling does not function upstream of svp. Consistent with this observation, neither did up-regulation of Hh signaling in ptcS2 clones cause ectopic (or elevated) svp expression (Figure S9B). It is also conceivable that there is a requirement for svp as a downstream component of cas to activate hh expression. Surprisingly, hh transcript remained detectable in the GMCs of svp1 mutant clones, whereas constitutive svp mis-expression clones did not trigger ectopic hh expression (Figure S9D and S9E). These results indicated an unlikely placement of hh downstream of svp as well. To gain a better perspective of Svp in the temporal series, we looked at its temporal expression pattern in the postembryonic central brain. We found that svp was expressed at a high level in the NBs at 24 h ALH, as early as the time when cas expression was detected (Figure S9F). However, unlike cas, which was expressed in a short pulse, svp had a much wider expression window where medium levels of Svp can be detected in the NBs until 50 h ALH and it continued to be expressed weakly till 96 h ALH (Figure S9G–S9I and S9K). Given the similar time frames of cas and svp expression at early larval stage, we generated embryonic cas24 clones and assayed for svp expression at 24 h ALH. It was found that svp expression was not abolished in cas mutant NBs, showing that svp might not be genetically downstream of cas in the NBs of the central brain (Figure S9J).
How does Hh signaling act to control NB cell cycle exit downstream of Cas? One of the positive targets of the temporal series is Grh, which is expressed as a terminal temporal series component in the late embryo. It is expressed upon activation by Cas and its expression persists during larval stages to maintain the mitotic activity of type I NBs [36],[38],[40],[44]. Down-regulation of Grh in the NBs coincides with cell cycle exit, which is accompanied by the reduction of NB size, delocalization of Mira from the cortex to the cytoplasm during early mitosis, as well as Pros translocation into the nucleus [44]. We noted that ptcS2 NBs also exhibited similar Mira and Pros mislocalization (Figures 1E and 2E), although there were rare events in which some ptcS2 NBs managed to reach telophase with poorly localized Mira and Pros along the entire mitotic phase (Figure S10A–S10D). Moreover, there was a noticeable enrichment of Pros within the nuclei of interphase ptcS2 NBs as compared to the neighbouring wt NBs with 22.8%±4.1% increase in the intensity of Pros (n = 72, p<0.0001) (Figures 2E, S10A, and S10E). In addition, the size of ptc mutant NBs at 96 h ALH (9.9±1.2 µm, n = 72 for ptcS2 mutant, and 9.2±1.1 µm, n = 36 for ptc13 mutant) was consistently smaller than that of wt NBs (11.2±0.9 µm, n = 61) (p = 1.0×10−10) (Figure S11A–S11C). We hypothesized that Hh signaling might function through Grh in promoting NB cell cycle exit, hence we examined the pattern of Grh expression in NB clones at various time intervals from early L3 until 24 h post-pupation.
In wt L3, within the window of 72–96 h ALH, a high level of Grh was always detected in the NBs in conjunction with relatively lower expression levels in some GMCs (Figure 6A and 6B). Down-regulation of Grh occurred at around 12 h APF during which the NBs retained low levels of Grh while the GMCs were devoid of Grh (Figure 6C). By 24 h APF, Grh became barely visible in the NBs (Figure 6D), consistent with the reported cessation of NB proliferation between 20–30 h APF [7]. However, in ptcS2 clones, NBs appeared to down-regulate Grh expression prematurely. At 72 h ALH, Grh expression was absence from GMCs and its level was significantly reduced in NBs at 96 h ALH, and subsequently lost completely from all NBs at 12 h APF (Figure 6E–6G). This was consistent with the lack of mitotic activity in ptcS2 NBs (Figure 1H), presumably most of the NBs had exited the cell cycle (or struggled with cell cycle progression) by 96 h ALH. As for smoIA3 clones, the NBs and some GMCs persistently expressed Grh from larval stages until 12 h into pupal stage (Figure 6H–6J). At 24 h APF, all NBs (n = 8) within smoIA3 clones remained Grh-positive but some GMCs began to down-regulate Grh expression (Figure 6K and 6L). More intriguingly, Grh expression persisted in all of the NBs for all the smoIA3 clones examined at 36 h APF (n = 16), and as much as 86.7% of the clones still retained some level of Grh expression in the GMCs (Figure S11D). The prolonged expression of Grh in the smoIA3 NBs could probably explain the presence of NB proliferative markers Mira and PH3 in some of the smoIA3 clones at 48 h APF (Figure 4B and 4J). Similar observation was obtained for ci94 homozygous clones in which 90.1% of the Dpn-positive NBs retained Grh expression, and up to 45.5% of these NBs were surrounded by Grh-expressing GMCs at 24 h APF (n = 22) (Figure S11E).
To substantiate a mechanistic link between Hh signaling, Grh expression as well as the temporally regulated cell cycle exit, we reduced the level of Grh by RNA-mediated interference in smoIA3 clones. At 24 h APF, 68.5%±7.8% of smoIA3 clones (n = 54) contained a NB that was positive for the proliferative marker Mira (Figure 7A). In contrast, expression of grhRNAi within smoIA3 background significantly reverted this NB cell cycle termination defect by bringing down the frequency of clones with a proliferating NB at 24 h APF to 17.6%±9.9% (n = 68, p = 0.009), a level that was comparable to that of the control grhRNAi clones (21.6%±10.9%, n = 51, p = 0.07) (Figure 7B and 7C). Moreover, the number of GMC-like cells at 96 h ALH, which is indicative of the proliferative capacity of the NBs, was reduced to 3.4±1.2 cells per grhRNAi expressing smoIA3 clone (n = 18, p<0.0001), as compared to 6.7±2.6 cells per smoIA3 clone and 2.7±1.3 cells per control grhRNAi clone (Figure 7D and 7E). On the other hand, the NB proliferative defect seen in ptcS2 clones was only marginally modulated by grh ectopic expression. Compared to ptcS2 mutant NBs with a mitotic index of 21.6%±8.4%, ptcS2 mutant NBs with forced grh expression exhibited a mitotic index of 26.1%±9.8% (n>45, p = 0.23). Hence, it is conceivable that Hh signaling might regulate NB behavior by other mechanisms in addition to Grh. Despite that, we noted that Mira delocalization defect was significantly rescued with 69.7%±18.4% of the ptcS2 ;grho/e NBs showing proper Mira localization (n = 96, p = 0.008) as compared to only 37.8%±19.6% among the corresponding ptcS2 NBs (n = 142) (Figure 7F–7I). Similarly, the percentage of ptcS2 NBs with wt-level of dpn-expression intensity improved from 32.2%±15.3% (n = 59) to 61.5%±10.1% with simultaneous induction of grh expression (n = 65, p = 0.02) (Figure 7F–7I). Thus, Hh signaling regulates NB cell cycle exit, partly via grh expression, in conjunction with the temporal series.
On one hand, Pros as a component of NB asymmetric division machinery, is normally tethered onto the cortex by Mira and kept out of the NB nucleus. On the other hand, a burst of interphase nuclear Pros (accompanied by cytoplasmic localization of Mira) is triggered by the temporal mechanism during the terminal division of NBs, indicative of cell cycle exit [44]. However, how the temporal mechanism can mediate nuclear Pros localization in NB is unknown. ptcS2 NBs were associated with a higher level of nuclear Pros during interphase and cytoplasmic Pros during mitosis (Figure 2E, S10A–S10D). Pros functions by repressing genes required for NB self-renewal and activating genes implicated in neuronal differentiation [28]. Hence, we reasoned that the slower mitotic cycle and premature cell cycle exit in ptcS2 NBs may be a result of mislocalized/nuclear Pros. Indeed, removal of one copy of functional pros largely suppressed the Mira mislocalization phenotype associated with ptc loss-of-function (Figure 8A and 8B). As many as 71.4% of the ptcS2; pros17/+ NBs (n = 21) expressed cortical Mira during interphase and strong Mira crescent during mitosis. Furthermore, the number of Dpn-negative, Elav-negative GMC-like cells was reverted to the wt level of 5.0±1.5 cells per clone (Figure S9A and S9B). As a control, clones with a single copy of functional pros had an average of 4.1±1.0 GMC-like cells (n = 40), showing that decreased level of pros itself in a wt background did not cause GMCs to over-proliferate (Figure S12C).
Like ptcS2, mutations in the subunits of PP4 complex had been reported to exhibit a NB under-proliferation phenotype with similar defects in the localization of Mira and Pros [32]. The majority of mitotic NBs in flflN42/flfl795 (the regulatory subunit of PP4) trans-heterozygotes exhibited nuclear Pros during interphase ([32] and unpublished data) and lacked a well-defined Mira crescent but instead, showed cytoplasmic accumulation of Mira during mitosis (Figure 8C and 8G). Like ptcS2 NBs, removal of one copy of pros in flflN42/flfl795 trans-heterozygotes could partially restore Mira crescent to the cortex as the percentage of NBs with strong Mira crescent increased from 13.0% to 34.6% (Figure 8D and 8G). As PP4 is known to be a Smo phosphatase that dephosphorylates Smo thus dampening Hh signaling [67], abolishing the function of PP4 will invariably lead to elevated Hh signaling. Supporting the idea that excess Hh signaling may in part be the cause for Mira/Pros delocalization in flfl mutants, we knocked down ci by RNAi in the flfl795 background and observed that the number of NBs with strong Mira crescent increased from 21.4% to 43.5%, while NBs expressing ciRNAi did not exhibit any noticeable phenotype in that respect (Figure 8E, 8G, and unpublished data). Thus, our data suggest that PP4 regulates NB divisions, in part, by dephosphorylating Smo and modulating Hh signaling to keep NBs in a proliferative state. Supporting this, compromising PP4 function suppressed the formation of expanded GMC-like pools and promoted cell cycle exit in a cas mutant background (Figure 4I–4K). Together, our results suggest that Hh signaling plays a role in NB asymmetric division via the regulation of Mira/Pros localization; it also acts both downstream (cas) and upstream (grh) of components of the temporal series to control NB cell cycle exit. Thus Hh signaling appears to be a key player in coordinating the asymmetric division machinery with the temporal cascade to schedule the termination of postembryonic neurogenesis in line with developmental timing.
Here, we show that Hh signaling functions during later postembryonic development and acts together with the NB temporal transcription factor cascade to regulate NB cell cycle exit (Figure 9). We further demonstrate that hh is a downstream target of Cas, a member of temporal series that determines the time at which NBs terminate proliferation via down-regulation of Grh. While increased Hh signaling results in increased cell cycle length and premature NB cell cycle exit, loss of Hh signaling decreases NB cell cycle length and also prolongs the duration of NB proliferation.
Hh family proteins can act as short- or long-range morphogens covering distances as few as ten cell diameters (20 µm), or as far as a field containing many more cell diameters (200 µm) [68],[69]. In the postembryonic brain, hh is expressed predominantly in the NBs and the newborn GMCs, whereas the expression of its target gene reporter, ptc-lacZ is observed in a narrow area covering the adjacent NB and the sibling GMCs, indicating a limited response to and suggesting a limited spread of Hh ligand. In addition, Hh protein is always found to be enriched and clustering around the NBs in a punctuated form rather than forming a gradient. These data, together with the lineage autonomous phenotype of hh mutant NB clones, strongly suggest that Hh acts locally at short range in the larval brain. This is consistent with the structural arrangement of the larval brain, where each NB lineage comprising of the NB itself, GMCs, and neurons, is encapsulated by a meshwork of glial processes that form a three-dimensional scaffold that potentially acts as a stem cell niche [70],[71]. Such a spatial arrangement may serve as a barrier to restrict spread of the ligand and confine signaling events within a particular lineage so that an individual NB lineage can development with considerable independence from its neighbouring lineages. Indeed, a NB clone derived from a hh null allele exhibits the GMC pool expansion phenotype even though GMCs from its neighbouring lineages are competent in producing Hh ligand.
While it is tempting to assume that Hh can also act on the GMCs in an autocrine mode of action judging from the presence of ptc-lacZ expression, we did not observe any noticeable GMC fate transformation or change in their proliferative capability in ptcS2 and smoIA3 clones. The higher mitotic rate in hh loss-of-function NBs could largely explain the amplification of the GMC pool and enlarged clone-size; however, we are unable to rule out a possible delay in GMC differentiation. The proposition that Hh ligand, which is produced by the NB and daughter GMCs, feeds back on the NB to control its own proliferative capacity and the timing of cell cycle exit is interesting but not totally unfamiliar. Similar feedback signalling mechanism has been demonstrated in the mouse brain in which post-mitotic neurons signal back to the progenitor to control cell fate decisions, as well as the number of neurons and glia produced during corticogenesis [72].
Hh signal reception is detectable in NBs as early as in L2 and persists throughout larval life and in early pupae when NBs undergo Pros-dependent cell cycle exit. This delay of approximately 96 h between the start of Hh reception and the ultimate outcome of cell cycle exit may be due to a requirement for cumulative exposure of NBs to increasing local concentrations of Hh. Such a graded response will enable the wt postembryonic NBs to progress from high to low proliferative stages before ceasing division, in line with the development of the larva. Evidence supporting this notion includes gradual accumulation of Hh on the NBs, lengthening of NB cell cycle time, as well as the necessity of high levels of Hh signaling to trigger cell cycle exit (unpublished data). It is worthwhile to note that even at pre-pupal stage during which most NBs are starting to undergo cell cycle exit, fewer than 20% of them are associated with Hh puncta at any point of time. One likely explanation is that not all the NB lineages within the larval central brain respond synchronously to Hh-mediated temporal transition. However, unlike the embryonic central nervous system in which hh expression is localized to rows 6–7 of the neuroectoderm [73], we find it difficult to pinpoint a specific expression pattern in the postembryonic central brain due to the disorganized array of NB lineages. It is equally possible that different NBs exit cell cycle progression at different time points. This is also consistent with the structural organization of individual NB into different “trophospongium” or stem cell niches. Nevertheless, we cannot rule out the possibility that Hh signal activation primes another yet-to-be-identified developmentally regulated signal/event to schedule NB cell cycle exit.
Interestingly, a recently proposed “cell cycle length hypothesis” postulates that cell cycle length, particularly the length of G1 phase in neural stem cells acts as a switch to trigger the transition from proliferative to neurogenesis mode [62]. In fact, experiments have shown that manipulation of cdk4/cyclinD1 expression and cdk2/cyclinE activity that result in the lengthening of G1 is sufficient to induce precocious neurogenesis; while inhibition of physiological lengthening of G1 delays neurogenesis and promotes expansion of intermediate progenitors [74],[75]. Our results show that Drosophila postembryonic NBs in the central brain exhibit a comparable trend of cell cycle lengthening from young to old larval stages. Interestingly, NBs with excess Hh signaling have an extended cell cycle time, consistent with the idea that there is a forward shift of the “perceived” age, leading to premature cell cycle exit. In contrast, Hh loss-of-function NBs have a shorter cell cycle time compared to their wt counterparts of the same actual age; hence, they have a younger “perceived” age and are able to maintain their proliferative phase over a longer period of time. Consistent with this, we showed that persistent NB proliferation in smoIA3 clones as well as the early termination of ptcS2 NBs proliferation, are always associated with the presence and absence of CycE expression, respectively. However, loss of Hh signaling in NBs merely extends their proliferative phase but is not sufficient to ensure perpetual proliferation as we failed to observe any mitotic NB in the adult brain. We also note that a previous report suggested that the cell cycle time of the larval NBs reduced during their growth and reached a peak at late third instar with a minimum cell cycle time of 55 min. However, this study was conducted on thoracic NBs from the neuromeres T1 to T3, which have a very distinctive proliferative profile to the central brain NBs assayed in the current study [6]. Indeed, it was shown in their study that abdominal NBs exhibited significantly different cell cycle times compared to their thoracic counterparts.
In Drosophila, the precise timing of NB cell cycle exit is governed by a highly regulated process that involves sequential expression of a series of transcription factors: Hb→Kr→Pdm1→Cas, known as the temporal series [36],[37],[76]. It is known that the temporal series probably utilizes Grh in the postembryonic NBs to regulate Pros localization or apoptotic gene activity, thus determining the time at which proliferation ends. In addition, the temporal series also regulate postembryonic Chinmo→Br-C neuronal switch, which specifies the size and the identity of the neurons [44],[77]. Our data show that Hh signaling does not regulate early to late neuronal transition as Chinmo and Br-C expression timings appear unaffected in both ptc and smo mutant clones (unpublished data). In contrast, excess Hh signaling leads to a variety of features associated with NB cell cycle exit: (1) premature down-regulation of Grh, (2) nuclear localization of Pros (in NBs), and (3) reduction of NB size. Taken together with the extended proliferative duration of Hh loss-of-function NBs, it is apparent that Hh signaling is a potent effector of the temporal series and functions late to promote NB cell cycle exit.
The results from our current genetic interaction assays with Hh pathway components and grh reaffirmed the conclusions from previous studies that Grh is necessary to maintain the mitotic activity of the postembryonic NBs [42],[44]. The loss of Hh signaling keeps the central brain type I NBs in their proliferative state and this is largely contributed by persistent grh expression past their normal developmental timing at around 24 h APF. Even though Grh is necessary to extend the proliferative phase of these NBs, it is not sufficient to rescue all aspects of the premature cell cycle exit phenotype seen in ptc mutant NBs. Hence, down-regulation of grh by over-activating Hh signaling is not solely responsible for NB proliferative defects, and this implies that Hh signaling may terminate NB cell cycle via other mechanisms in addition to Grh.
The expression of hh appears to be dependent on the pulse of Cas expression at the transition between L1 and L2, as induction of cas mutant clones after that stage does not significantly affect hh expression. Moreover, ChIP assays suggest that Cas binds the hh genomic region, thereby placing Hh as a direct downstream target of the temporal series. However, it is intriguing to speculate on how the early pulse of Cas can mediate hh expression, which only comes on later during larval development. One possible explanation involves a relay mechanism in which that pulse of Cas activates an (or a cascade of) unknown components, which persist and eventually turns on the later hh expression. Yet, in such a model, Cas need not interact directly with the hh locus as our ChIP assay clearly suggests. Moreover, there are at least two pulses of hh expression during larval brain development, and the earlier, shorter pulse that is required for the activation of quiescent NBs appear to be independent from Cas regulation as Cas is only switched on in the larval NBs upon reactivation [70]. Most importantly, our data show that mis-expression of cas abolishes, rather than triggers ectopic hh expression. Thus, our findings do not favour the continuous expression of a hh activator downstream of Cas. Alternatively, Cas may be involve in the epigenetic modifications of the hh locus such that it is primed for expression at a much later stage. This may also explain why saturating the system with Cas for prolonged period of time via mis-expression can negatively affect subsequent hh expression because of to its potential aberrant association with the chromatin. Although such a function has not been reported for Cas, previous studies have postulated that components of the temporal series, such as Hb (or mammalian homolog Ikaros) and Svp (or mammalian homolog COUP-TFI/II), play a role in modulating chromatin structure, hence modifying the competency of downstream gene expression subsequently [78]–[80].
The relationship between svp and Hh signaling within the postembryonic temporal series cascade is interesting yet unexpected. svp was thought to be a downstream component of cas on the basis of studies in postembryonic NBs in the thoracic segment of the ventral nerve cord [44]. This is supported by the observations that the pulse Svp occurs at 40–60 h ALH following the pulse of Cas at 30–50 h ALH. Moreover, both svp and cas mutant clones affect Chinmo/Br-C neuronal target transition, apart from causing NBs' failure to exit the cell cycle at early pupal stage. However, examinations of Svp and Cas expression patterns in the central brain region in this study reveal that the Cas expression window overlaps with the peak of the Svp expression window, even though the latter has a much wider expression window in which low expression levels can still be detected in the NBs at 96 h ALH. Moreover, our data show that abolishment of cas function starting from the embryonic stage does not reduce Svp expression in the NBs at 24 h ALH. Hence, previous interpretation that svp functions downstream of cas in the thoracic postembryonic NBs may not be easily extrapolated to NBs in other brain regions. On the basis of our results, it is tempting to postulate that Cas and Svp constitute two parallel pathways within the temporal series and Hh signaling is regulated by Cas but not Svp. Nevertheless, such a hypothesis warrants more in depth studies.
The precise generation of diverse cell types with distinct function from a single progenitor is important for the formation of a functional nervous system during animal development. It has been shown that, in Drosophila, the developmental timing mechanism (the temporal series) is tightly coupled with the asymmetric machinery [44]. However, the underlying mechanism of this coordination remains elusive. Our data suggest that on the one hand, Hh signaling is under the control of the temporal series (hh expression is directly regulated by Cas), while on the other hand, Hh signaling participates in asymmetric segregation of Mira/Pros during NB division. Introduction of ectopic/premature Hh signaling (in ptc mutant clones) during developmental stages in which NBs are proliferating results in cytoplasmic localization of Mira/Pros during mitosis, reduction of NB size, and slow-down of NB cell cycle progression, reminiscent of the final division of NBs in early pupa just before cessation of proliferation. Consequently, these NBs exit the cell cycle prematurely. We speculate that Pros may be a direct or indirect target of Hh signaling as elevated pathway activity invariantly leads to increased pros expression in the NBs. Furthermore, reducing the level of Pros protein by removing one copy of function pros is able to rescue the Mira delocalization phenotype seen in ptc mutant NBs. Thus, it is plausible that Hh signaling impinges on the asymmetric division apparatus, likely through Pros, to diminish NB fate gradually (as seen with the absence of Dpn and Mira delocalization) prior to the final cell cycle exit. Despite our results indicating a tight correlation between nuclear entry of Pros into the NBs and the eventual cell cycle exit of these NBs during pupal stage, we would like to caution the readers that Pros may not be the direct causative agent in controlling NB cell cycle exit. Therefore the actual role of Pros in this process is purely speculative as far as this study is concerned.
On the other hand, loss of Hh signaling (e.g., in Smo mutant clones) maintains NBs in their “younger” proliferating stage far beyond the time when they normally exit the cell cycle. Thus, Hh signaling couples the developmental timing mechanism (the temporal series) with the NB intrinsic asymmetric machinery for the generation of a functional nervous system.
In vertebrates, constitutive activation of the Sonic hedgehog (SHH, a homologue of Drosophila Hh), signaling pathway through inactivation mutations in PTCH1, activating mutations in SMO, as well as other mutations involving SHH, IHH, GLI1, GLI2, GLI3, and SUFU, has been implicated in a vast array of malignancies [81],[82]. The proven association of Hh signaling pathway with tumourigenesis and tumour cell growth fuel the view that Hh constitutes a mitogenic signal that promotes pro-proliferative responses of the target cells. Moreover, Hh acts as a stem cell factor in somatic stem cells in the Drosophila ovary, human hematopoietic stem cells, and mouse embryonic stem cells, possibly by exerting its effects on the cell cycle machinery [83]–[86].
Our report here provides an opposing facet of Hh signaling where it is required for timely NB cell cycle exit in the postembryonic pupal brain. This may sound astonishing, but the essential roles of Hh signaling as a negative regulator of the cell cycle has been eclipsed by the common bias that it stimulates proliferation, given the many examples of malignancies with the Hh pathway dysregulation. Indeed, studies have indicated that cell cycle exit and differentiation of a number of cell types, such as absorptive colonocytes of the mammalian gut, zebrafish, and Drosophila retina, require Hh activities [87]. A more recent article also showed that SHH signaling pathway is highly activated in human embryonic stem cell (hESC) and such activity is crucial for hESC differentiation as embryoid bodies [88]. The opposing functions of Hh signaling pathway in different cell types reveal that the ultimate effect of this pathway is likely to be tissue specific, depending on its interaction with other regulatory pathways. Our data indicate that in Drosophila postembryonic NBs of the brain this does indeed appear to be the case, because in this system, Hh signaling pathway interacts with NB-specific temporal series and likely the asymmetric cell division machinery to promote pros nuclear localization to trigger cell cycle exit.
All fly stocks and crosses were maintained at 25°C. Stocks used were FRT40A, FRT42D, FRT82B, ptcS2, ptc13 (P. Ingham), smoIA3, ci94 (K. Basler), cas24 (A. Gould), pros17, flflN42, flfl795, G147, svp1, grhRNAi (Bloomington, 33678/GD), ptcRNAi (Bloomington, JF03223), PP4-19CRNAi (VDRC, 25317/GD), ciRNAi (VDRC, 51479/GD), elav-GAL4, UAS-CD8::GFP, UAS-ciNc5m5m (D. Kalderon), ptc-LacZ (J. Hooper), and UAS-grh (S. Thor), UAS-histone2AvRFP (M. Gonzalez-Gaitan), UAS-svp1.12 (Y. Hiromi), UAS-stg.N4, UAS-cycE.L, UAS-smoRA1234 (J. Jiang), UAS-histone::RFP (J. Bellaiche), UAS-pon::GFP (B. Lu), Ay-GAL4 (act flip-out). All stocks were obtained from Bloomington Stock Center unless otherwise stated.
Embryos were collected over a period of 6 h, heat-shocked in 37°C water bath at 24 h and 48 h ALH for all experiments unless otherwise specified, and larvae and pupae of desired genotype (see below) were dissected at specific time points and processed for immunochemistry analysis. Under our culture conditions, Drosophila larvae underwent approximately 108 h of postembryonic development. After hatching from the embryo, the 1st-instar larva (L1) stage lasted for 24 h before molting into 2nd-instar larva (L2). The L2 to 3rd-instar larva (L3) transition occurred at approximately 48–60 h after hatching, and finally L3 larva pupate at 96–108 h after hatching. MARCM clones were generated according to the technique reported previously [55].
Brains were fixed for 15 min in 3.7% formaldehyde in PBS with 0.1% Triton-X. The following antibodies were used: mouse anti-Mira (F. Matsuzaki), 1/50; rabbit anti-Mira (generated in our lab), 1/1,000; chicken anti-green fluorescent protein (GFP) (Abcam), 1/2,000; guinea-pig anti-Dpn (J. Skeath), 1/500; mouse anti-Pros (DSHB), 1/10; rat anti-Elav (DSHB), 1/5; mouse anti-BrdU (Sigma), 1/50; rat anti-Hh (I. Guerrero), 1/20; rabbit anti-Grh (B. Bello), 1/200; rabbit anti-Pon (Y.N. Jan), 1/500; rabbit and mouse anti-phosphohistone H3 (Abcam); rabbit anti-aPKCζ C20 (Santa Cruz Biotechnologies), 1/1,000; rabbit anti-Pins, 1/1,000; rabbit anti-Insc, 1/1,000; guinea-pig anti-Numb (J. Skeath), 1/1,000. Secondary antibodies were conjugated to either Alexa Fluor 488, Alexa Fluor 555, or Alexa Fluor 633 (Molecular Probes), and used at 1/500, 1/1,000, and 1/250, respectively. DNA stain was To-PRO-3 (Molecular Probes), 1/5,000 and samples were mounted in Vectashield (Vector Laboratories). Images were obtained using Zeiss LSM 510 upright microscope and processed in Adobe Photoshop CS3 and Adobe Illustrator CS3.
Brains were dissected from larvae at 48 h, 72 h, and 96 h ALH, and were prepared for live imaging using the clot method as describe previously [89],[90]. Image acquisition was performed at 25°C on Leica SP5 inverted microscope. Multiple z-sections were recorded with step-size of 2–4 µm. Each z-stacks was recorded every 5 min over a period of 6–8 h in order to capture at least one complete cell cycle (only the first cell cycle will be considered for calculation of cell cycle length). Images obtained were processed using Adobe Photoshop CS3 and ImageJ. Refer to Text S1 for genotypes of the larvae used.
Larvae at the age of 82–88 h ALH were picked up from the fly food and starved for 1 h on a clean Petri dish. They were then fed with yeast paste infused with 0.1 mM BrdU (Roche) for 4 h before being dissected and analyzed.
hh genomic region was amplified with Expand High Fidelity PCR system (Roche) to yield a ∼1.9-kb intronic and a ∼0.7-kb exonic template using the following primer-pairs:
Intronic: 5′-GTGGATTTGGATCTGGCTATC-3′ and 5′-CAATTAGCCGCGATACAGCAC-3′
Exonic: 5′-ATTCGTCGATCAGTTCCCACGTGC-3′ and 5′-GATGGAATCCTGGAAGAGCGATCC-3′
Digoxigenin-labeled probes were generated using DIG RNA Labeling kit according to the manufacturer's instructions (Roche). Larval brains of specific genotype were rinsed thoroughly with PBS and dissected at specific age. They were fixed for 20 min in 3.7% formaldehyde in PBS supplemented with 0.1% Tween-20. In situ hybridization was performed as previously described [91],[92].
Full length cDNA of cas was amplified from DGC gold collection (LD36057, BDGP) using Expand High Fidelity PCR system (Roche) with the primers: 5′-ATGTCCAACCAAATGGAGTTTA-3′ and 5′-CTACTCCTTAAACTCTGGCTTAAAGCT-3′. The resultant PCR product was TOPO-cloned into pENTR vector (Invitrogen) and switched into pAFW vector (Drosophila Gateway Vector collection) using the pre-existing protocol (T. Murphy). The flag-tagged Cas construct was fully sequenced.
2×106 S2 cells were seeded onto a 75-ml culture flask at 25°C a day prior to transfection. 2.5 µg of Flag epitope tagged Cas construct was transfected into these cells using Qiagen Effectene transfection reagent. DNA to effectene ratio was maintained at 1∶20. 24 h post-transfected cells were used for ChIP.
ChIP was performed according to the manufacturer's protocol for EZ-Magna ChIP G (Milipore).
qPCR was performed using KAPA SYBR FAST qPCR kit (KAPA Biosystems) according to the standard protocol on 7900HT Fast Real-Time PCR system (Applied Biosystems). The sequences of the primer-pairs used are listed in Text S1.
|
10.1371/journal.ppat.1006664 | EBF1 binds to EBNA2 and promotes the assembly of EBNA2 chromatin complexes in B cells | Epstein-Barr virus (EBV) infection converts resting human B cells into permanently proliferating lymphoblastoid cell lines (LCLs). The Epstein-Barr virus nuclear antigen 2 (EBNA2) plays a key role in this process. It preferentially binds to B cell enhancers and establishes a specific viral and cellular gene expression program in LCLs. The cellular DNA binding factor CBF1/CSL serves as a sequence specific chromatin anchor for EBNA2. The ubiquitous expression of this highly conserved protein raises the question whether additional cellular factors might determine EBNA2 chromatin binding selectively in B cells. Here we used CBF1 deficient B cells to identify cellular genes up or downregulated by EBNA2 as well as CBF1 independent EBNA2 chromatin binding sites. Apparently, CBF1 independent EBNA2 target genes and chromatin binding sites can be identified but are less frequent than CBF1 dependent EBNA2 functions. CBF1 independent EBNA2 binding sites are highly enriched for EBF1 binding motifs. We show that EBNA2 binds to EBF1 via its N-terminal domain. CBF1 proficient and deficient B cells require EBF1 to bind to CBF1 independent binding sites. Our results identify EBF1 as a co-factor of EBNA2 which conveys B cell specificity to EBNA2.
| Epstein-Barr virus (EBV) infection is closely linked to cancer development. At particular risk are immunocompromised individuals like post-transplant patients which can develop B cell lymphomas. In healthy individuals EBV preferentially infects B cells and establishes a latent infection without causing apparent clinical symptoms in most cases. Upon infection, Epstein-Barr virus nuclear antigen 2 (EBNA2) initiates a B cell specific gene expression program that causes activation and proliferation of the infected cells. EBNA2 is a transcription factor well known to use a cellular protein, CBF1/CSL, as a DNA adaptor. CBF1/CSL is a sequence specific DNA binding protein robustly expressed in all tissues. Here we show that EBNA2 can form complexes with early B cell factor 1 (EBF1), a B cell specific DNA binding transcription factor, and EBF1 stabilizes EBNA2 chromatin binding. This EBNA2/EBF1 complex might serve as a novel target to develop future small molecule strategies that act as antivirals in latent B cell infection.
| CBF1/CSL (C promoter binding factor, Suppressor of Hairless, and lag1 also called RBPJ or RBPJκ) is a cellular DNA binding protein, ubiquitously expressed in all mammalian tissues. CBF1 serves as a DNA adaptor molecule that recruits either repressors or activators to transcriptional control elements like enhancers and transcription start sites of genes and is described as the major downstream effector of the cellular Notch signal transduction pathway [1]. Notch signaling controls the development and differentiation of diverse organs and tissues. Despite the ubiquitous expression of its chromatin anchor CBF1, target gene control by Notch is context dependent and requires tissue and lineage specific cooperating transcription factors [2]. In B cells, latently infected with Epstein-Barr virus (EBV), CBF1 anchors the viral transactivator protein EBV nuclear antigen 2 (EBNA2) to chromatin and thereby initiates a cascade of signaling events that coordinate B cell activation and proliferation of infected cells [3–6]. Thus, EBNA2 is considered to mimic Notch signaling [7]. In contrast to the universal expression and pleiotropic activities of Notch, the expression and the biological activity of EBNA2 is strictly confined to EBV infected B cells, characterized by a transcription program that phenocopies antigen activated B cell blasts [8, 9].
CBF1 and EBNA2 frequently co-occupy cellular enhancer and super-enhancer regions reinforcing the concept that CBF1 is the major adaptor for EBNA2 to chromatin [10]. In addition, EBNA2 bound regions are co-occupied with multiple additional transcription factors including IRF4, BATF, NFκB, Runx, and ETS family members as well as the B cell lineage defining and pioneer factors PU.1/SPI1 and EBF1 [10, 11]. While the adaptor function of CBF1 is well defined, a potential functional contribution of these co-occurring factors to EBNA2 function has not been studied thoroughly. These proteins are active transcription factors which carry transactivation domains and can actively promote or impair transcription of target genes. PU.1/SPI1 promotes B cell development and is expressed throughout B cell differentiation, but also controls T cell, myeloid and dendritic cell differentiation [12]. PU.1/SPI1 DNA binding sites are critical for LMP1 promoter luciferase activation [3, 13–15]. However, its contribution to LMP1 expression in the context of the entire viral genome is surprisingly weak [16]. Most recently it has been shown that EBNA2 enhances the binding of CBF1 and EBF1 to chromatin and EBF1 is critical for expression of the EBNA2 viral target gene LMP1 [16, 17]. By sequential chromatin immunoprecipitation, EBF1 and EBNA2 have been shown to bind to the same chromatin fragment in the same cell [17]. Importantly, within the hematopoietic compartment EBF1 is exclusively expressed in B cells and their lymphocytic precursors. The other EBF gene family members EBF2, 3, and 4 are expressed at very low or undetectable levels in B cells. EBF1 initiates B cell lineage commitment, development and differentiation as a pioneer factor that promotes chromatin accessibility and DNA demethylation in lymphocyte precursors [18, 19].
Strong EBNA2 binding correlates with extended regions of extraordinarily high histone 3 lysine 27 acetylation (H3K27ac) and H3K4 mono-methylation (H3K4me1) marks which are characteristic features of activated super enhancers [11]. In addition, EBNA2 modulates the formation of chromatin loops to connect enhancers and promoters of its target genes [20]. In theory, EBNA2 co-occurring factors, like PU.1/SPI1 and EBF1 could function as pioneer factors for EBNA2 by modulating the chromatin state and thereby promoting access of EBNA2 to chromatin, indirectly. Alternatively, EBNA2 co-occurring factors might serve as alternate adaptors that promote DNA binding of EBNA2.
CBF1 is ubiquitously expressed in all mammalian cells including primary human B cells and EBV infected and non-infected human B cell lines. For this study, we used a CBF1 deficient human B cell line, which had been generated by homologous recombination in the somatic B cell line DG75, to screen for CBF1 independent functions of EBNA2. The parental DG75 B cell line is an EBV negative Burkitt's lymphoma cell line. While the knock-down of EBF1 and CBF1 in EBV immortalized B cells severely impairs cellular viability [17], DG75 cells tolerate inactivation of the CBF1 gene without loss of viability [21, 22]. We compared EBNA2 induced cellular genes in CBF1 proficient and deficient DG75 cells and found the majority of EBNA2 target genes to be CBF1 dependent. A minor fraction of EBNA2 target genes is regulated CBF1 independently. By chromatin immunoprecipitation and genome wide sequencing of EBNA2 bound DNA fragments (ChIP-Seq), we identified a subpopulation of CBF1 independent EBNA2 binding sites that was significantly enriched for EBF1 binding motifs. We show that CBF1 independent EBNA2 binding to chromatin is dependent on EBF1 protein expression. Importantly, we demonstrate that EBNA2 and EBF1 can form protein complexes in CBF1 positive and negative cells, indicating that EBF1 serves as B cell specific DNA anchor for EBNA2.
In order to rigorously test if EBNA2 can exert any functions in the absence of its DNA adaptor CBF1, a microarray based genome wide screen for EBNA2 target genes in DG75 B cells that are either proficient (wt) or deficient (ko) for CBF1 was performed (Fig 1A, 1B and 1F and S1 Table). Both cell lines constitutively express an estrogen receptor (ER) hormone binding domain EBNA2 fusion protein (ER/EBNA2). ER/EBNA2 is retained in the cytoplasm of the cell but is rapidly activated and translocated to the nucleus in response to estrogen [21, 22]. For expression profiling, DG75ER/EBNA2 CBF1 wt and CBF1 ko cells were cultured in estrogen supplemented media for 24 h, total cellular RNAs were harvested and processed for the hybridization of gene arrays that detect 30645 coding transcripts, 11086 lincRNAs (long intergenic non-coding RNA transcripts) and 148 miRNAs (micro RNAs). Cell cultures of the parental DG75 CBF1 wt and CBF1 ko cell lines, which do not express ER/EBNA2, were treated with estrogen and processed for the microarray analysis as specificity controls. Neither in DG75 CBF1 wt nor in DG75 CBF1 ko cells statistically significant changes (p ≤ 0.05) of cellular transcript abundance in response to estrogen treatment were observed, proving that target gene activation is strictly dependent on ER/EBNA2 (S1A Fig). In addition, estrogen responsive target genes described in the literature did not change expression levels proving that the estrogen receptor response is not functional in DG75 B cells (S1B Fig) [23–27]. It is important to note that EBNA2 not only activates a set of direct target genes but thereby initiates a cascade of secondary events, which are included in our target gene lists and in total reflect EBNA2 functions.
Based on expression level changes of 950 transcripts (≥ 2-fold, p≤ 0.05) which are regulated in DG75ER/EBNA2 CBF1 wt and expression levels of the same transcripts in DG75ER/EBNA2 CBF1 ko cells 12 clusters of transcripts were defined and illustrate the complex patterns that arise. Transcripts in cluster I, II, IV and VI are upregulated in CBF1 wt cells. Cluster VII transcripts are upregulated in both, wt and ko cells. Cluster III transcripts are down-regulated in CBF1 wt only, while cluster VIII transcripts are down-regulated in both, wt and ko cell (S2 Fig and S2 Table). Multiple previously characterized EBNA2 target genes were significantly upregulated in DG75ER/EBNA2 CBF1 wt cells (S1C Fig). In total, 99 cellular transcripts were up- and 37 cellular transcripts were downregulated ≥ 4-fold (p ≤ 0.001) (Fig 1A). Importantly, 15 transcripts were upregulated and 6 transcripts were downregulated in CBF1 deficient DG75ER/EBNA2 ≥ 4-fold (Fig 1B). Although the number of differentially expressed EBNA2 target genes was markedly higher in CBF1 proficient cells, the mean changes of the response were similar in CBF1 proficient and deficient cells as illustrated for genes regulated ≥ 2-fold (p ≤ 0.05) (Fig 1C).
The majority of EBNA2 target genes identified in CBF1 deficient cells were also regulated in CBF1 proficient cells, while a small group of targets is regulated by EBNA2 in CBF1 deficient cells, only (ko only, Fig 1D and 1E). On average, the transcripts which are regulated in CBF1 proficient and deficient cells (wt & ko) showed a stronger response than those regulated in proficient cells, only (wt only). In order to verify the microarray results, a panel of 12 CBF1 dependent (S2 Fig cluster 2) and independent targets (S2 Fig cluster 7) was selected for re-testing. RT-qPCR experiments confirmed that most CBF1 independent targets also responded to EBNA2 in CBF1 proficient cells. As already seen in the microarray experiment, the degree to which individual targets responded in CBF1 proficient cells varied considerably but was faithfully reproduced by RT-qPCR (S4 Fig). Interestingly, the CBF1 dependent target genes included a substantial number of miRNAs that are up- or downregulated by EBNA2 (S5 Fig).
To functionally characterize EBNA2 target genes, biological processes associated with individual subsets of genes were analyzed. The online tool GOrilla was used to test whether differentially expressed genes in comparison to all other genes on the array were enriched in any of the GO terms in the “Biological Process” category. The subsets considered here consisted of genes that were on average induced or repressed in the CBF1 proficient and deficient cell lines, or genes where induction or repression was dependent or independent of CBF1. Only genes significantly (q < 0.01) regulated in at least one of the two cell lines were considered. Thresholds on fold changes were chosen by the online tool GOrilla in a data dependent manner to identify subsets enriched in GO terms in the “Biological Process” category (S6 Fig).
Neither genes repressed in CBF1 proficient cells only (repressed/CBF1 dependent) nor induced in CBF1 deficient cells (induced/CBF1 independent) were significantly (q ≤ 10−4) enriched for any biological process. Genes induced in CBF1 proficient cells only (induced/CBF1 dependent) were strongly and most significantly enriched for immunoglobulin receptor binding and moderately enriched for biological processes involving several enzymatic activities (Table 1).
Target genes repressed by EBNA2 in the absence of CBF1 (repressed/ CBF1 independent) showed a remarkable profile (Table 2). They map to several GO terms that cover diverse immune responses. Since the study had been performed in B cells, the enrichment for genes involved in immune responses and B cell receptor biology could have been expected. However, our study indicates that EBNA2 also represses immune response genes and this feature of EBNA2 is CBF1 independent.
In summary, our differential expression analysis of EBNA2 target genes shows that EBNA2 can regulate a small fraction of its target genes without using CBF1 as a DNA anchor. In order to uncover alternative strategies of EBNA2 to bind to chromatin, we performed chromatin immunoprecipitation (ChIP) studies to identify genomic loci that are bound by EBNA2 in CBF1 negative cells. In ER/EBNA2 expressing cells, EBNA2 shuttles from the cytoplasm to the nucleus in response to estrogen. In order to avoid a potential impact of cytoplasmic ER/EBNA2 contamination on our biochemical studies, we switched to a doxycycline inducible HA-EBNA2 expression system (doxHA-E2) in DG75 (S7A Fig). In the absence of doxycycline, EBNA2 is not expressed and cannot interfere with the immunoprecipitation procedure in DG75doxHA-E2/CBF1 wt and DG75doxHA-E2/CBF1 ko cells. Up to 90% of the cells express EBNA2 when treated with doxycycline (S7B Fig). EBNA2 protein signal detected by immunostaining was 5- to10-fold stronger than EBNA2 in LCLs (S7C Fig). In comparison to LCLs, some of the EBNA2 co-occurring transcription factors like BATF and IRF4 were expressed at very low levels while EBF1 and PU.1/SPI1 were robustly expressed (S7C Fig). ChIP followed by high throughput sequencing (ChIP-seq) was performed to determine EBNA2 genome occupancy. 1,789 EBNA2 binding sites were identified in CBF1 proficient DG75dox-HA2 (Fig 2A), while 22,500 EBNA2 peaks were identified in LCLs, which had been performed in parallel. 1,325 (74%) of the EBNA2 peaks in DG75doxHA-E2 cells were also present in LCLs (shared peaks), while 464 binding sites occurred exclusively in DG75doxHA-E2 cells. EBNA2 signal intensity was most prominent at LCL/DG75doxHA-E2 shared EBNA2 binding sites (Fig 2B and S8 Fig).
In LCLs, EBNA2 is preferentially recruited to enhancer elements which pre-exist in peripheral CD19 positive B cells before they are infected by EBV to generate LCLs [10]. Chromatin marks characteristic for activated enhancer elements are H3K27ac in combination with H3K4me1 signals that are stronger than H3K4me3. We speculated that DG75 specific chromatin signatures in the absence of EBV infection might influence EBNA2 binding. We thus compared H3K4me1, H3K4me3, and H3K27ac signal intensities at EBNA2 binding sites i) shared by LCLs and DG75doxHA-E2, ii) unique for LCLs and iii) unique for DG75 in naïve CD19 positive B cells with those in non-transfected DG75.
EBNA2 binding sites, shared by LCLs and DG75doxHA-E2, stand out as the subset with the most prominent enrichment for all three investigated histone modifications associated with the chromatin state of active enhancers (Fig 2C). In contrast, DG75doxHA-E2 unique EBNA2 binding sites were highly enriched for active chromatin marks in the DG75 precursor only, while LCL unique EBNA2 peaks showed significantly lower signal intensities in DG75. These data indicate that a set of enhancers, which are pre-activated in DG75 cells but not in the CD19 positive LCL precursors, might allow the formation of "DG75 unique" EBNA2 binding sites. DG75 lack pre-formed enhancer signatures at "LCL unique" binding sites. ChIP signals for CBF1 and EBF1 are most highly enriched at EBNA2 binding peaks shared by LCLs and DG75doxHA-E2, while enrichment at LCL unique peaks was attenuated. In contrast, BATF and IRF4 were strongly enriched at shared LCLs and DG75doxHA-E2as well as LCL unique peaks. Thus, low abundance of IRF4 and BATF proteins or other co-occurring transcription factors in DG75doxHA-E2 might limit EBNA2 occupancy in DG75 at these LCL unique sites.
The comparison of ChIP-Seq data between DG75doxHA-E2 CBF1 wt and ko cells identified 1,789 EBNA2 binding sites in CBF1 proficient and 271 in CBF1 deficient DG75. 243 (81%) were found in both cell lines and thus constitute CBF1 independent EBNA2 peaks (Fig 3A). The majority of EBNA2 binding sites found in CBF1 proficient (74%) and deficient (83%) DG75 were shared with LCLs (S8 Fig). A small group of 28 binding sites were only identified in CBF1 deficient cells and were not analyzed further. 1,546 EBNA2 sites were not detected in CBF1 deficient cells and thus defined as "CBF1 dependent". The mean EBNA2 signal intensity at EBNA2 binding sites was elevated 1.4-fold in wt compared to ko cells (Fig 3B and 3C). Remarkably, EBNA2 binding to CBF1 independent peaks was significantly enriched compared to CBF1 dependent peaks in CBF1 wt cells (2.5-fold, Fig 3D and 3E). The quantitative re-analysis of the subclasses of EBNA2 peaks in LCLs confirmed that CBF1 independent peaks are characterized by stronger EBNA2 enrichment (Fig 3D, 3E and 3F, right panel). Since CBF1 independent EBNA2 binding obviously contributes to EBNA2 occupancy in LCLs, we conclude that our CBF1 deficient B cell line is a valid model system to study mechanisms which drive EBNA2 chromatin interactions.
To better characterize CBF1 dependent and independent EBNA2 binding sites prior to EBNA2 binding we could use H3K4me1, H3K4me3, and H3K27ac ChIP-Seq data published for DG75 [29]. Signal intensities of H3K4me1, H3K4me3, and H3K27ac were reanalyzed separately plotted for the CBF1 dependent and independent peak subpopulations and compared to the mean signal peak intensities of the respective chromatin modification in DG75 (Fig 4). All three activation marks showed almost the same high enrichment profiles for both subpopulations indicating that chromatin signatures are most probably not the trigger either for CBF1 dependent or independent binding.
To further investigate CBF1 independent EBNA2 binding to chromatin, de novo motif enrichment analyses of the two subclasses of EBNA2 binding sites were performed separately. Strikingly, the motif of EBF1, an important player in B cell development, was identified as the only and also highly enriched TF motif in the CBF1 independent EBNA2 peak subset, while CBF1 and EBF1 motifs as well as a CBF1/EBF1 composite core motif, show up in the top five motifs of the CBF1 dependent EBNA2 peak set (Fig 5A). In order to look at peak sets of similar size, 243 out of 1546 CBF1 dependent peaks were randomly selected and re-analyzed. For this reduced set, only the CBF1 and EBF1 motifs were significantly enriched. Since the majority of EBNA2 binding sites are also present in LCLs, we could use publicly available ChIP-seq data for EBF1 in LCLs to investigate EBF1 enrichment at CBF1 independent compared to dependent sites (Fig 5B, 5C and S8 Fig). Average CBF1 signal enrichment at EBNA2 binding sites did not significantly differ between CBF1 independent and dependent sites. However, the EBF1 signal was highly and significantly enriched at CBF1 independent compared to CBF1 dependent sites, indicating a potential role for EBF1 in mediating CBF1 independent EBNA2 binding to chromatin. Further quantitative correlation analyses focusing on signal intensities of EBNA2, CBF1, EBF1, and PU.1/SPI1 (Fig 5D and 5E) were performed to rank these co-occurring factors in a quantitative manner. PU.1/SPI1 was included since it had been suggested to serve as a DNA anchor for EBNA2 in the past. As expected, CBF1 showed the highest correlation in signal distribution with EBNA2 at EBNA2 peaks (rs = 0.46) as well as genome wide (rs = 0.5). Most strikingly, EBF1 highly correlated with EBNA2 signals at EBNA2 peaks (rs = 0.4) as well as genome wide (rs = 0.42). However, PU.1/SPI1 and EBNA2 signal intensities correlated weakly at EBNA2 peaks (rs = 0.19) as well as genome wide (rs = 0.17).
To test, if EBF1 can bind EBNA2, we performed co-immunoprecipitation (Co-IP) studies in DG75doxHA-E2 CBF1 wt and ko cells. After ectopic expression of EBF1 in DG75, an EBF specific antibody co-immunoprecipitated EBNA2 from cellular extracts in both CBF1 proficient as well as CBF1 deficient cells. This indicates that, CBF1 is not required for complex formation of EBF1 with EBNA2 (Fig 6A). The interaction of EBNA2 and CBF1 is well characterized. Two tryptophan residues (WW) within conserved region 6 are absolutely critical for EBNA2 to bind to CBF1 [4]. In order to test, whether the same region might also confer EBF1 binding, we generated DG75 cells expressing an EBNA2 WW325FF mutant (DG75doxHA-E2-WW). EBF1 was readily co-precipitated with EBNA2 WW325FF suggesting that EBF1 and CBF1 might target different regions of EBNA2 (Fig 6B).
The N-terminal region of EBNA2, comprising residues 1–58, appears to mediate multiple molecular functions including self-association and transactivation [31, 32]. We have recently described the three-dimensional structure of the EBNA-2 N-terminal dimerization (END) domain by heteronuclear NMR-spectroscopy. The END domain monomer comprises a small fold of four β-strands and an α-helix which form a parallel dimer by interaction of two β-strands from each protomer [33]. To further delineate the EBNA2 region that mediates the interaction with EBF1, we expressed glutathione S-transferase (GST) -END domain fusion proteins in bacteria and used the purified recombinant GST-END proteins as baits to affinity capture EBF1. GST-END specifically pulled down EBF1 in EBF1 transfected cells proving that the N-terminal domain of EBNA2 is sufficient to bind to EBF1 (Fig 6C).
Since CBF1 was neither required nor inhibitory for EBF1/EBNA2 complex formation, we asked if EBNA2 needs EBF1 to bind to either CBF1 independent or dependent chromatin sites. To this end, EBF1 protein levels were strongly reduced by siRNA mediated knock down (S9 Fig). EBF1 and EBNA2 binding to chromatin was tested by ChIP followed by quantitative PCR (ChIP-qPCR) for six selected enhancer loci, three CBF1 independent and three CBF1 dependent (Fig 7) sites, which also bind CBF1 and EBF1 in LCLs. While EBNA2 binding to CBF1 independent peaks was significantly reduced after EBF1 knock-down, CBF1 dependent EBNA2 binding was not significantly changed at reduced EBF1 levels. Thus, although EBF1 can bind to CBF1 dependent peaks it does not contribute to EBNA2 recruitment in this context.
Despite the ubiquitous expression of its anchor protein CBF1, EBNA2 is preferentially recruited to B cell specific enhancers and super enhancers [10, 11, 20, 34, 35]. The underlying mechanism that recruits EBNA2 specifically to these sites in B cells is still not understood and hard to study in the constitutive presence of CBF1. Since it was expected and also shown by other labs that CBF1 knock-down is not compatible with long term proliferation of LCLs [16, 17], we used a CBF1 deficient EBV negative B cell line to study whether EBNA2 can activate cellular genes and bind to chromatin in the absence of CBF1. This CBF1 deficient B cell line had been generated by targeted homologous recombination in DG75, a somatic cell line derived from an EBV negative Burkitt's lymphoma [36]. The proliferation of DG75 cells is driven by the reciprocal t (8;14) translocation which hyper-activates c-MYC expression and which renders proliferation of this cell line CBF1 independent [21].
EBNA2 target gene expression has been intensively studied but due to the omnipresence of CBF1, CBF1 independent target genes, direct or indirect, have never been discovered [21, 22, 37–46]. Unfortunately, a direct comparison of target gene lists across all studies and biological systems that have been published would be misleading, since different methodologies, thresholds and time points were applied. In addition, the use of different gene array systems does not allow the re-analysis of primary data sets as we have done for ChIP-Seq results taken from published data sets. Under these circumstances, the absence of evidence does not provide evidence for absence.
The comparison, however, can be made for selected EBNA2 target genes which were analyzed under similar conditions. While some target genes were identified in all studies, others appear specifically in distinct B cell lines as exemplified by the EBNA2 target gene CXCR7 which is induced in LCLs and BL41, a Burkitt's lymphoma cell line, but not in BJAB, a human lymphoblastoid B cell line [22, 46]. These findings suggest that activation of a subset of EBNA2 target genes requires specific cellular factors which, unlike CBF1, are not ubiquitously expressed. The DG75 cell lines used here express extremely low levels of the cellular transcription factors IRF4 and BATF, which are both well expressed in LCLs (S7C Fig) and highly enriched at LCL unique EBNA2 binding sites (Fig 2D). In addition, chromatin signatures at enhancer positions that can be bound by EBNA2 are distinct for DG75 and naïve B cells (Fig 2C). Thus, EBNA2 target gene activation is fine-tuned by multiple factors in B cells. We expect that additional rate limiting transcription factors apart from EBF1 control EBNA2 functions. A comparative analysis of CBF1 proficient and deficient B cells with distinct transcription factor signatures will be required to identify these additional factors.
Our genome wide gene expression studies confirm previously described EBNA2 cellular target genes which are also induced in LCLs like CD21, SLAMF1, RHOH, HEY1 or CCR7 [22] and in addition identify novel cellular EBNA2 target genes including long non-coding RNAs and micro RNAs. Notably, EBNA2 also controls a smaller but well defined set of CBF1 independent target genes. A selection of targets was validated by qPCR and confirmed the robust regulation of targets in both cell lines proving a strong biological activity of EBNA2 in CBF1 deficient B cells.
It is important to note that EBNA2 not only activates a set of direct target genes but thereby initiates a cascade of secondary events, which are included in our target gene lists and in total reflect EBNA2 functions. CBF1 dependent induced targets were strongly enriched for biological processes involved in immunoglobulin receptor binding functions and a broad array of enzymatic activities. While CBF1 independent EBNA2 induced targets were not significantly enriched for any biological processes, repressed and CBF1 independent targets could be assigned to multiple biological processes involving immune responses. Some of these repressed B cell specific genes like CD79A/mb1, CD79B/B29, VpreB3 have been described previously [21, 22, 47]. These targets are well characterized EBF1 induced target genes in mice [19, 48–51] and have been confirmed in human cells [52]. Recently, it has been demonstrated that EBNA2 promotes the formation of new CBF1 and EBF1 chromatin binding sites [17]. We speculate that EBNA2 might redirect EBF1 to novel chromatin sites and thereby deplete EBF1 activities required for target gene activation.
Several lines of evidence support a dynamic model for CBF1/DNA complex formation. Rather than functioning as a pre-bound DNA anchor, this dynamic model suggests that CBF1 is recruited to its DNA binding sites when complexed to cellular or viral binding partners. Notch [53], EBNA2 [17, 44], the EBV viral protein EBNA3C [54] and also RTA [55], the KSHV derived CBF1 binding protein, all promote CBF1/chromatin complex formation and influence chromatin site recognition. We propose that additional tissue-specific cellular or viral factors guide CBF1 associated activator or repressor proteins to functional regulatory elements in the cell.
Our genome-wide EBNA2 ChIP-Seq studies revealed that EBNA2 can bind to chromatin in a CBF1 independent manner. We used publicly available information on transcription factor occupancy in LCLs or peripheral human B cells to characterize different subpopulations of EBNA2 binding sites: i) EBNA2 binding sites shared by or unique to either LCLs or DG75 and ii) CBF1 independent and dependent binding sites. The total number of EBNA2 binding sites found in DG75 cells was significantly smaller than the number of binding sites found in LCLs, although EBNA2 was expressed abundantly in DG75 transfectants. Most EBNA2 binding sites initially identified in DG75 cells were shared with LCLs (S8 Fig). In LCLs, CBF1 independent binding sites score as strong EBNA2 binding sites (Fig 3D and 3E) and EBF1 is significantly enriched (Fig 5B and 5C).
In silico transcription factor binding analysis predicted CBF1, EBF1 and MEF2 to be bound at CBF1 dependent binding sites while CBF1 independent EBNA2 binding sites were predicted to bind EBF1 only. Thus these latter binding sites might have low affinity for CBF1 suggesting that EBF1 might be a B cell specific chromatin co-factor for EBNA2, which enhances complex formation also in CBF1 proficient LCLs and DG75 at sites with low affinity for CBF1. Interestingly, we observed that EBNA2 can enhance EBF1 expression and thus might further support complex formation (S1 Table).
For our study, we re-analyzed publicly available primary data sets and correlated signal intensities of transcription factors either at a genome wide level or by focusing on EBNA2 binding sites. These quantitative correlation studies on CBF1, PU.1/SPI1, EBF1, and EBNA2 signal intensities revealed a strong positive correlation of CBF1 and EBF1 to EBNA2 and weak correlation of CBF1 and EBF1 to each other. Surprisingly, PU.1/SPI1 binding activity correlated with neither EBNA2, nor CBF1 nor EBF1 binding activity (Fig 5D and 5E). A physical interaction of PU.1/SPI1 and EBNA2 has been described, but was never characterized in detail [56, 57]. Transient promoter reporter studies had previously suggested that both, PU.1/SPI1 and CBF1 are critical for transactivation of the viral LMP1 promoter by EBNA2 [13, 15, 58]. However, inactivation of the PU.1/SPI1 binding site at the LMP1 promoter in the viral genome did not grossly change the transformation potential of the viral mutants. LMP1 expression and proliferation was diminished but not abolished while inactivation of the EBF1 binding site ablated LMP1 expression [16]. To date, there is no experimental proof indicating that EBNA2 is recruited to chromatin by PU.1/SPI1 [17]. If the pioneer factor PU.1/SPI1 does not serve as chromatin anchor for EBNA2, it could facilitate the access of transcription factors to compacted chromatin or prevent chromatin silencing at the respective enhancer regions [59].
In order to define the contribution of EBF1 to EBNA2 chromatin binding, EBF1 protein expression was downregulated by siRNA. These knock down experiments proved that EBNA2 needs EBF1 to bind efficiently to CBF1 independent chromatin sites in both, CBF1 proficient and deficient cells. In contrast, EBNA2 binding to CBF1 dependent sites was not impaired by EBF1 siRNA knock down and thus was defined to be EBF1 independent although EBF1 is present (Fig 7B). EBF1 and EBNA2 binding is consistently weaker in CBF1 deficient compared to CBF1 proficient DG75 cells (Fig 7). Surprisingly, EBF1 binding is elevated at CBF1 independent sites in CBF1 proficient LCLs (Fig 5). Thus, CBF1 might contribute to the assembly of EBNA2/EBF1 complexes on chromatin, a concept which is consistent with findings of Lu et al. (16), describing EBF1/CBF1 co-occupied binding sites which are preferentially formed in the presence of EBNA2.
Here we show that EBNA2 and EBF1 can form complexes in cells and thus provide the first evidence that EBF1 interacts with a viral protein. Only a few cellular binding partners of EBF1 have been described so far. EBF1 can bind DNA as a homodimer [60], but can further interact and cooperate with other transcription factors like MEF2C [61], the deoxygenase TET2, an enzyme involved in the DNA demethylation process [62], or the histone acetyltransferase CBP [63]. EBF1 also binds to CNOT3, a subunit of the CCR4-NOT complex [64] which regulates multiple steps in RNA metabolism including transcription, nuclear RNA export and RNA decay [65], and thereby also modulates target gene profiles of EBF. In addition, two multi-zinc finger proteins, ZNF423 and ZNF521, antagonize the biological activity of EBF1 and thereby might promote tumorigenesis [66]. It should be mentioned that in B cells, with a single exception (CNOT3), these interactions have been described after expressing at least one binding partner ectopically or using cross-linking reagents before co-immunoprecipitations have been performed [61]. Thus, it appears that EBF1 protein-protein interactions are particularly difficult to detect at endogenous expression levels in B cells. To date we and others have tried and failed to detect EBNA2/EBF1 complexes expressed at endogenous levels in LCLs while EBNA2/CBF1 complexes could be readily detected in LCLs [17]. Here we detect EBF1/EBNA2 complexes after overexpression of both binding partners in cells. Importantly, the purified END domain consisting of 58 amino acids of EBNA2 is sufficient to specifically affinity capture EBF1 from cellular extracts.
Future studies on purified proteins of both binding partners will reveal whether the interaction of EBNA2 and EBF1 is direct or whether so far unknown factors, proteins or DNA, support complex formation. If additional cellular factors promote complex formation, they need to be expressed at very high levels in the cell to efficiently bridge viral and cellular proteins.
In summary, the genetic ablation of CBF1 expression in B cells provides novel valuable insights into the molecular mechanisms of EBNA2 activity. At this point of our study, we can define EBNA2 functions in the absence of CBF1. Chromatin conformation capture techniques performed in DG75 cells will be required to link EBNA2 binding sites to the respective target genes. Since EBNA2/EBF1 complex formation could be demonstrated in CBF1 proficient and deficient cells and EBF1 and CBF1 bind to different regions of EBNA2, heterotrimeric complexes might be formed. Whether these complexes activate or repress transcription might depend on their composition and the chromatin context of enhancer and promoters they bind to. Any working hypothesis to be tested will have to take into account the dimeric nature of EBNA2 and EBF1 as well as the fact that CBF1 and EBF1 are co-expressed and their binding motifs might overlap [67]. Our future studies will need to explore the architecture of these complexes in order to understand if pre-formed EBNA2/CBF1 complexes can use EBF1 to guide EBNA2 to B cell specific enhancers and thereby provide B cell specificity to EBNA2 activities.
pcDNA3 (pCDNA3) and EBF1-myc expression plasmid (pCDNA3.EBF1-5xmyc) were kindly provided by Mikael Sigvardsson [68]. pCKR74.2 is a Dox (doxycycline) inducible HA- (haemagglutinin) tagged EBNA2 expression plasmid (pCKR74.2) based on pRTR [69, 70].
The cells were maintained as suspension cultures in RPMI 1640 medium (Gibco Life Technologies) supplemented with 10% FCS (fetal calf serum, Bio&Sell), 4 mM L-Glutamine and 1 x penicillin/streptomycin (Gibco Life Technologies). 721 is an EBV positive LCL cell line [71]. The DG75 ko cell line (SM224.9), DG75ER/EBNA2 CBF1 wt and ko cells (SM295 and SM296) have been described before [21, 22]. The ER/EBNA2 (estrogen receptor hormone binding domain EBNA2) fusion protein was activated by cultivating the cells in cell culture medium supplemented with 1 μM ß-estradiol. The DG75doxHA-E2/CBF1 wt (CKR128-34) and the DG75doxHA-E2/CBF1 ko (CKR178-10) cell lines carry the Dox inducible HA-EBNA2 expression plasmid (pCKR74.2). DG75doxHA-E2WW/CBF1 wt (CKR436) expresses a Dox inducible HA-EBNA2 WW325FF mutant (pCKR421). They were cultivated in 1 μg/ml puromycin containing media. EBNA2 expression was induced by doxycycline treatment (1μg/ml).
Total RNA was extracted from 1x107 cells induced for 24 h with 1 μM ß-estradiol using the Qiagen RNeasy Mini Kit. Expression analysis starting from 100 ng of total cellular RNA was performed using the Ambion WT Expression Kit (Applied Biosystems) and subsequently the GeneChip WT Terminal Labeling and Hybridization Kit (Affymetrix) followed by the GeneChip Human Gene 2.0 ST array (Affymetrix) according to the manufacturer's protocol. All affymetrix CEL files have been processed in Bioconductor/R using robust multiarray average (RMA) for normalization and summarization and limma for differential expression and significance. Quality has been checked using the array QualityMetrics package. Additional filtering based on the fold change between the two conditions was applied with different stringency, individually described in the legend of the tables and figures. Analyzation and visualization of the Microarray was performed using Genesis, available at http://genome.tugraz.at. Quantitative RT-PCR analysis was performed as described previously [72]. Primers used for RT-qPCR were designed applying Primer3 software (http://primer3.ut.ee/) and selection of mature transcripts was ensured by amplification across exon-exon junctions. Primers used for quantitative RT-PCR are summarized in S3 Table. All data were normalized for the relative abundance of the Actin B transcript.
GOrilla is a tool to identify and visualize enriched GO terms in ranked lists of genes (http://cbl-gorilla.cs.technion.ac.il/) [73]. Enrichment is defined as E = (b/n) / (B/N), with N = the total number of genes, B = the total number of genes associated with a specific GO term, n = the number of genes in the top of the user's input list and b = the number of genes in the intersection. The threshold for n is selected by GOrilla by maximizing E and statistical significance is computed taking into account the multiple hypothesis tests arising due to the maximization.
All GO terms for which B < 10 were ignored. GO terms with a q-value (FDR) ≤ 10−4 were selected and ranked for their enrichment score given by GOrilla.
As induction and repression was stronger in DG75ER/EBNA2 CBF1 wt cells than in CBF1 ko cells, principal component analysis (PCA) was used to identify genes regulated on average or differentially between wt and ko (S6 Fig). PCA was performed for all genes significantly regulated in CBF1 wt or ko cells (limma q < 0.01). The first principal component corresponded to average regulation while the second principal component represented CBF1 dependence. Genes were first ranked according to the first principal component, i.e. top entries corresponded to genes that were induced on average in CBF1 wt and ko cells. This was repeated after reversing the list to analyze genes repressed on average. Furthermore, from each of these two lists, the top 2000 genes were selected and both were ranked according to the second principal component. Both lists were additionally reversed. Therefore, in these four additional lists, genes that are either induced or repressed on average were ranked according to their degree of CBF1 dependence.
1x107 DG75doxHA-E2/CBF1 wt or ko cells were lysed in 500 μl NP-40 lysis buffer (1% NP-40, 150 mM NaCl, 10 mM Tris-HCL pH 7.4, 1mM EDTA pH 8.0, 3% Glycerol) supplemented with complete protease inhibitor cocktail (Roche) for 1h (30 min rolling at 4°C, 30 min on ice). Precleared protein lysates were used for co-immunoprecipitation by adding 100 μl of hybridoma supernatant (E2: α-HA R1 3F10; E.Kremmer) or 1 μg of purified antibody (α-EBF Santa Cruz Biotechnology, sc-137065) at 4°C under rotation overnight. Subsequently, 50 μl of 50% suspension of pre-blocked, equilibrated protein G-coupled Sepharose beads (GE Healthcare) were added to the lysates and incubated for 2h at 4°C under rotation. Immunoprecipitates were washed 5 times with NP-40 lysis buffer, Laemmli buffer was added to the beads, and the samples were boiled, submitted to electrophoresis by SDS-PAGE and analyzed by immunoblotting.
5x 106 cells were lysed in 200 μl NP-40 lysis buffer (1% NP-40, 150 mM NaCl, 10 mM Tris-HCL pH 7.4, 1mM EDTA pH 8.0, 3% Glycerol) for 2 h on ice. 30 μg of total cell lysate were submitted to SDS-PAGE under reducing conditions. Immunoblotting was performed on polyvinylidene difluoride (PVDF) membranes. Western blots were probed with the following primary antibodies: rat α-EBNA2 (R3; IgG2A; E. Kremmer), rat α-CBF1 (RBP-J 7A11, E. Kremmer), rat α-GST (GST 6G9, IgG2A, E. Kremmer), mouse α-EBF (Santa Cruz Biotechnology, sc-137065), goat α-BATF (B-ATF H-19, Santa Cruz Biotechnology, sc-15280), rabbit αIRF4 (IRF4H-140), and-GAPDH (EMD Millipore MAB374). HRP-coupled secondary antibodies (Santa Cruz Biotechnology) and an ECL kit (GE Healthcare) were used for visualization. For subsequent quantification of protein levels, exposed films were scanned in transmission mode and protein band intensities were determined by densitometry using ImageJ software (http://rsbweb.nih.gov/ij/) [74].
5x 106 DG75 cells were transfected by electroporation at 250 V and 950 μF in 250 μl reduced serum media (Opti-MEM, Gibco Life Technologies; without supplements) using 0.4 cm-electrode-gap cuvettes (Bio-Rad) and the Bio-Rad Gene Pulser.
5x 106 cells were transfected with 100 pmol control siRNA-A or EBF1 siRNA (both Santa Cruz Biotechnology, sc-37007 and sc-10695) by electroporation. 24 h after transfection, 1x 107 induced, siRNA treated cells were harvested for chromatin isolation and 5x106 cells for protein isolation.
This ChIP protocol is based on reference (59) with minor modifications as indicated below. In brief, 2x 107 DG75doxHA-E2 cells were harvested and washed twice in ice cold PBS, resuspended in 20 ml RPMI 1640 (Gibco Life Technologies) and formaldehyde (1% final) was added for cross-linking. The reaction was stopped by addition of glycine (125 mM final) after 7 min and gentle shaking for 5 min at RT. Cells were pelleted and washed twice in ice cold PBS. Nuclei were isolated by washing the cells 3x with 10 ml of ice cold Lysis Buffer (10 mM Tris-HCl, pH 7.5, 10 mM NaCl, 3 mM MgCl2, 0.5% NP-40, 1x proteinase inhibitor cocktail (PIC, Roche)) and subsequent centrifugation (300 g for 10 min at 4°C). Nuclei were resuspended in 1 ml Sonication Buffer (50 mM Tris-HCl, pH 8.0, 10 mM EDTA, pH 8.0, 0.5% SDS, 1x PIC) and incubated on ice for 10 min. Chromatin was sheared to an average size of 200–300 bp by four rounds of sonication for 10 min (30 sec pulse, 30 sec pause) using a Bioruptor device (Biogenode). Cell debris was separated by centrifugation at maximum speed for 10 min at 4°C and chromatin containing supernatants were stored at -80°C or directly used for IP. To prepare input DNA, 25 μl aliquots (1/10 of the amount used per IP) were saved at -80°C. For IPs 250 μl chromatin (equals 5x 106 cells) were diluted 1:4 with IP Dilution Buffer (12.5 mM Tri-HCl, pH 8.0, 212.5 mM NaCl, 1.25% Triton X-100, 1 x PIC) and incubated with 100 μl of hybridoma supernatant on a rotating platform at 4°C overnight. A combination of EBNA2 and HA-tag specific antibodies (⅓ α-E2 R3 (rat IgG2a, ⅓ α-E2 1E6 (rat IgG2a), and ⅓ α-HA R1-3F10 (rat IgG1)) was used to precipitate EBNA2 and an isotype-matched unspecific antibody mixture (⅔ α- GST 6G9 (rat IgG2a) and ⅓ α-CD23 Dog-CD3 (rat IgG1) both by E. Kremmer) was used as isotype control. The EBF antibody (C-8) (sc-137065, Santa Cruz Biotechnology) was used to precipitate EBF1 and an antibody specific for ovalbumin (M-Ova 3D2, E. Kremmer) was used as an isotype control. Protein G sepharose (GE Healthcare) was equilibrated with IP Dilution Buffer, added to the lysate and incubated at 4°C for 4 h with constant rotation. Beads were extensively washed with: 2x Wash Buffer I (20 mM Tris-HCl, pH 8.0, 2 mM EDTA, pH 8.0, 1% Triton X-100, 150 mM NaCl, 0.1% SDS, 1x PIC), 1x Wash Buffer II (20 mM Tris-HCl, pH 8.0, 2 mM EDTA, pH 8.0, 1% Triton X-100, 500 mM NaCl, 0.1% SDS, 1x PIC), 1x Wash Buffer III (10 mM Tris-HCl, pH 8.0, 1 mM EDTA, pH 8.0, 250 mM LiCl, 1% NP-40, 1% sodium deoxycholate, 1x PIC) for 5 min under rotation, and 2x with TE (10 mM Tris-HCl, pH 8.0, 1 mM EDTA, pH 8.0) for 1 min. Protein-DNA complexes were eluted with 2x 150 μl Elution Buffer (25 mM Tris-HCl, pH 7.5, 10 mM EDTA, pH 8.0, 1% SDS) at 65°C for 15 min. Input samples were adjusted to 300 μl with Elution Buffer. Eluates and input samples were incubated with Proteinase K (1.5 μg/μl final, Roche) for 1 h at 42°C. Cross-linking was reversed by incubation at 65°C overnight. DNA was recovered using QIAquick PCR purification kit (Qiagen).
The EBNA2 specific ChIP in LCL was performed as described above with the following modifications: Protein-protein interactions were fixated by adding disuccinimidyl glutarate (DSG, Pierce #20593, 2 mM final, using freshly prepared 0.5 M stock solution in DMSO) for 23 min at RT and prior to formaldehyde (1% final) cross-link for additional 7 min. Sonication Buffer was composed of 50 mM Tris-HCl, pH 8.0, 5 mM EDTA, pH 8.0, 0.5% SDS, 0.5% Triton X-100, 0.05% sodium deoxycholate, and 1x PIC. IP Dilution Buffer was composed of 12.5 mM Tri-HCl, pH 8.0, 187.5 mM NaCl, 1.25 mM EDTA, pH 8.0, 1.125% Triton X-100, and 1 x PIC. For EBNA2 specific IP 50 μl of α-E2 R3 (rat IgG2a) and 50 μl α-E2 1E6 (rat IgG2a) hybridoma supernatant were applied and the same volume of isotype-matched nonspecific antibody (α- GST 6G9 (rat IgG2a) E. Kremmer) was used as negative control.
For sequencing purposes DNA concentration was measured using the Qubit dsDNA HS Assay Kit (Thermo Fisher). A maximum of 100 ng ChIP or input derived DNA were used for library preparation (NEBNext Ultra DNA Library Prep Kit for Illumina) and subsequently subjected to deep sequencing using a HiSeq 1500 device (Illumina).
The amount of recovered DNA in input samples and after IP with specific antibody or an unspecific isotype-matched IgG control was quantified by qPCR using primers listed in S3 Table.
qPCR was performed using LightCycler 480 SYBR Green I Master (Roche) on a LightCycler 480 II instrument (Roche) as described previously [72]. 2 technical replicates were analyzed for each biological replicate. Amplification was always conducted at 63°C. To account for differences in amplification efficiencies a standard curve was generated for each primer pair using serial dilutions of sheared DNA (input) as template. DNA quantities detected in input samples were adjusted to the amount of chromatin used per IP by multiplication with 20. Values obtained from IP samples with unspecific IgG control were subtracted from the DNA amounts recovered by IP with specific antibody. The percent of input was calculated as (DNA from specific IP corrected for IgG control background/ DNA input) x 100. To validate the ChIP, qPCR at a known (ChIP-Seq) positive locus was performed. To compromise divergent EBNA2 inducibility in wildtype and knockout cells, the percent input was calculated relative to a known negative locus (ChIP-Seq; percent input at tested locus/percent input of known negative locus). To display the change in binding, the mean relative input of the wildtype cells treated with control siRNA was set to one. A paired t-test was performed to assess significance of differences of means.
Expression plasmids were transformed into E.coli strain BL21. Bacteria were cultured in 400 ml of LB medium containing antibiotics at 37°C until an OD of 0.5–0.7 was reached. Expression of proteins was induced with 1 mM IPTG for 3 h at 30°C. After induction, bacteria were suspended in 20 ml ice cold binding buffer (25 mM HEPES, pH 7.6, 0.1 mM EDTA, pH 8, 12.5 mM MgCl2, 10% Glycerol, 0.1% NP-40, 100 mM KCl, 1 mM PMSF, 1 mM DTT, 1 mM PMSF) and lysed by sonication. Lysates were cleared by centrifugation at 48,000 x g at 4°C for 20 min. Glutathione Sepharose 4B beads (GE Healthcare) were washed and resuspended in binding buffer to prepare a 50% slurry. To coat the beads with GST or GST fusion protein, 100 μl of the 50% slurry were incubated with the 20 ml of cleared lysates for 1 hat 4°C and washed 3 times with 20 ml binding buffer.
1x107 DG75 cells were transfected with EBF1 expression plasmids or empty vector controls. 24 h after transfection, cells were harvested and lysed in 500 μl lysis buffer (50 mM HEPES, pH 7.6, 5 mM EDTA, pH 8, 150 mM NaCl, 0.1% NP-40, 1 mM PMSF) followed by sonication. Cell lysates were centrifuged for 15 min at 16,000x g, 4°C, and the protein concentration was measured by Bradford assay. To pull down EBF1, the supernatants were incubated with the GST or GST fusion protein coated beads for 3 hat 4°C. Subsequently, beads were washed 5 times with binding buffer and protein complexes were dissolved in 2x Lämmli buffer (4% SDS, 20% Glycerol, 120 mM Tris/HCl, pH 6.8, 5% β-Mercaptoethanol, Bromphenol-blue). Samples were analyzed by SDS-PAGE and Western Blot.
All bioinformatic analyses of ChIP-Seq data were conducted by using the galaxy bioinformatics platform [75] hosted and maintained by the Bioinformatics Department of the University of Freiburg. For all sequenced samples, at least 17 million reads were obtained and biological duplicates of EBNA2 ChIP and input samples were sequenced. Reads were mapped to the human genome using Bowtie2 [76]. For all samples, at least 95% of reads were mappable to the human genome including at least 69% of uniquely mapping reads with one distinct location (S4 Table). Biological duplicates of mapped reads were merged and subsequently significant EBNA2 binding sites were identified using MACS2 [77] by normalizing ChIP to input samples (S4 Table). In a second step, the peaks were further filtered and “negative peaks” (negative amplitude, significantly higher read count in the input sample), peaks located at black-listed regions [78], peaks with a very low enrichment score, and such located on chromosomes not included in the ENCODE data for GM12878 (e.g. chrY, chrUn) were excluded (S4 Table). Normalized EBNA2 ChIP signal tracks were generated by subjecting duplicate-merged ChIP and input read files to bamCompare of the deepTool package [79] and normalizing ChIP to input samples by subtraction as well as normalizing to fragments (reads) per kb per million (RPKM) to account for genome coverage. Mean signal intensities at specific peak sets were calculated using computeMatrix of the deepTools package. This workflow for transcription factor peak calling and signal track generation was applied to all ChIP-seq data sets analyzed in this manuscript. A separate workflow for the analysis of histone modification ChIP-seq data was generated to account for the typical broader signal distribution and applied to all such data sets analyzed in this study. Data provided by public resources were reanalyzed using the same pipeline as described above and references are listed in S5 Table. The details of all analyses steps are captured in a Galaxy workflow which can be downloaded at github (https://github.com/bgruening/galaxytools/tree/master/workflows/peak_calling) and re-run and analyzed in Galaxy. Cluster analysis was performed using the k-means algorithm tool (numbers of clusters expected = 12, max. iterations = 50) provided by Genesis (release 1.7.7), available at http://genome.tugraz.at. Genesis was also used to generate heatmaps [80].
Inducibility of EBNA2 expression in DG75doxHA-E2/CBF1 wt and ko cell lines was evaluated by monitoring the expression of the eGFP surrogate marker of pCKR74.2. Cells were induced for 16 h or 24 h with doxycycline, washed and fixed with 0.5% PFA in PBS. For quantification of induced cells, the FACSCalibur system (BD Biosciences) and CellQuest Pro software (BD Biosciences) were applied.
|
10.1371/journal.pmed.1002119 | Sex Differences in Tuberculosis Burden and Notifications in Low- and Middle-Income Countries: A Systematic Review and Meta-analysis | Tuberculosis (TB) case notification rates are usually higher in men than in women, but notification data are insufficient to measure sex differences in disease burden. This review set out to systematically investigate whether sex ratios in case notifications reflect differences in disease prevalence and to identify gaps in access to and/or utilisation of diagnostic services.
In accordance with the published protocol (CRD42015022163), TB prevalence surveys in nationally representative and sub-national adult populations (age ≥ 15 y) in low- and middle-income countries published between 1 January 1993 and 15 March 2016 were identified through searches of PubMed, Embase, Global Health, and the Cochrane Database of Systematic Reviews; review of abstracts; and correspondence with the World Health Organization. Random-effects meta-analyses examined male-to-female (M:F) ratios in TB prevalence and prevalence-to-notification (P:N) ratios for smear-positive TB. Meta-regression was done to identify factors associated with higher M:F ratios in prevalence and higher P:N ratios. Eighty-three publications describing 88 surveys with over 3.1 million participants in 28 countries were identified (36 surveys in Africa, three in the Americas, four in the Eastern Mediterranean, 28 in South-East Asia and 17 in the Western Pacific). Fifty-six surveys reported in 53 publications were included in quantitative analyses. Overall random-effects weighted M:F prevalence ratios were 2.21 (95% CI 1.92–2.54; 56 surveys) for bacteriologically positive TB and 2.51 (95% CI 2.07–3.04; 40 surveys) for smear-positive TB. M:F prevalence ratios were highest in South-East Asia and in surveys that did not require self-report of signs/symptoms in initial screening procedures. The summary random-effects weighted M:F ratio for P:N ratios was 1.55 (95% CI 1.25–1.91; 34 surveys). We intended to stratify the analyses by age, HIV status, and rural or urban setting; however, few studies reported such data.
TB prevalence is significantly higher among men than women in low- and middle-income countries, with strong evidence that men are disadvantaged in seeking and/or accessing TB care in many settings. Global strategies and national TB programmes should recognise men as an underserved high-risk group and improve men’s access to diagnostic and screening services to reduce the overall burden of TB more effectively and ensure gender equity in TB care.
| Global health initiatives have tended to treat “gender” issues in health as being synonymous with women’s health. However, for infectious diseases, policy and practice need to be guided by epidemiological data and consideration of transmission dynamics.
Many more men than women are diagnosed with, and die from, tuberculosis (TB) globally.
Data from population-level surveys for undiagnosed TB, carried out in a number of countries during the last two decades, can be combined with data on diagnosed (notified) cases to provide more complete insight into the magnitude and nature of sex differences in TB.
Surveys conducted to identify adult cases of TB in communities in low- and middle-income countries between 1993 and 2016 were analysed by sex.
TB prevalence among men was over twice as high as among women and was substantially higher even in settings with high HIV prevalence.
Case notification rates were also higher for men, and the ratio of prevalent-to-notified cases of TB—an indication of how long patients take to be diagnosed, on average—was 1.5 times higher among men than women, suggesting that men are less likely than women to achieve a timely diagnosis.
Given that undiagnosed TB is the key driver for transmission in communities, our data show that greater effort and investment are needed to improve awareness of TB in men as an individual and public health issue.
Policies on gender and TB should place greater emphasis on the high burden of disease in men and the need to invest in male-friendly diagnostic and screening services, with the aim of reducing undiagnosed TB.
| Over the past twenty years, tuberculosis (TB) case notifications among men have exceeded those among women in most settings [1]. In 2014, the male-to-female (M:F) ratio in smear-positive pulmonary TB case notification was 1.7 globally and ranged from 1.0 in the Eastern Mediterranean Region to 2.1 in the Western Pacific Region [2]. The excess of notified cases among men has often been explained as a result of barriers faced by women in seeking care for and being diagnosed with TB [3,4]. However, notification data alone are insufficient to determine whether this is true, or whether sex differences in case notifications reflect an excess in the burden of disease among men and even a disadvantage among men in seeking and accessing TB care.
Prevalence surveys offer a robust measure of disease burden in the community, reducing or eliminating the care-seeking biases that affect case notifications: a higher proportion of men in case notifications could reflect either higher incidence of TB disease or more complete registration for treatment by men. Prevalence surveys predominantly identify infectious TB patients with previously undiagnosed TB disease who have, therefore, not contributed to routine notification data before participation in the survey. As such, comparison of the characteristics of diagnosed TB patients (notification data) with those of undiagnosed TB patients (prevalence survey data) provides a unique insight into diagnosis and treatment access barriers. For example, finding a similar male predominance in undiagnosed TB (prevalence surveys) patients as in notified TB cases would support the hypothesis that men genuinely have a higher burden of TB disease, while finding a greater male predominance in undiagnosed TB patients than in notified TB cases would suggest male-specific access barriers or male sex being a risk factor for TB disease.
A previous analysis in 2000 found that male TB prevalence exceeded female TB prevalence in 27 (93%) of 29 prevalence surveys conducted in 14 countries between 1953 and 1997 [5]. The same analysis calculated the patient diagnostic rate (the inverse of the prevalence-to-notification ratio) and found that female cases were more likely to be notified than male cases in 21 (72%) surveys.
Despite these findings, men are often overlooked in discussions of gender and TB. While global TB reports and meetings on gender acknowledge the fact that the majority of TB cases and TB-associated deaths occur among men, greater focus is usually placed on women. More broadly in global health discussions, there is a tendency to use the word “gender” when really “women” is meant, as exemplified by the Millennium Development Goals [6] and Sustainable Development Goals [7]. Subsequently, an emphasis on men runs contrary to global norms [8], and strategies to assess and address men’s barriers to TB care are notably absent from the global research agenda.
The World Health Organization’s End TB Strategy emphasises the importance of equity in access to diagnosis and treatment [9]; men should not be excluded from this target. The End TB Strategy has also prioritised systematic screening of high-risk groups to ensure early diagnosis of individuals with TB [10]. If TB prevalence remains higher among men than women, as in previous analysis [5], men should be considered a high-risk group for TB [11], and national TB programmes should more actively target men with routine diagnostic and/or screening services. This action is necessary to reduce the burden of TB in the whole population more effectively [12] and to ensure that principles of gender equity are upheld.
This review set out to systematically investigate sex differences in the prevalence of bacteriologically positive TB and smear-positive TB in adult participants in cross-sectional surveys conducted in low- and middle-income countries to determine whether sex ratios in adult case notifications reflect population sex differences in disease and to compare prevalence-to-notification (P:N) ratios for men and women. The current study adds to previous analysis [5] by including surveys conducted since the widespread availability of anti-TB chemotherapy in low-resource settings and the implementation of the directly observed treatment short course (DOTS) strategy, as well as the rise of the HIV/AIDS pandemic and the implementation of interventions against it—all factors that may have different effects on TB in men and women. The current study also provides more detailed meta-analyses of sex differences in TB prevalence and P:N ratios.
In accordance with the published protocol [13], studies describing national and sub-national TB prevalence surveys in adult populations (age ≥ 15 y) in low- and middle-income countries published between 1 January 1993 and 15 March 2016 were identified through searches of PubMed, Embase, Global Health, and the Cochrane Database of Systematic Reviews (Table 1). The WHO Global Tuberculosis Report 2015 [2] and abstract books from the Union World Conference on Lung Health (2012–2015) were also searched by hand, as were the reference lists of included studies. Researchers in the field and at WHO were contacted to assist with identification of relevant studies.
Two authors (K. C. H. and P. M.) independently reviewed titles and abstracts in parallel to identify relevant studies for full-text review. A third author (E. L. C.) resolved any discrepancies. The same authors reviewed full texts to determine whether studies met inclusion criteria and then extracted data on study methodology and TB prevalence in parallel using piloted electronic forms.
Study authors were contacted for additional information if studies did not report the number of participants and the number of bacteriologically positive and/or smear-positive TB cases by sex for adult participants. Authors were also contacted if sex-specific prevalence data were not available by age group.
The review included cross-sectional prevalence surveys conducted in low- and middle-income countries [15]. Studies conducted among symptomatic or care-seeking individuals, children, individuals of a single sex, occupational settings, or other sub-populations (e.g., only HIV-positive individuals) were excluded. Studies reporting prevalence of Mycobacterium tuberculosis infection but not TB disease were excluded. Individuals under 15 y of age were excluded since diagnosis of childhood TB is more complicated than diagnosis of adult disease, especially within the context of community-based surveys [16]. Studies including both adults and children were included in the qualitative review but were excluded from quantitative analyses unless the study reported participation and prevalence for adults. Studies published in languages other than English were excluded due to limited resources for translation. Where more than one report was identified for a single survey, the most complete source was included and the others were excluded.
The risk of bias in included studies was assessed in parallel by K. C. H. and P. M. Each study was ranked on eight criteria from a tool developed to assess the risk of bias in prevalence surveys [17]. These criteria assessed factors related to the selection of the study population, the risk of nonresponse bias, data collection methods, and case definitions. The eight criteria were summarised to give an assessment of the overall risk of bias.
Study participants were defined as individuals who were interviewed and/or underwent initial screening procedures, according to study-specific procedures. Participation was defined as the number of participants divided by the number of individuals who were eligible or invited to participate. High relative male participation was defined as a M:F ratio in participation ≥ 0.90.
Case definitions for TB were based on internationally recognised terminology, where available, and study-specific definitions otherwise. Bacteriologically positive TB was defined as positive smear microscopy, culture, or WHO-approved rapid diagnostic results (such as from Xpert MTB/RIF) [18].
Sex-specific prevalence of bacteriologically and smear-positive TB was defined as the number of individuals with bacteriologically or smear-positive TB divided by the number of study participants, by sex. Reported prevalence was used to estimate the number of cases or the number of participants where one of these values was missing. No adjustments were made for nonparticipation or nonsampling.
Sex-specific P:N ratios were calculated as the ratio of smear-positive TB prevalence per 100,000 individuals to smear-positive TB case notifications per 100,000 individuals among adults [5,19]. WHO case notification data [20] and United Nations population estimates [21] were matched to each prevalence survey by country and year. For surveys that took place over more than one calendar year, the annual case notification rate was averaged over all survey years (excluding years with no reported data). No adjustments were made for sub-national surveys.
National estimates of TB and HIV burden were matched to each prevalence survey by country and year. For surveys that took place over more than one calendar year, estimates were averaged over all survey years (excluding years with no reported data). High TB prevalence was defined using the median value for included studies, which was an estimated national TB prevalence ≥ 300 per 100,000 individuals [22]. High HIV prevalence was defined as estimated national HIV prevalence ≥ 1% in the general population [23,24], and high HIV prevalence in incident TB was defined as estimated HIV prevalence ≥ 20% in new and relapse TB cases [22,25].
Prevalence of bacteriologically positive TB and smear-positive TB was calculated for included studies by sex. Prevalence of bacteriologically positive TB by sex and age was also calculated, where possible. Sub-group prevalence was estimated for sub-groups based on survey characteristics including WHO geographical region, survey setting (national versus sub-national), national estimates of TB and HIV burden (both in the general population; the latter also in incident TB), study quality, initial screening procedures, case definitions, and relative male participation. Clopper-Pearson confidence intervals [26] and M:F ratios were calculated for all prevalence estimates. P:N ratios for smear-positive TB were estimated with confidence intervals based on the estimated variance using a continuity correction of 0.5 in the corresponding prevalence estimates.
Heterogeneity was assessed using the I2 statistic [27]. Due to substantial heterogeneity between studies, random-effects models were used for meta-analyses, weighting for the inverse of the variance. Random-effects weighted summary M:F ratios were calculated for participation, prevalence of bacteriologically positive TB and smear-positive TB, age-specific prevalence of bacteriologically positive TB, and P:N ratios.
Meta-regression was performed for M:F ratios in prevalence and M:F ratios in P:N ratios to examine associations with the survey characteristics mentioned above, plus the starting year of each survey. Univariate meta-regression of M:F ratios in prevalence was conducted separately for bacteriologically positive TB and smear-positive TB. If either univariate meta-regression suggested evidence of an association with a particular characteristic, that characteristic was included as a variable in the multivariate meta-regression models for both bacteriologically positive and smear-positive TB. Similarly, multivariate meta-regression of M:F ratios in P:N ratios was based on evidence of associations in univariate analysis.
All analyses were performed using R version 3.2.2 [28] (S1 Data; S1 Analysis).
Of 7,502 potentially relevant English-language studies screened by title and abstract, 148 were reviewed in full; of these, 65 were excluded after full-text review (S1 Table) and 83 were eligible for inclusion (Fig 1; S2 Table) [29–111]. Included studies describe 88 surveys in 28 countries: 36 surveys in 13 countries in the African Region, three surveys in two countries in the Region of the Americas, four surveys in two countries in the Eastern Mediterranean Region, 28 surveys in five countries in the South-East Asia Region, and 17 surveys in six countries in the Western Pacific Region (Fig 2). There were 22 nationally representative surveys and 66 sub-national surveys, with at least 20 of the latter conducted in urban settings and eight among tribal populations. Over 3.1 million adult participants were included; 16 surveys did not report the number of adult participants.
The risk of bias assessment identified 33 (43%) surveys with low risk of bias, 32 (42%) with moderate risk of bias, and 12 (16%) with high risk of bias (S1 Fig). Eleven surveys for which only an abstract was available were characterised as unknown risk of bias due to limited information on study methodology [34,54,57,62,63,75,76,79,80,95,104]. The quantitative analyses included a slightly higher proportion of surveys with low risk of bias than the qualitative summary. In all, 84% to 94% of the surveys in the quantitative analyses had low to moderate risk of bias (S2 Fig).
Female participation equalled or exceeded male participation in all of the 28 surveys for which participation was reported by sex (Fig 3). Of 687,926 men eligible or invited to participate, 521,934 (75.9%) participated, while 611,901 (82.5%) of 741,705 eligible or invited women participated. The overall random-effects weighted M:F ratio in participation was 0.90 (95% CI 0.86–0.93; range 0.50 to 1.00).
The prevalence of bacteriologically positive TB was reported by sex in 56 surveys with 2.2 million participants in 24 countries [29,30,32,33,35,36,38–44,47–51,53,55,56,58–60,65–67,69–74,82,84,85,87,89–94,97,101,102,104,105,107,110–112]. Forty surveys with 1.7 million participants in 22 countries reported the prevalence of smear-positive TB by sex [35,40,43,44,48–51,53,55,56,58–60,65–67,69–71,73,74,85,87,89,90,92,94,97,101,102,105,107,110,111]. The overall random-effects weighted prevalence per 100,000 individuals was 488 (95% CI 382–623) among men and 231 (95% CI 166–321) among women for bacteriologically positive TB and 314 (95% CI 245–403) among men and 129 (95% CI 89–189) among women for smear-positive TB (S3 Table).
Excluding the Region of the Americas—because it had only two small sub-national surveys included in the quantitative analysis—the prevalence of bacteriologically positive TB and smear-positive TB was highest in the African Region. There was strong evidence that male and female prevalence of bacteriologically positive TB per 100,000 individuals was higher in settings with high HIV prevalence in the general population (high versus low HIV prevalence settings: for men, 1,162, 95% CI 735–1,834, versus 360, 95% CI 275–471, p < 0.001; for women, 735, 95% CI 448–1202, versus 157, 95% CI 110–223, p < 0.001). This same relationship (higher prevalence of undiagnosed TB in settings with high HIV prevalence) was also apparent when HIV data from diagnosed TB patients, rather than the general population, were used (for men: 907, 95% CI 582–1,413, versus 359, 95% CI 270–477, p = 0.001; for women: 553, 95% CI 341–896, versus 153, 95% CI 105–224, p < 0.001) (S4 Table). Prevalence of smear-positive TB per 100,000 individuals was also higher in settings with high HIV prevalence in the general population (high versus low HIV prevalence settings: for men, 548, 95% CI 303–990, versus 275, 95% CI 208–364, p = 0.039; for women, 273, 95% CI 131–568, versus 110, 95% CI 71–169, p = 0.036) and in settings with high HIV prevalence in diagnosed TB patients for women (229, 95% CI 126–416, versus 103, 95% CI 64–165, p = 0.040) but not for men (459, 95% CI 289–727, versus 270, 95% CI 200–366, p = 0.060) (S4 Table).
The overall random-effects weighted M:F prevalence ratio was 2.21 for bacteriologically positive TB (95% CI 1.92–2.54; range 0.62 to 6.18; 56 surveys in 24 countries) and 2.51 for smear-positive TB (95% CI 2.07–3.04; range 0.25 to 5.91; 40 surveys in 22 countries). Random-effects weighted M:F prevalence ratios for bacteriologically positive TB and smear-positive TB were significantly greater than one in all regions except the Region of the Americas, where analyses included only two small sub-national surveys (Fig 4).
Among countries with multiple surveys, an excess of male TB cases was observed in all studies in eight (73%) of 11 countries. Exceptions with inconsistent results were Ethiopia, South Africa, and Viet Nam, although overall random-effects weighted M:F prevalence ratios exceeded one for each of these countries.
In univariate meta-regression of M:F ratios in bacteriologically positive TB (Table 2), there was strong evidence that M:F prevalence ratios were 1.95 times higher in the South-East Asia Region than in the African Region (95% CI 1.54–2.48; 56 surveys). M:F prevalence ratios were lower in settings with high HIV prevalence in the general population (0.67, 95% CI 0.49–0.90; 54 surveys) or in incident TB (0.69, 95% CI 0.53–0.93; 54 surveys).
M:F prevalence ratios were also higher in the South-East Asia Region than in the African Region in univariate meta-regression of smear-positive TB (1.91, 95% CI 1.33–2.75; 39 surveys). In this analysis there was also evidence that M:F prevalence ratios were lower in surveys that required individuals to report signs or symptoms of TB during initial screening procedures (0.63, 95% CI 0.42–0.96; 39 surveys) compared to surveys within which initial screening procedures included criteria such as chest X-ray, self-reported history of TB, or self-reported contact with a TB case, instead of or in addition to self-reported signs or symptoms.
In univariate meta-regression models for M:F ratios in bacteriologically positive TB and M:F ratios in smear-positive TB, none of the following survey characteristics were associated with differences in M:F ratios in TB prevalence: survey setting (national versus sub-national), survey starting year, TB prevalence, risk of bias, case definitions, or relative sex ratios in participation.
In multivariate meta-regression of M:F ratios in bacteriologically positive TB, there was evidence that M:F ratios remained higher in the South-East Asia Region than in the African Region after adjusting for HIV prevalence and initial screening procedures, although the relative M:F ratio between these two regions was slightly lower than in univariate analysis (1.78, 95% CI 1.13–2.80; 54 surveys).
There was evidence in the multivariate meta-regression of M:F ratios in smear-positive TB that M:F ratios were 2.21 times higher in the South-East Asia Region than in the African region (95% CI 1.23–4.04; 38 surveys). There was also evidence in the multivariate meta-regression that M:F ratios in surveys that required individuals to self-report signs or symptoms of TB in initial screening procedures were lower than those in surveys with broader initial screening procedures (0.65, 95% CI 0.45–0.93; 38 surveys).
Data on the prevalence of bacteriologically positive TB by sex and age were available for 19 surveys in 13 countries [32,33,35,36,43,44,50,51,53,58,60,65–67,70,71,97,101,107]. Random-effects weighted M:F ratios in prevalence appear to increase with age from 1.28 (95% CI 0.85–1.92; range 0.29 to 5.06) among individuals aged 15–24 y to 3.18 (95% CI 2.24–4.53; range 0.57 to 11.34) among individuals aged 45–54 y (Fig 5).
P:N ratios for smear-positive TB exceeded one for both men and women in 25 (74%) of 34 surveys in 20 countries with available data (Fig 6). The median number of prevalent cases per notified case was 2.6 (interquartile range 1.3–3.4) for men and 1.6 (interquartile range 1.2–2.7) for women, and the overall random-effects weighted M:F ratio for P:N ratios was 1.55 (95% CI 1.25–1.91).
There was no evidence in univariate meta-regression that any of the study or setting characteristics examined were associated with M:F ratios in P:N ratios (S5 Table). Due to the lack of evidence of associations in univariate analyses, multivariate meta-regression was not performed for M:F ratios in P:N ratios.
Meta-analysis of 56 TB prevalence surveys including 2.2 million participants in 28 countries provides strong evidence that TB prevalence is higher among men than women, with a higher M:F ratio than that reported for case notification data. The number of prevalent cases per notified case of smear-positive TB was also higher among men than women, adding evidence that men may be less likely than women to seek or access care in many settings. Further evidence of men’s barriers to seeking or accessing care is provided by results showing that men were less likely than women to participate in prevalence surveys and that relatively fewer prevalent cases were found among men in surveys that required participants to self-report signs or symptoms in initial screening procedures.
The excess male prevalence observed in surveys conducted between 1953 and 1997 [5] persists in more recent surveys, despite widespread implementation of the DOTS strategy and interventions against the HIV pandemic that have decreased overall TB prevalence. Regional summary M:F ratios in the current study were similar to those previously reported for South-East Asia (3.8 versus 3.2), where sex differences were greatest, and the Western Pacific (1.9 versus 2.0). However, in the current study, the summary M:F ratio for the African Region was twice that previously reported (2.0 versus 1.0), suggesting that sex disparities in TB prevalence in this region have increased over the past fifty years. The emergence of HIV during this time has had a substantial impact on TB epidemiology, especially in the African Region. However, while the prevalence of HIV is slightly higher among women than men [113], this study shows that the prevalence of TB is higher among men, even in countries with generalised HIV epidemics. Men also face a relative disadvantage in accessing and remaining in HIV care [114–117], and so men’s risk of TB is likely to be further increased as a result of undiagnosed and untreated HIV co-infection and missed opportunities for TB screening within HIV care.
Comparisons of sex ratios in TB prevalence and notification highlight sex differences in time to diagnosis and imply that in many settings women are more likely than men to have a timely TB diagnosis. While these results could be attributed to men seeking care in private facilities and therefore being less likely to be included in case notification numbers, this explanation would require that the proportion of men who seek care in the public sector be only two-thirds the proportion of women who seek care in that sector. Instead, there is wider evidence that men are less well-served than women by health services [118,119]. Within the context of HIV, which has a similarly lengthy pathway to diagnosis, there is also substantial evidence that men experience greater attrition and worse outcomes [114–117]. Men are less likely than women to access antiretroviral therapy, and in many countries this disparity has increased over time [114]. Similar evidence showing men’s disadvantage in the TB care pathway is building [120–122]. Focusing specifically on access to diagnosis, male TB patients often delay care-seeking longer than female TB patients [123], and this review adds support that timely entry into the TB care pathway may be more difficult for men than women in many settings.
Lower prevalence survey participation among men and evidence of lower M:F prevalence ratios in studies that require individuals to self-report signs or symptoms of TB in initial screening procedures imply that symptom screening in community-based active case finding may be a less effective tool for identifying TB disease in men than women. It is not known whether this is due to men refusing to report symptoms or whether the sub-clinical phase of disease may be longer for men [124]. Further investigation is needed to examine men’s acceptance of screening and reporting of symptoms, even when barriers related to visiting a healthcare facility are removed.
Findings from this review suggest that case detection efforts, whilst not ignoring women, should be greatly strengthened for men. This will require a detailed understanding of the barriers that men face in accessing care. Previous studies have highlighted factors such as loss of income and financial barriers, as well as stigma, that affect men’s healthcare decisions [125,126]. Care-seeking decisions are further influenced by perceptions of masculinity that discourage admission of illness, and systems of care that take away men’s sense of control and leave feelings of inadequacy [127,128]. Interventions to improve case detection among men must recognise and address these barriers. Healthcare providers should be sensitive to men’s needs and consider offering dedicated clinic times and outreach services for men. TB diagnostic services that incorporate men’s peer networks or workplaces to promote wellness and reduce stigma may also be effective. In South Africa, a men-only after-hours clinic situated close to a transport hub has been effective in improving men’s uptake of HIV testing and adherence to antiretroviral therapy [129]. Comparable opportunities for TB strategies that offer convenient access to care while maintaining men’s sense of control should be explored.
This review summarises evidence on sex ratios in TB prevalence from a large number of prevalence surveys across geographic regions, an approach which introduces a number of potential sources of bias. Surveys varied greatly in their methodology, particularly in screening criteria and case definitions, and levels of participation varied within and between studies. However, over 84% of the surveys in the analyses had low to moderate risk of bias.
Prevalence as a measure of disease burden has limitations as it provides an estimate at a single point in time and cannot distinguish between disease as a result of recent infection and disease from reactivation, limiting understanding of current transmission. Comparing the rate of prevalent cases to notified cases is a crude measurement, especially comparing all surveys to national case notification rates, regardless of study setting. Stratifying by age and rural or urban setting would improve P:N ratios; however, data on these characteristics were not available at the time of analysis. Prevalence data by sex and HIV status were too infrequently available to be reported here. To our knowledge, no surveys that conducted drug susceptibility testing reported the results of those analyses by sex, so it is not possible to comment on whether the sex differences reported here are also relevant to drug-resistant TB. Given the significant sex differences reported in prevalence, future surveys should analyse and report all results by sex to facilitate greater understanding of the relationship between gender and TB.
Men have a higher prevalence of TB and, in many settings, remain infectious in the community for a longer period of time than women. Men are therefore likely to generate a greater number of secondary infections than women, and social mixing patterns have suggested that, as a result, men are responsible for the majority of infections in men, women, and children [12]. Addressing men’s burden of disease and disadvantage in TB care is therefore an issue not only for men’s health but for broader TB prevention and care. Given the compelling evidence presented here, global discourse and policy on key underserved populations need to include a focus on men. Recommendations to address issues of gender and TB cannot continue to insist on addressing the needs of women and girls [130] while ignoring the inequity faced by men and boys, who carry the higher burden of disease, often with less access to timely diagnosis and treatment. With a clear need and high burden, improving diagnosis and treatment among men is essential to achieving the ambitious targets of the End TB Strategy.
|
10.1371/journal.pgen.1000200 | Sex-Specific Genetic Structure and Social Organization in Central Asia: Insights from a Multi-Locus Study | In the last two decades, mitochondrial DNA (mtDNA) and the non-recombining portion of the Y chromosome (NRY) have been extensively used in order to measure the maternally and paternally inherited genetic structure of human populations, and to infer sex-specific demography and history. Most studies converge towards the notion that among populations, women are genetically less structured than men. This has been mainly explained by a higher migration rate of women, due to patrilocality, a tendency for men to stay in their birthplace while women move to their husband's house. Yet, since population differentiation depends upon the product of the effective number of individuals within each deme and the migration rate among demes, differences in male and female effective numbers and sex-biased dispersal have confounding effects on the comparison of genetic structure as measured by uniparentally inherited markers. In this study, we develop a new multi-locus approach to analyze jointly autosomal and X-linked markers in order to aid the understanding of sex-specific contributions to population differentiation. We show that in patrilineal herder groups of Central Asia, in contrast to bilineal agriculturalists, the effective number of women is higher than that of men. We interpret this result, which could not be obtained by the analysis of mtDNA and NRY alone, as the consequence of the social organization of patrilineal populations, in which genetically related men (but not women) tend to cluster together. This study suggests that differences in sex-specific migration rates may not be the only cause of contrasting male and female differentiation in humans, and that differences in effective numbers do matter.
| Human evolutionary history has been investigated mainly through the prism of genetic variation of the Y chromosome and mitochondrial DNA. These two uniparentally inherited markers reflect the demographic history of males and females, respectively. Their contrasting patterns of genetic differentiation reveal that women are more mobile than men among populations, which might be due to specific marriage rules. However, these two markers provide only a limited understanding of the underlying demographic processes. To obtain an independent picture of sex-specific demography, we developed a new multi-locus approach based on the analysis of markers from the autosomal and X-chromosomal compartments. We applied our method to 21 human populations sampled in Central Asia, with contrasting social organizations and lifestyles. We found that, in patrilineal populations, not only the migration rate but also the number of reproductive individuals is likely to be higher for women. This result does not hold for bilineal populations, for which both the migration rate and the number of reproductive individuals can be equal for both sexes. The social organization of patrilineal populations is the likely cause of this pattern. This study suggests that differences in sex-specific migration rates may not be the only cause of contrasting male and female differentiation in humans, and that differences in effective numbers do matter.
| Understanding the extent to which sex-specific processes shape human genetic diversity has long been a matter of great interest for human population geneticists [1],[2]. To date, as detailed in Table 1, the focus has mainly been on the analysis of uniparentally inherited markers: mitochondrial DNA (mtDNA) and the non-recombining portion of the Y chromosome (NRY). A large number of studies have found that the level of differentiation was greater for the Y chromosome than for mtDNA, both at a global [3] and a local scale [4]–[11], for a review see [12]. This result has mainly been explained by patrilocality, a widespread tendency for men to stay in their birthplace while women move to their husband's house [13] (see Table 1 for more detailed interpretations). This hypothesis of a higher migration rate of women has been especially strengthened by the comparison of patrilocal and matrilocal populations at a local scale [14]–[17]. These studies have shown that in patrilocal populations, genetic differentiation is stronger among men than among women, while the reverse is observed in matrilocal populations. It is also noteworthy that the absolute difference between male and female genetic structure is more pronounced in patrilocal than in matrilocal populations [16]. Interestingly, while social practices seem to consistently influence the sex-specific demography at a local scale, the robustness of a sex-specific genetic structure at a global scale is still a challenging issue (see Table 1). A recent analysis of mtDNA and NRY variation at a global scale, which used the same panel of populations for both categories of markers (an omission that was criticized in Seielstad et al.'s [3] study [18]) showed no difference between the male and female genetic structure [19]. Consistent with this result, an analysis of the autosomal and X-linked microsatellite markers in the HGDP-CEPH Human Genome Diversity Cell Line Panel showed no major differences between the demographic history of men and women [20]. The apparent paradox between local and global trends can be resolved though, since the geographical clustering of populations with potentially different lifestyles may minimize the differences in sex-specific demography at a global scale [21],[22]. It may also be that the global structure reflects more ancient, pre-agricultural, social patterns, as patrilocality may only have increased in human societies only with the recent transition to agriculture [12].
The higher differentiation level found on the NRY as compared to mtDNA at a local scale could also be the consequence of a higher effective number of women, for example through the practice of polygyny, a tendency for men (but not for women) to have multiple mates [4], [7], [15], [23]–[25], and/or through the paternal transmission of reproductive success [11]. However, the influence of such processes on genetic structure has often been considered as negligible, since realistic rates of polygyny cannot create large differences in male and female genetic structure [3],[5],[14]. Hence, until now, the effect of local social processes on male and female effective numbers has not been investigated directly, possibly because current methods fail to unravel the relative contribution of effective number and migration rate on the differentiation level [26]. The consequence is that the vast majority of studies fail to show whether the observed differentiation arises from sex-specific differences in migration rate, effective numbers, or both (see Table 1). New methods need therefore to be developed in order to appreciate the relative influence of sex-biased dispersal and differences in effective numbers on genetic structure.
Another limitation to the use of uniparentally inherited markers stems from the fact that each of them is, in effect, a single genetic locus. For that reason, we cannot test for the robustness of the sex-specific genetic structure on these markers. We cannot either rule out the possibility that mtDNA and NRY, which contain multiple linked genes, may be shaped by selection [27],[28]. This raises the question of whether results based on uniparentally inherited markers simply reflect stochastic variation, or real differences in sex-specific demography. To answer this question, we propose a novel approach based on the joint analysis of autosomal and X-linked markers. This multi-locus analysis has the potential of providing more robust information, as these markers give an independent picture of sex-specific demography. This approach also aims to disentangle the effects of sex-biased dispersal and effective numbers on genetic structure.
In order to recognize the impact of social organization on these differences, we investigate sex-specific genetic structure in human populations of Central Asia (Figure 1), where various ethnic groups, characterized by different languages, lifestyles and social organizations, co-exist. Although all groups share a patrilocal organization, Tajiks (sedentary agriculturalists) are bilineal, i.e. they are organized into nuclear or extended families where blood links and rights of inheritance through both male and female ancestors are of equal importance, and they preferentially establish endogamous marriages with cousins. By contrast, Kazaks, Karakalpaks, Kyrgyz and Turkmen (traditionally nomadic herders) are patrilineal, i.e. they are organized into paternal descent groups (tribes, clans, lineages), and they practice exogamous marriages, in which a man chooses a bride from a different clan.
We sampled 780 healthy adult men from 10 populations of bilineal agriculturalists and 11 populations of patrilineal herders from West Uzbekistan to East Kyrgyzstan, representing 5 ethnic groups (Tajiks, Kyrgyz, Karakalpaks, Kazaks, and Turkmen) (see Figure 1 and Table 2). We genotyped all bilineal populations, and 8 out of 11 patrilineal populations at the HVS-I locus of mtDNA, and at 11 microsatellite markers on the NRY (for more details on the markers used, see Table 3). The overall genetic differentiation was higher for NRY, as compared to mtDNA, both among the 10 bilineal agriculturalist populations , and among the subset of 8 patrilineal herder populations . Assuming an island model of population structure, this implies that female migration rate (mf), and/or the effective number of females (Nf), is higher than of the corresponding parameters for males (mm and Nm). These results also suggest that the differences in sex-specific genetic structure are much more pronounced in the patrilineal herders than in the bilineal agriculturalists. From the above FST estimates, we obtained the female-to-male ratio of the effective number of migrants per generation (see the Methods section for details): Nfmf/Nmmm≈2.1 for bilineal populations and Nfmf/Nmmm≈21.6 for patrilineal populations. The ratio in patrilineal populations is thus one order of magnitude higher than in bilineal populations. However, since each of these markers is a single genetic locus, we cannot test for the robustness of the sex-specific genetic structure on these markers. We therefore examined the amount of information contained in multi-locus data on autosomal and X-linked markers, both of which average over male and female histories.
In the infinite island model of population structure with two classes of individuals (males and females), we obtained the following expressions of FST (see the Methods section for details):(1)for autosomal genes, and(2)for X-linked genes. A special case of interest occurs when , i.e. when the differentiation of X-linked genes exactly equals that of autosomal genes. Combining eqs (1) and (2), we find that this occurs for , with N = Nf+Nm and m = mf+mm. Furthermore, as shown in Figure 2, if we observe a lower genetic differentiation of autosomal markers, as compared to X-linked markers (blue zone in Figure 2), this suggests that . This may happen, e.g., for Nf = Nm and mf = mm, i.e. for equal effective numbers of males and females and unbiased dispersal. But if autosomal markers are more differentiated than X-linked markers (, see the red upper-right triangle in Figure 2), this implies that . In this case, since mf/m and Nf/N are ratios varying between 0 and 1, the effective number of females must be higher than that of males (Nf>Nm), and the female migration rate must be higher than half the male migration rate (mf>mm/2). Hence, a prediction from this model is that when , the effective number of females is higher than that of males, whatever the pattern of sex-specific dispersal. This suggests that it is indeed possible to test for differences in effective numbers between males and females from the joint analysis of autosomal and X-linked data. We note however that when , we cannot conclude on the relative male and female effective numbers and migration rates.
We tested the above prediction in the 10 bilineal agriculturalist populations and 11 patrilineal herder populations sampled in Central Asia by comparing the genetic structure estimated from 27 unlinked polymorphic autosomal microsatellite markers (AR = 16.2, He = 0.803 on average) to that from 9 unlinked polymorphic X-linked microsatellite markers (AR = 12.6, He = 0.752 on average) (for more details on the markers used, see Table 4). Overall heterozygosity was not significantly different between X-linked and autosomal markers, neither in the pooled sample (two-tailed Wilcoxon sum rank test; p = 0.09), nor in the bilineal agriculturalists (p = 0.13) or the patrilineal herders (p = 0.12). The overall population structure was significantly higher for autosomal as compared to X-linked markers among patrilineal herders: (one-tailed Wilcoxon sum rank test; ; p = 0.02). Among bilineal agriculturalists, the result was not significant: (p = 0.36). From these results, and following our model predictions, we conclude that in patrilineal herders (where ), the effective number of females is higher than that of males. This conclusion does not hold for the bilineal agriculturalists.
From our model, it is possible to get more precise indications on the sets of (Nf/N, mf/m) values that are compatible with our data. Rearranging eqs (1–2), we get:(3)i.e.:(4)
For any given set of (Nf/N, mf/m) values, we can therefore calculate from eq. (4) the expected value of for each estimate in the dataset. We can then test the null hypothesis by comparing the distribution of observed and expected values. If the hypothesis can be rejected at the α = 0.05 level, then the corresponding set of (Nf/N, mf/m) values can also be rejected. Following Ramachandran et al. [20], we varied the values of the ratios Nf/N and mf/m (respectively, the female fraction of effective number, and the female fraction of the total migration rate) from 0 to 1, with an interval of 0.01 between consecutive values. For each set of (Nf/N, mf/m) values, we applied the transformation in eq. (4) to each of the 27 locus-specific values observed. Thus, for each set of (Nf/N, mf/m) values, we obtained 27 expected values of , given our data. These expected values of were then compared to the 9 observed locus-specific in our dataset, and we calculated the p-value for a two-sided Wilcoxon sum rank test between the list of 27 expected values and the 9 observed in the dataset. The results are depicted in Figure 3. Significant p-values (p≤0.05) correspond to a significant difference between the observed and expected values, thus to sets of (Nf/N, mf/m) values that are rejected, given our data (see the blue region in Figure 3). Conversely, non-significant p-values (p>0.05) correspond to sets of (Nf/N, mf/m) values that cannot be rejected (see the red region in Figure 3).
For the patrilineal herder populations (Figures 3A–3B), most sets of (Nf/N, mf/m) values are rejected, except those corresponding to larger effective numbers for females (from Figures 3A–3B: Nf/N>0.55, i.e. Nf>1.27Nm) and mf>0.67mm. Because the multi-locus estimate of is significantly higher than the estimate of , we expected to find such patterns of non-significant values (see Figure 2). For the bilineal agriculturalist populations, we could not reject the hypothesis that the effective numbers and migration rates are equal across males and females or even lower in females (see Figures 3C–3D). This is also reflected by the fact that the estimates of were not significantly higher than the estimates of in those populations.
Finally, we have shown that the effective number of women is higher than that of men among patrilineal herders, but not necessarily among bilineal agriculturalists. Furthermore, a close inspection of the results depicted in Figures 3A and 3B reveals that, among herders, we reject all the sets of (Nf/N, mf/m) values for which mf<mm at the α = 0.10 level. This is not true for agriculturalists. This suggests that the migration rates are also likely to be higher for women than for men in patrilineal populations, as compared to bilineal populations (compare Figures 3B and 3D). Although both groups are patrilocal, such a difference in sex-specific migration patterns might be expected, since patrilineal herders are exogamous (among clans) and bilineal agriculturalists are preferentially endogamous. For example, it was observed that in patrilocal and matrilocal Indian populations, where migrations are strictly confined within endogamous groups, sex-specific patterns were not influenced by post-marital residence [21].
While an influence of post-marital residence on the migration rate of women and men has already been widely proposed [14]–[17] (see also Table 1), the factors that may locally affect the effective number of women, relatively to that of men, are not well recognized. As seen in Table 1, although a number of studies have compared matrilocal and patrilocal populations, few have compared contrasting groups of populations with respect to other factors as, e.g., the tendency for polygyny [15]. Furthermore, a number of these studies lack ethnological information a priori, concerning social organization, marriage rules, etc., which makes interpretation somewhat difficult (see Table 1). Here, we compared two groups of patrilocal populations with contrasting social organizations, and at least five non-mutually exclusive interpretations for a larger effective number of females can be invoked:
There might also be non-biological explanations of our results, however, as they are based on the simplifying assumptions of Wright's infinite island model of population structure [39]. This model assumes (i) that there is no selection and that mutation is negligible, (ii) that each population has the same size, and sends and receives a constant fraction of its individuals to or from a common migrant pool each generation (so that geographical structure is absent), and (iii) that equilibrium is reached between migration, mutation and drift. On the first point, we did not find any evidence of selection, for any marker, based on Beaumont and Nichols' method [40] for detecting selected markers from the analysis of the null distribution generated by a coalescent-based simulation model (data not shown). As for the second point, we tested for the significance of the correlation between the pairwise FST/(1−FST) estimates and the natural logarithm of their geographical distances [41]. We found no evidence for isolation by distance, either for X-linked markers (p = 0.47 for agriculturalists, p = 0.24 for herders), or for autosomal markers (p = 0.92 for agriculturalists, p = 0.45 for herders). As for the third point, the X-to-autosomes (X/A) effective size ratio can significantly deviate from the expected three-quarters (assuming equal effective numbers of men and women) following a bottleneck or an expansion [42]. This is because X-linked genes have a smaller effective size, and hence reach equilibrium more rapidly. After a reduction of population size, the X/A diversity ratio is lower than expected, while after an expansion, the diversity of X-linked genes recovers faster than on the autosomes, and the X/A diversity ratio is then closer to unity. In the latter case, would be reduced and could then tend towards . However, neither reduction nor expansion should lead to , as we found in herder populations of Central Asia. Therefore, we do not expect the limits of Wright's island model to undermine our approach.
We aimed to investigate to what extent the approach proposed here is able to detect differences in male and female effective numbers. To do this, we performed coalescent simulations in a finite island model, for a wide range of (Nf/N, mf/m) values. The simulation parameters were set to match those of our dataset: 11 sampled demes, 30 males genotyped at 27 autosomal and 9 X-linked markers per deme (for further details concerning the simulations, see the Methods section). We used 1421 sets of (Nf/N, mf/m) values, covering the whole parameter space (represented as white dots in Figure 4B). For each set of (Nf/N, mf/m) parameter values, we simulated 100 independent datasets. For each dataset, we calculated the estimates of at all loci, and we calculated the p-value for a one-sided Wilcoxon sum rank test for the list of estimates . Hence, for each set of (Nf/N, mf/m) parameter values, we could calculate the proportion of significant tests at the α = 0.05 level, among the 100 independent datasets. Figure 4 shows the distribution of the percentage of significant tests in the (Nf/N, mf/m) parameter space. Theory predicts that in the upper-right triangle where , we should have . One can see from Figure 4 that, given the simulation parameters used, the method is conservative: the proportion of significant tests at the α = 0.05 level is null outside of the upper-right triangle. However, we find a fairly large proportion of significant tests for large Nf/N and mf/m ratios which indicates (i) that the method presented here has the potential to detect differences in male and female effective numbers, but (ii) that only strong differences might be detected, for similarly sized datasets as the one considered here.
We also aimed to investigate whether the results obtained here were robust to our sampling scheme, and that our results were not biased by the inclusion of particular populations. To do this, we re-analyzed both the bilineal agriculturalists and the patrilineal herders datasets, removing one population at a time in each group. For each of these jackknifed datasets, we calculated the p-value of a one-sided Wilcoxon sum rank test , as done on the full datasets. The results are given in Table 5. We found no significant test for any of the bilineal agriculturalist groupings (p>0.109), which supports the idea that, in those populations, both the migration rate and the number of reproductive individuals can be equal for both sexes. In patrilineal herders, the tests were significant at the α = 0.05 level for 8 out of 11 population groupings. For the 3 other groupings, the p-values were 0.068, 0.078 and 0.073 (see Table 5). Overall, the ratio of multi-locus estimates ranged from 1.7 to 3.5 in patrilineal herders (and from 0.9 to 1.2 in bilineal agriculturalists). Although in some particular groupings of patrilineal herder populations, the difference in the distributions of may not be strong enough to be significant, we can clearly distinguish the pattern of differentiation for autosomal and X-linked markers in patrilineal and bilineal groups. Results from coalescent simulations (see above) suggest that this lack of statistical power might be expected for ratios close to unity. Indeed, we found that the tests were more likely to be significant for fairly large Nf/N and mf/m ratios (the upper-right red region in Figure 4) which would correspond to ratios much greater than one.
Importantly, our results on X-linked and autosomal markers are consistent with those obtained from NRY and mtDNA (see Figures 3B–3D): in these figures, the dashed line gives all the sets of (Nf/N, mf/m) values that are compatible with the observed estimates. These are the sets of values that satisfy for the bilineal populations, and for the patrilineal populations, since we inferred Nfmf/Nmmm≈2.1 and Nfmf/Nmmm≈21.6, respectively, for the two groups. For the bilineal agriculturalists (Figure 3D), the set of (Nf/N, mf/m) values inferred from the estimates fall within the range that was not rejected, given our data on X-linked and autosomal markers. For the patrilineal herders (Figure 3B), the overlap is only partial: from the NRY and mtDNA data only, low Nf/N ratios associated with high mf/m ratios are as likely as high Nf/N ratios associated with low mf/m ratios. Yet, it is clear from this figure that a large set of (Nf/N, mf/m) values inferred from the single-locus estimates can be rejected, given the observed differentiation on X-linked and autosomal markers. All genetic systems (mtDNA, NRY, X-linked and autosomal markers) converge toward the notion that patrilineal herders, in contrast to bilineal agriculturalists, have a strong sex-specific genetic structure. Yet, the information brought by X-linked and autosomal markers is substantial, since we show that this is likely due to both higher migration rates and larger effective numbers for women than for men.
Our results, based on the X chromosome and the autosomes, also confirm previous analyses based on the mtDNA and the NRY, showing that men are genetically more structured than women in other patrilocal populations [3]–[10], [14]–[17] (see also Table 1). A handful of studies have also shown a reduced effective number of men compared to that of women, based on coalescent methods [23],[24], but none have considered the influence of social organization on this dissimilarity (see Table 1).
In some respects, our results contrast with those of Wilder and Hammer [25], who studied sex-specific population genetic structure among the Baining of New Britain, using mtDNA, NRY, and X-linked markers. Interestingly, they found that Nf>Nm, but mf<mm, and claimed that a similar result, although left unexplored by the authors, was to be found in a recent study by Hamilton et al. [16]. This raises the interesting point that sex-specific proportions of migrants (m) are likely to be shaped by factors that may only partially overlap with those that affect the sex-specific effective numbers (N). Further studies of human populations with contrasted social organizations, as well as further theoretical developments, are needed to appreciate this point.
In order to ask to what extent our results generalize to other human populations, we investigated sex-specific patterns in the 51 worldwide populations represented in the HGDP-CEPH Human Genome Diversity Cell Line Panel dataset [43], for which the data on the differentiation of 784 autosomal microsatellites and 36 X-linked microsatellites are available (data not shown). By doing this, we found a larger differentiation for X-linked than for autosomal markers . Therefore, we confirmed Ramachandran et al.'s [20] result that no major differences in demographic parameters between males and females are required to explain the X-chromosomal and autosomal results in this worldwide sample. Ramachandran et al.'s approach [20] is based upon a pure divergence model from a single ancestral population, which is very different from the migration-drift equilibrium model considered here. In real populations, however, genetic differentiation almost certainly arises both through divergence and limited dispersal, which places these two models at two ends of a continuum. Yet, importantly, if we apply Ramachandran et al.'s [20] model to the Central Asian data, our conclusions are left unchanged. In their model, the differentiation among populations is , where t is the time since divergence from an ancestral population and Ne the effective size of the populations (see, e.g., [44]). Hence, we get for autosomal and X-linked markers, respectively. Therefore, our observation that implies that , which requires that Nf>7Nm since (see, e.g., [45]). In this case, the female fraction of effective number is larger than that of males, which is consistent with our findings in a model with migration.
The HGDP-CEPH dataset does not provide any detailed ethnic information for the sampled groups, and we can therefore not distinguish populations with different lifestyles. However, at a more local scale in Pakistan, we were able to analyze a subset of 5 populations (Brahui, Balochi, Makrani, Sindhi and Pathan), which are presumed to be patrilineal [46]. For this subset, we found a higher differentiation for autosomal than for X-linked markers , although non-significantly (p = 0.12). This result seems to suggest that other patrilineal populations may behave like the Central Asian sample presented here. Therefore, because the geographical clustering of populations with potentially different lifestyles may minimize the differences in sex-specific demography at a global scale [21],[22], and/or because the global structure may reflect ancient (pre-agricultural) marital residence patterns with less pronounced patrilocality [12], we emphasize the point that large-scale studies may not be relevant to detect sex-specific patterns, which supports a claim made by many authors.
In conclusion, we have shown here that the joint analysis of autosomal and X-linked polymorphic markers provides an efficient tool to infer sex-specific demography and history in human populations, as suggested recently [12],[47]. This new multilocus approach is, to our knowledge, the first attempt to combine the information contained in mtDNA, NRY, X-linked and autosomal markers (see Table 1), which allowed us to test for the robustness of a sex-specific genetic structure at a local scale. Unraveling the respective influence of migration and drift upon neutral genetic structure is a long-standing quest in population genetics [48],[49]. Here, our analysis allowed us to show that differences in sex-specific migration rates may not be the only cause of contrasted male and female differentiation in humans and that, contrary to the conclusion of a number of studies (see Table 1), differences in effective numbers may also play an important role. Indeed, we have demonstrated that sex-specific differences in population structure in patrilineal herders may be the consequence of both higher female effective numbers and female effective dispersal. Our results also illustrate the importance of analyzing human populations at a local scale, rather than global or even continental scale [2],[19],[21]. The originality of our approach lies in the comparison of identified ethnic groups that differ in well-known social structures and lifestyles. In that respect, our study is among the very few which compare patrilineal vs. bilineal or matrilineal groups (see Table 1), and we believe that it contributes to the growing body of evidence showing that social organization and lifestyle have a strong impact on the distribution of genetic variation in human populations. Moreover, our approach could also be applied on a wide range of animal species with contrasted social organizations. Therefore, we expect our results to stimulate research on the comparison of X-linked and autosomal data to disentangle sex-specific demography.
We sampled 10 populations of bilineal agriculturalists and 11 populations of patrilineal herders from West Uzbekistan to East Kyrgyzstan, representing 780 healthy adult men from 5 ethnic groups (Tajiks, Kyrgyz, Karakalpaks, Kazaks, and Turkmen) (see Table 2). The geographic distribution of the samples and information about lifestyle is provided in Figure 1. Also living in Central Asia, Uzbeks are traditionally patrilineal herders too, but they have recently lost their traditional social organization [11], and we therefore chose not to include any sample from this ethnic group for the purpose of this study. We collected ethnologic data prior to sampling, including the recent genealogy of the participants. Using this information, we retained only those individuals that were unrelated for at least two generations back in time. All individuals gave their informed consent for participation in this study. Total genomic DNA was isolated from blood samples by a standard phenol-chloroform extraction [50].
The mtDNA first hypervariable segment of the mtDNA control region (HVS-I) was amplified using primers L15987 (5′TCAAATGGGCCTGTCCTTGTA) and H580 (5′TTGAGGAGGTAAGCTACATA) in 18 populations out of 21 (674 individuals, see Table 2). The amplification products were subsequently purified with the EXOSAP standard procedure. The sequence reaction was performed using primers L15925 (5′TAATACACCAGTCTTGTAAAC) and HH23 (5′AATAGGGTGATAGACCTGTG). Sequences from positions 16 024–16 391 were obtained. Eleven Y-linked microsatellite markers (see Table 3) were genotyped in the same individuals, following the protocol described by Parkin et al. [51].
27 autosomal and 9 X-linked microsatellite markers (see Table 4) were genotyped in the same individuals. We used the informativeness for assignment index In [52] to select subsets of microsatellite markers on the X chromosome and the autosomes from the set of markers used in Rosenberg et al.'s worldwide study [43]. This statistic measures the amount of information that multiallelic markers provide about individual ancestry [52]. This index was calculated among a subset of 14 populations, chosen from the Rosenberg et al.'s dataset [43] to be genetically the closest to the Central Asian populations (Balochi, Brahui, Burusho, Hazara, Pathan, Shindi, Uygur, Han, Mongola, Yakut, Adygei, Russian, Druze and Palestinian). The rationale was to infer the information provided by individual loci about ancestry from this subset of populations, and to extrapolate the results to the populations studied here. For the X chromosome data, we pooled the ‘Screening Set10’ and ‘Screening Set52’ from the HGDP-CEPH Human Genome Diversity Cell Line Panel [53] analyzed by Rosenberg et al. [43] which represented a total of 36 microsatellites. We chose 9 markers among the 11 with the highest In. For autosomal data, we used the ‘Screening Set10’, which represented a total of 377 microsatellites, and chose 27 markers among the 30 with the highest In. All markers were chosen at a minimum of 2 cM apart from each others [54]. PCR amplifications were performed in a 20 µl final volume composed of 1× Eppendorf buffer, 125 µM each dNTP, 0.5U Eppendorf Taq polymerase, 125 nM of each primer, and 10 ng DNA. The reactions were performed in a Eppendorf Mastercycler with an initial denaturation step at 94°C for 5 min; followed by 36 cycles at 94°C for 30 s, 55°C for 30 s, 72°C for 20 s, and 72°C for 10 min as final extension. Forward primers were fluorescently labeled and reactions were further analyzed by capillary electrophoresis (ABI 310, Applied Biosystems). We used the software package Genemarker (SoftGenetics LLC) to obtain allele sizes from the analysis of PCR products (allele calling).
We calculated the total allelic richness (AR) (over all populations), the unbiased estimate of expected heterozygosity He [55], the total number of polymorphic sites and FST for mtDNA using Arlequin version 3.1. [56]. Genetic differentiation among populations for the autosomes, the X and the Y chromosome was measured both per locus and overall loci using Weir and Cockerham's FST estimator [57], as calculated in Genepop 4.0. [58]. The 95% confidence intervals were obtained by bootstrapping over loci [58], using the approximate bootstrap confidence intervals (ABC) method described by DiCiccio and Efron [59]. Isolation by distance (i.e. the correlation between the genetic and the geographic distances) was analyzed by computing the regression of pairwise FST/(1−FST) estimates between pairs of populations to the natural logarithm of their geographical distances, and rank correlations were tested using the Mantel permutation procedure [60], as implemented in Genepop 4.0. [58]. All other statistical tests were performed using the software package R v. 2.2.1 [61].
Let us consider an infinite island model of population structure [62], with two classes of individuals (males and females), which describes a infinite set of populations with constant and equal sizes that are connected by gene flow. Then the expected values of FST for uniparentally inherited markers depend on the effective number Nm (resp. Nf) of adult males (resp. females) per population and the migration rate mm (resp. mf) of males (resp. females) per generation, as: (see, e.g., [63]). We can therefore calculate the female-to-male ratio of the effective number of migrants per generation as: .
In this model, we can also compute for the autosomes and the X chromosome the reproductive values for each class (sex), which are interpreted here as the probability that an ancestral gene lineage was in a given class in a distant past [64]. From these, we can obtain the well-known expressions of effective size Ne for autosomal and X-linked genes: , respectively [45]. Note that Ne is expressed here as a number of gene copies (i.e., twice the effective number of diploid individuals for autosomes). Likewise, the effective migration rate, i.e. the average dispersal rate of an ancestral gene lineage, is given by for autosomal genes, and for X-linked genes, respectively. Substituting these expressions into the well-known equation: FST≈1/(1+2Neme) [64], we get:(5)for autosomal genes, and(6)for X-linked genes.
We performed coalescent simulations, using an algorithm in which coalescence and migration events are considered generation-by-generation until the common ancestor of the whole sample has been reached (see [65]). We simulated a finite island model with 50 demes, each made of N = Nf+Nm = 500 diploid individuals, with a migration parameter m = mf+mm = 0.2. Using these total values for N and m, we then varied the sex-specific parameters to cover the (Nf/N, mf/m) parameter space evenly. Note that the parameter m is the total migration rate, which corresponds to twice the effective migration rate for autosomal markers. Hence, for each set of (Nf/N, mf/m) values, the total number of individuals is 500 (although the number of females may vary from 1 to 499) and the effective migration rate for autosomal markers is . We chose these total values for N and m such that, for a ratio Nfmf/Nmmm = 21.6 (as observed for the herder populations), the distribution of FST estimates on uniparentally-inherited markers in the simulations were close to the observations: for mtDNA, the 95% highest posterior density interval (see [66], pp. 38–39) for the distribution of FST estimates in the simulations was [0.007; 0.033] with a mode at 0.014 (estimated value from the real dataset: among the herders) while for the NRY, the 95% highest posterior density interval was [0.088; 0.374] with a mode at 0.187 (estimated value from the real dataset: ).
Each simulated sample consisted in 330 sampled males from 11 populations (30 males per population), genotyped at 27 autosomal, 9 X-linked markers as well as 10 Y-linked markers and a single mtDNA locus. Each locus was assumed to follow a Generalized Stepwise Model (GSM) [67] with a possible range of 40 contiguous allelic states, except the mtDNA, which was assumed to follow an infinite allele model of mutation. The average mutation rate was 5.10−3, and the mean parameter of the geometric distribution of the mutation step lengths for microsatellites was set to 0.2 [67],[68].
|
10.1371/journal.pntd.0006254 | Hyperendemic dengue transmission and identification of a locally evolved DENV-3 lineage, Papua New Guinea 2007-2010 | Dengue is endemic in the Western Pacific and Oceania and the region reports more than 200,000 cases annually. Outbreaks of dengue and severe dengue occur regularly and movement of virus throughout the region has been reported. Disease surveillance systems, however, in many areas are not fully established and dengue incidence is underreported. Dengue epidemiology is likely least understood in Papua New Guinea (PNG), where the prototype DENV-2 strain New Guinea C was first isolated by Sabin in 1944 but where routine surveillance is not undertaken and little incidence and prevalence data is available.
Serum samples from individuals with recent acute febrile illness or with non-febrile conditions collected between 2007–2010 were tested for anti-DENV neutralizing antibody. Responses were predominantly multitypic and seroprevalence increased with age, a pattern indicative of endemic dengue. DENV-1, DENV-2 and DENV-3 genomes were detected by RT-PCR within a nine-month period and in several instances, two serotypes were identified in individuals sampled within a period of 10 days. Phylogenetic analysis of whole genome sequences identified a DENV-3 Genotype 1 lineage which had evolved on the northern coast of PNG which was likely exported to the western Pacific five years later, in addition to a DENV-2 Cosmopolitan Genotype lineage which had previously circulated in the region.
We show that dengue is hyperendemic in PNG and identify an endemic, locally evolved lineage of DENV-3 that was associated with an outbreak of severe dengue in Pacific countries in subsequent years, although severe disease was not identified in PNG. Additional studies need to be undertaken to understand dengue epidemiology and burden of disease in PNG.
| Dengue virus (DENV) was first identified in Papua New Guinea (PNG) in 1944. Dengue is currently assumed to be an endemic disease in PNG although there is little incidence or prevalence data, and the evidence consensus for dengue presence is low. Routine surveillance is not undertaken and dengue is not a notifiable disease. Severe dengue is rarely identified by local clinicians and the reasons for this are unclear but may be related to poor recognition of dengue and a low index of suspicion, despite high incidence and prevalence rates in neighbouring countries. For example, Indonesia shares borders with PNG and regularly reports outbreaks of severe dengue and transmission of multiple DENV serotypes. DENV infection is identified in travellers from PNG however there are no data on locally circulating strains and how they may compare to viruses associated with severe dengue epidemics in other countries in the Asia Pacific region. We identified evidence for previous infection with all four DENV serotypes among people living on the northern coast of PNG, in Madang, and on Lihir Island in the Bismarck Archipelago off the northeastern coast. We also detected DENV-1, DENV-2, and DENV-3 virus in febrile patients, and we describe the first whole genome sequences of endemically circulating DENV since the prototype 1944 DENV-2New Guinea C strain was characterized. Of note, severe dengue was not diagnosed in any patient infected with these viruses in PNG although introduction of the PNG DENV-3 strain into the Solomon Islands five years later resulted in a large outbreak of severe dengue with hospitalizations and deaths in that country. Dengue epidemiology and burden of disease should be investigated in PNG.
| The dengue viruses (DENV) are the most important arboviral pathogens of humans causing an estimated 390 million infections annually, of which approximately one quarter are symptomatic [1]. Infection with DENV causes a spectrum of clinical outcomes ranging from self-limiting febrile illness (dengue fever, DF) to potentially fatal severe dengue, characterized by plasma leakage, thrombocytopenia, and hypovolemic shock. Dengue is endemic in more than 100 tropical and subtropical countries, where the principal mosquito vectors Aedes aegypti and Aedes albopictus are found [2, 3].
DENV is a single-stranded positive-sense RNA virus of the Flaviviridae family. Like other RNA viruses, DENV displays considerable genetic diversity and is grouped into four antigenically distinct serotypes (DENV-1-DENV-4) which may be distinguished on the basis of serum neutralization tests. The four serotypes are more precisely classified, using phylogenetic approaches, into distinct genotypes which have been defined as clusters with nucleotide sequence divergence of not more than 6% [4]; lineages within the genotypes may represent strains with similar geographic origins [5]. Certain genotypes have been associated with more [6,7] or less [8] virulent disease, and there is some evidence for humoral and cellular immune selection focused on viral B- and T-cell epitopes [9,10]. DENV genetic diversity thus appears to impact host mechanisms shown to mediate pathogenesis [11] and ultimately, disease severity.
Dengue was first identified in Papua New Guinea (PNG) when Sabin isolated the prototype DENV-2New Guinea C strain from febrile soldiers deployed on the northern coast of New Guinea in 1944 [12]. A DENV strain with similar biologic features as the prototype DENV-1Hawaii that Sabin had recently isolated from febrile soldiers in Hawaii, was also isolated from soldiers in the same area of New Guinea in 1944 [13], suggesting that at least two serotypes may have circulated in the north of PNG in the 1940s. More recently, DENV-1-3 genetic data derived from viremic travellers returning to northern Australia from PNG between 1999–2010 [14] indicate that dengue is endemic in PNG, supporting an earlier serological study demonstrating a seroprevalence rate of 8% among patients presenting to clinics in Madang with acute febrile illness [15]. Despite the likely transmission of multiple DENV serotypes and the potential for introduction of DENV from endemic neighbouring countries which experience large-scale epidemics of severe dengue [16], little is known about the epidemiology and transmission dynamics of dengue in PNG, where dengue surveillance is not undertaken, individuals with acute febrile illness are not routinely tested for DENV infection, and where severe disease is rarely reported.
We sought to determine DENV serotype and genotype prevalence in local populations presenting with febrile illness, or with a range of non-febrile conditions. We sequenced whole genomes of DENV and conducted a phylogenetic analysis to determine the evolutionary origin of PNG DENV. In addition we determined anti-DENV-1-4 neutralizing antibody profiles in adults and children in order to assess prevalence and serotype diversity.
Samples analysed in this study were collected from Madang, on the northern coast of PNG, and from Lihir Island in New Ireland Province. Madang is a town of about 30,000 people with a sea port that is a major hub for domestic and international shipping, and an airport where domestic flights from throughout PNG arrive several times each day. Lihir Island is 800 kilometres northeast of Madang in the Bismarck Archipelago, in the western Pacific Ocean. The island’s population doubled to more than 12,000 people after establishment of a gold mine in 1997 and although many residents still live a predominantly traditional subsistence lifestyle, in recent years there has been an influx of PNG-national and expatriate mine workers and development of an international airport and sea port.
Madang sera were collected from febrile patients presenting to the outpatient clinic of Yagaum rural hospital or to Jomba town clinic from September 2007 through June 2008, and who were enrolled in a malaria study [15]. Sera excluded for malaria antigens were tested for anti-DENV IgG and IgM, and NS1 antigen; 8% (46/578) were identified as probable acute DENV infection ie. NS1 antigen-positive and/or anti-DENV IgM-positive. A total of 55 acute phase sera (46 sera identified serologically as probable acute DENV infection plus 9 additional febrile sera that were not tested for DENV), were assessed in the present study for the presence of DENV by DENV E gene RT-PCR and virus isolation was attempted on RT-PCR positive samples. Whole genomes of isolated viruses were sequenced using Illumina. Convalescent sera from 119 patients excluded for acute DENV infection and who presented for recollection were collected an average of 29.6 days (range 5–159 days) after the first patient visit and tested for the presence of anti-DENV neutralizing antibody (NAb) to all four serotypes simultaneously using a microneutralization assay optimized for small sample volumes [17].
Lihir sera (55 in total) were collected from patients presenting to the outpatient clinic of Lihir rural health centre from May through November 2010 during pre-employment medical visits, at antenatal screening visits, or from patients presenting for a range of conditions including joint pain, diabetes and fever. The sera were also assessed for DENV genomes (11/55 sera from febrile patients) and for anti-DENV NAb (44/55 sea from patients with non-febrile conditions).
Patient data are summarized in Table 1.
Ethics approval for this study was granted by the Medical Research Advisory Committee, Ministry of Health, Government of Papua New Guinea (2010) and the Human Research Ethics Committee, University of Western Australia (2010). All data analysed were anonymized.
DENV was isolated from serum by inoculation onto monolayers of Vero cells [5]. Briefly, 100μl of acute phase serum was inoculated onto a Vero cell monolayer in minimal media in a 2.5 ml culture tube, and incubated overnight. On the following day the inoculum was removed and 3ml of DMEM with 2% FBS (supplemented with L-glutamine and antibiotics) was added to the cells, and the culture was maintained at 37°C with 5% CO2 for 7 days or until cytopathic effect (CPE) was observed; for most samples a blind passage into a second 2.5 ml culture tube was required in sorder to isolate virus. Successful virus isolation was identified by NS1 antigen ELISA (Platelia Dengue NS1 Antigen ELISA; Bio-Rad, Australia).
Viral RNA was extracted from 140 μl of culture supernatant using QIAmp viral RNA Mini kits (Qiagen), according to the manufacturer’s instructions. cDNA was synthesized from extracted RNA using SuperScript III First-Strand Synthesis System for RT-PCR (Invitrogen) as per the manufacturer’s instructions. DENV serotype was identified by RT-PCR using serotype-specific primers [5], and the LongRange PCR Kit (Qiagen) (thermocycling conditions are available on request).
A serum microneutralization (MN) assay was used to measure serum anti-DENV antibodies [17]. This approach was selected to allow simultaneous assessment of antibody to all four DENV serotypes in samples with limited volumes. Standard anti-DENV-1-4 sera NIBSC 05/248 (National Institute for Biological Standards and Control [NIBSC], Potter’s Bar, Hertfordshire, United Kingdom) were assayed against the homologous DENV prototype strains DENV-1Hawaii2001; DENV-2NGC; DENV-3H-87 and DENV-4H-241 and consistently produced MN titres of 10–20 and thus, the cut-off value for a positive test result was a reciprocal serum dilution of 10. Subject results were summarized and presented as geometric mean titres (GMT). Cross-neutralization experiments in which Standard DENV-1-4 sera were each tested against heterologous prototype DENV consistently produced negative results. Standard anti-JE serum (NIBSC 02/182) and serum samples from individuals with diagnosed other flavivirus infection (JEV, MVEV, and KUNJV) were tested against DENV-1-4 and were always negative.
All sequences have been deposited in GenBank and assigned accession numbers KY794785-KY794790.
Three circulating DENV serotypes: DENV-1, DENV-2 and DENV-3 were identified by full length E gene PCR of acute phase serum samples from 17 febrile patients sampled in Madang between 2007–2008 and Lihir in 2010 (Table 2).
DENV-1, DENV-2, and DENV-3 infections were identified in adults and children from Madang town or from villages and rural settlements around the town. In several instances, two serotypes were identified in individuals sampled within a period of 10 days. Two DENV cases originated in Lihir in May and October 2010 –a 47 year-old male Australian traveller infected with DENV-2 and 32 year-old female resident of Lihir infected with DENV-3. These data clearly identify hyperendemic DENV transmission on the northern coast of PNG, in Madang, and on Lihir Island in the Bismarck Archipaleago.
Whole genomes were sequenced from one DENV-2 and five DENV-3 isolates. Phylogenetic analysis of the five DENV-3 isolates indicated that the four from Madang group together to form a previously unidentified lineage within Genotype I (Fig 1). The Madang lineage included other DENV-3 from the region, including a group exported to Solomon Islands and Fiji and collected in 2013 [21] and 2014, respectively. The DENV-3 strain originating in Lihir in 2010 also grouped within Genotype I and clustered with a lineage formed by DENV-3 first found in travellers between PNG and northern Australia [14]. Collectively, these data identify a lineage of DENV-3 endemic to the northern coast of PNG, closely related to DENV-3 circulating in neighbouring Indonesia, and the subsequent introduction of this lineage into the western and south Pacific. The entire lineage is well supported and distinguished from other viruses within Genotype I by seven amino acid substitutions spanning several genes [prM, E, NS2A, NS3, NS5] (Table 3) and are indicated by the highlighted branch in Figure I. Several of these substitutions were non-conservative, involving the replacement of one encoded amino acid with another of very different properties. At least two of these substitutions were fixed through the process of positive selection according to at least one statistical test implemented in HyPhy [22] and the online version Data Monkey (http://www.datamonkey.org/ [23, 24].
The Lihir 2010 DENV-2 virus grouped with the Cosmopolitan genotype (Fig 2) and clustered with DENV-2 originating in Makassar, Indonesia in 2007 [25], and identified in travellers between PNG and northern Australia [14]. Cosmopolitan DENV-2 circulates widely in South and Southeast Asia, Africa, the Middle East, and northern Australia, and these data illustrate introduction of DENV-2 into the Western Pacific region from Southeast Asia.
Overall dengue seroprevalence at the two study sites was 85.3%. Assessment of convalescent sera from residents of Madang identified previous DENV infection in the great majority of samples tested (101/119 sera; 84.9%), most of which showed multitypic NAb responses to all four DENV serotypes (Table 4). DENV-2 and DENV-3 GMTs were of greatest magnitude. Lihir sera showed a similar profile (Table 5) to that seen in Madang where the majority of sera were seropositive (38/44; 86.4%), multitypic DENV-1-DENV-4 NAb responses predominated and a majority neutralized all four serotypes. In both locations, monotypic NAb responses were identified in 11% of sera and overall were directed against each of the four serotypes. Seroprevalence increased with age (Table 6, Fig 3), a pattern that is typically seen in dengue-endemic areas.
Although dengue was first described in Papua New Guinea more than 70 years ago when Sabin isolated the prototype DENV-2 strain New Guinea-C from US soldiers deployed along the northern PNG coast during the Second World War [12], the lack of reported case data and DENV transmission data since that time has meant that the distribution of dengue in PNG is not understood. In this study we identified infection with three DENV serotypes among febrile patients presenting to health clinics on the northern coast of PNG, in Madang, over a nine month period in 2007–2008, and infection with two serotypes within 6 months in 2010, in Lihir Island. These findings are consistent with hyperendemic DENV transmission between 2007–2010 and confirm that dengue is endemic in this country.
Our understanding of DENV transmission in PNG has previously been largely limited to detection of DENV in febrile travellers to northern Australia [14]. The Madang DENV-3 Genotype 1 lineage evolved locally and circulated over the year the study was conducted. Our phylogenetic analysis showed that it was most closely related to DENV-3 detected 3.5 years later in the Solomon Islands in January 2103 and then in the Fiji Islands in 2014, likely exported by travellers and having moved over distances of thousands of kilometres. PNG shares geographic borders with Indonesia, where dengue epidemics occur regularly and endemic, locally evolved DENV-3 lineages have been described [5, 25]. The Madang lineage is most closely related to DENV-3 originating in Indonesia, as is the DENV-3 originating in Lihir Island which clustered with DENV-3 identified in travellers to northern Australia.
Similarly, the DENV-2 virus identified in this study in a local resident of Lihir Island in 2010 is representative of strains known to circulate in the region, clustering with a highly similar strain originating in Makassar, Indonesia, in 2007; both viruses belong to a DENV-2 (Cosmopolitan genotype) clade also circulating in neighbouring Singapore in 2008 and identified in travellers to northern Australia in 2004 and 2006. Our results provide further evidence of DENV movement between endemic countries in the Asia Pacific region that has been described by ourselves and others [26–29], and also illustrate the value of sequence data as a means of understanding virus dispersal. PNG may be a source, or at least a place of transit, for dengue to enter the Pacific and further disseminate to other Pacific Island nations.
DENV epidemic virulence has been linked to introduction and transmission of specific serotypes and lineages [6,7]. The DENV-3 lineage we identified to be circulating in Madang in 2007–2008 and which our analysis showed to have evolved locally was subsequently associated with hospitalizations and deaths when it was introduced into the Solomon Islands in 2013 [30], although severe disease was not identified among patients in Madang. This entire lineage was distinguished by several amino acid substitutions that may influence virus phenotypic characteristics and ultimately, disease outcome. Severe dengue has also been linked to genetic polymorphisms including HLA type [31]; PNG and the Solomon Islands both belong to the Melanesian subgroup of Pacific Islanders and although relatively few HLA data are available for this population group, certain allele frequencies are known to be shared. Studies to assess the immunopathogenesis of dengue in PNG would be informative and should be undertaken in future.
Dengue seroprevalence was very high with an overall rate of 85.3%, and increased with age. More than half of seropositive individuals from Madang were greater than 11 years of age and about 40% were older than 20 years. This age distribution reflects opportunities for multiple exposures over the lifetime of the individual in a setting where dengue is endemic or is regularly introduced and indeed, anti-DENV neutralising antibody profiles were predominantly multitypic. A majority of individuals in Madang and in Lihir demonstrated responses to two or more DENV serotypes and most sera neutralized all four DENV serotypes. Serum samples included in this analysis were from two main groups–convalescent patients excluded for acute DENV infection in Madang, and clinic attendees in Lihir (predominantly women attending antenatal clinics) therefore these neutralization data likely reflect multiple DENV infections in the years prior to sampling, and corroborate the genetic evidence for circulation of multiple serotypes. Studies in rural Haiti and Nicaragua [32, 33] have shown that in endemic populations anti-DENV antibody prevalence increases with age and begins to plateau in adolescence reflecting long-term exposure to, and infection with, endemically circulating DENV. In the present study further support for endemic DENV transmission in PNG is indicated by the young age (median 2 years) of individuals in Madang with monotypic anti-DENV NAb responses to DENV-1, DENV-2, DENV-3 or DENV-4 whereas the majority of multitypic infections were in adults. Three individuals with monotypic NAb profiles were babies less than 1 year old and anti-DENV NAb were likely passively transferred maternal antibody. We do not know the age of the mothers but it is possible they were relatively young and/or had experienced a single DENV infection in the past, or that they had experienced multiple infections but that only antibodies to a single serotype were present at sufficiently high levels in the infant to be detectable in our assay.
Hyperendemic DENV transmission is associated with symptomatic dengue infection and with greater incidence of severe dengue [3]. No febrile patients were diagnosed with severe dengue in Madang [15] or Lihir and indeed, severe dengue is rarely identified in PNG despite the hyperendemic transmission we have identified and which has likely been occurring for a significant period of time. In previous studies we demonstrated high seroprevalence and predominantly multitypic NAb responses to all four serotypes in sera collected in New Guinea between 1959–1963 [17] indicating DENV transmission in the decades prior to sampling. The reasons for the rarity of severe disease are unclear but may be related to poor recognition of dengue and the lack of routine dengue surveillance. It is also possible that undefined DENV-specific immune mechanisms may contribute to the apparent rarity of severe dengue disease.
In summary we have identified hyperendemic dengue transmission in PNG in the period up to 2010, which has likely been occurring for many years. We also demonstrate circulation of DENV which has evolved locally, and shown by others to have been introduced into the western and south Pacific in subsequent years. The absence of regular sampling for DENV in PNG and the potential for misdiagnosis of febrile individuals has meant that the evidence consensus for dengue presence is low [2] and the true burden of disease is unknown. A dengue outbreak in Port Moresby was confirmed by the PNG Department of Health in 2016 [34] highlighting the need for additional studies to be undertaken to understand the epidemiology and impact of dengue in this country. An important aspect of this is to understand the origin and transmission patterns of PNG DENV, define endemic and introduced genotypes and lineages, and to characterize epidemic virulence associated with circulation of these viruses among the PNG population and within the Asia Pacific region.
|
10.1371/journal.pntd.0001409 | A Deep Sequencing Approach to Comparatively Analyze the Transcriptome of Lifecycle Stages of the Filarial Worm, Brugia malayi | Developing intervention strategies for the control of parasitic nematodes continues to be a significant challenge. Genomic and post-genomic approaches play an increasingly important role for providing fundamental molecular information about these parasites, thus enhancing basic as well as translational research. Here we report a comprehensive genome-wide survey of the developmental transcriptome of the human filarial parasite Brugia malayi.
Using deep sequencing, we profiled the transcriptome of eggs and embryos, immature (≤3 days of age) and mature microfilariae (MF), third- and fourth-stage larvae (L3 and L4), and adult male and female worms. Comparative analysis across these stages provided a detailed overview of the molecular repertoires that define and differentiate distinct lifecycle stages of the parasite. Genome-wide assessment of the overall transcriptional variability indicated that the cuticle collagen family and those implicated in molting exhibit noticeably dynamic stage-dependent patterns. Of particular interest was the identification of genes displaying sex-biased or germline-enriched profiles due to their potential involvement in reproductive processes. The study also revealed discrete transcriptional changes during larval development, namely those accompanying the maturation of MF and the L3 to L4 transition that are vital in establishing successful infection in mosquito vectors and vertebrate hosts, respectively.
Characterization of the transcriptional program of the parasite's lifecycle is an important step toward understanding the developmental processes required for the infectious cycle. We find that the transcriptional program has a number of stage-specific pathways activated during worm development. In addition to advancing our understanding of transcriptome dynamics, these data will aid in the study of genome structure and organization by facilitating the identification of novel transcribed elements and splice variants.
| Lymphatic filariasis, also known as elephantiasis, is a tropical disease affecting over 120 million people worldwide. More than 40 million people live with painful, disfiguring symptoms that can cause severe debilitation and social stigma. The disease is caused by infection with thread-like filarial nematodes (roundworms) that have a complex parasitic lifecycle involving both human and mosquito hosts. In the study, the authors profiled the transcriptome (the set of genes transcribed into messenger RNA rather than all of those in the genome) of the human filarial worm Brugia malayi in different lifecyle stages using deep sequencing technology. The analysis revealed major transitions in RNA expression from eggs through larval stages to adults. Using statistical approaches, the authors identified groups of genes with distinct life stage dependent transcriptional patterns, with particular emphasis on genes displaying sex-biased or germline-enriched patterns and those displaying significant changes during larval development. This study presents a first comprehensive analysis of the lifecycle transcriptome of B. malayi, providing fundamental molecular information that should help researchers better understand parasite biology and could provide clues for the development of more effective interventions.
| Wuchereria bancrofti, Brugia malayi and Brugia timori are mosquito-borne filarial nematode parasites that cause the tropical disease lymphatic filariasis (LF). The manifestation of the disease ranges from swelling of the lymph nodes to elephantiasis and hydrocele. LF is a major cause of clinical morbidity and disability, leading to significant psychosocial and psychosexual burden in endemic countries. B. malayi is the primary organism for the study of LF because it has a tractable lifecycle that can be replicated in a laboratory setting. Like other filarial nematodes it is a heteroxenous parasite alternating between arthropod vectors and vertebrate hosts. Filarial nematodes are dioecious and reproduce sexually via copulation. Inseminated adult female worms are ovoviviparous and release live larvae (microfilariae) into the lymph, where they eventually circulate in the bloodstream to be taken up by mosquitoes during blood feeding. After a microfilaria (MF) successfully penetrates the midgut of a susceptible vector, it migrates to the thoracic muscles, and develops intracellularly through two molts to achieve the developmentally arrested third-stage larva (L3) that exits the mosquito proboscis during bloodfeeding and subsequently penetrates the mammalian host. Once L3s enter the definitive host, they undergo two additional molts and mature to adults in the lymphatics.
Characterization of the transcriptional program over the complete lifecycle is necessary to clearly understand the development of the parasite and could help devise better target strategies for control. From the standpoint of possibly designing drug-based or vaccine interventions that prevent infection or curtail parasite transmission, there is particular interest in understanding the biology of the L3 to L4 transition in the mammalian host, and the reproductive biology of filarial worms. The completion of the draft genome of B. malayi [1] has ushered in the possibility to use whole-genome gene expression profiling. With that goal in mind, we used next-generation sequencing to comparatively analyze the transcriptome of seven B. malayi lifecycle stages: eggs & embryos, immature MF (of less than 3 days of age), mature MF, L3, L4, adult male and adult female. We find that the transcriptional program has a number of stage-specific pathways activated during worm development and that a number of these are potential targets for drugs or vaccines.
All animal work was conducted according to relevant national and international guidelines outlined by the National Institutes of Health Office of Laboratory Animal Welfare, and was approved under UWO Institutional Animal Care and Use Protocol 0-03-0026-000246-4-6-11; and UWM Research Animal Resource Center Protocol V00846-0-10-09.
Brugia malayi adults and MF were obtained from the peritoneal cavities of patently infected dark-clawed Mongolian gerbils (Meriones uguiculatus) by peritoneal flush with prewarmed RPMI media (Fisher Scientific, Piscataway, NJ). MF were purified by centrifugation through Ficoll-Paque® lymphocyte isolation media (Amersham Pharmacia Biotech, Piscataway, NJ), and washed in PBS three times prior to flash freezing at −80°C. Adult worms were separated by gender, washed three times in RPMI, and flash frozen. Egg and embryo preparations were made by repeated cutting of 10 female worms with a scalpel to release eggs and embryos into a small volume of cold RPMI. The sample was examined microscopically and pieces of uterine tissue were removed using watchmaker's forceps. The sample was washed three times in cold RPMI prior to flash freezing. Immature MF (≤3 days old) were generated and purified as previously described [2]. L4s were isolated from gerbils 12–13 days post peritoneal infection and were processed as described for adult worms. L3s were obtained from the NIAID-NIH Filariasis Research Reagent Resource Center at University of Georgia, Athens, GA.
Total RNA was isolated from the majority of samples using a previously described protocol [2] that combines organic extraction with Trizol LS (Invitrogen, Carlsbad, CA) and column purification (RNAqeous-Micro®, Applied Biosystems, Foster City, CA). Samples were treated with DNase I (Ambion, Austin, TX, USA) according to the manufacturer's instructions, and the absence of background DNA confirmed by using a portion of each sample in a PCR designed to amplify the B. malayi GPX gene [GenBank:X69128] (data not shown). Isolation of RNA from L3s often produces low yields therefore we used a modified protocol employing homogenization of tissue combined with organic extraction in RNAzol [3] followed by cleaning, concentration and DNase treatment using a Zymo Research RNA column (Zymo Research Corp, Orange, CA). For all samples RNA integrity was confirmed visually by agarose gel electrophoresis (data not shown) and purity and concentration determined spectrophotometrically (NanoDrop ND-1000, ThermoFisher Scientific); samples were stored at −80°C. Total RNA was lyophilized under vacuum for transport on dry ice to the Wellcome Trust Sanger Institute Genome Facility.
Polyadenylated mRNA was purified from total RNA using oligo-dT dynabead selection followed by metal ion hydrolysis fragmentation with the Ambion RNA fragmentation kit. First strand synthesis, primed using random oligonucleotides, was followed by 2nd strand synthesis with RNaseH and DNApolI to produce double-stranded cDNA using the Illumina mRNA Seq kit. Template DNA fragments were end-repaired with T4 and Klenow DNA polymerases and blunt-ended with T4 polynucleotide kinase. A single 3′ adenosine was added to the repaired ends using Klenow exo- and dATP to reduce template concatemerization and adapter dimer formation, and to increase the efficiency of adapter ligation. Adapters (containing primer sites for sequencing) were then ligated and fragments size-selected (200–275 bp) by agarose gel electrophoresis. DNA was extracted using a Qiagen gel extraction kit protocol but with dissolution of gel slices at room temperature (rather than 50°C) to avoid heat induced bias. Libraries were then amplified by PCR to enrich for properly ligated template strands, to generate enough DNA, and to add primers for flowcell surface annealing. AMPure SPRI beads were used to purify amplified templates before quantification using an Agilent Bioanalyser chip and Kapa Illumina SYBR Fast qPCR kit.
Libraries were denatured with 0.1 M sodium hydroxide and diluted to 6 pM in a hybridization buffer to allow the template strands to hybridize to adapters attached to the flowcell surface. Cluster amplification was performed on the Illumina cluster station or the Illumina cBOT using the V4 cluster generation kit following the manufacturer's protocol. A SYBRGreen QC was performed to measure cluster density and to determine whether to pass or fail the flowcell for sequencing. This was followed by linearization, blocking and hybridization of the R1 sequencing primer. The hybridized flowcells were loaded onto the Illumina Genome Analyser IIx for 54 cycles of sequencing-by-synthesis using Illumina's v4 or v5 SBS sequencing kit then, in situ, the linearization, blocking and hybridization step was repeated to regenerate clusters, release the 2nd strand for sequencing and to hybridize the R2 sequencing primer followed by another 54 cycles of sequencing to produce paired end reads. These steps were performed using proprietary reagents according to manufacturer's recommended protocol (https://icom.illumina.com/). Data were analyzed using the RTA1.6 or RTA1.8 Illumina pipeline and submitted to Array Express (http://www.ebi.ac.uk/arrayexpress/) under the accession number E-MTAB-811.
Each lane of Illumina sequence was assessed for quality based on %GC content, average base quality and Illumina adapter contamination. To assess the quality of the lane, the mean base quality at each base position in the read was computed over all reads from the lane. To assess %GC content of the reads a frequency distribution of values was plotted. For a single sample in a lane, a GC plot with a normal distribution around the expected GC for the organism would be expected. Any lanes containing a contamination could therefore be identified by the presence of multiple peaks in the %GC plot. To screen for adapter contamination, the sequence reads were aligned to the set of Illumina adapter sequences using BLAT v.34 with default parameters [4]. Any reads matching these sequences were reported as being contaminated with adapter sequence.
Sequence reads from each lifecycle stage were aligned to the genome assembly [GenBank:DS236884–DS264093] using TopHat v1.0.14, a splice junction mapper built upon the short read aligner Bowtie [5], [6]. The pipeline utilized exon records in the genome annotation [1] to build a set of known splice junctions for each gene model, complementing its de novo junction mapping algorithm. Default parameters were used except for the following: minimum intron length was set to 50; minimum isoform fraction filter was disabled; closure-search, coverage-search, microexon-search and butterfly-search were enabled for maximum sensitivity. The resulting alignment files were converted to BAM format and low quality alignments with mapping quality scores less than 5 were removed before downstream analyses [7], [8]. No replicate samples were sequenced and all data were combined per lifecycle stage. Reads aligned to exonic regions were enumerated for each gene model using the HTSeq package (v0.4.7) in Python (www-huber.embl.de/users/anders/HTSeq). Reads overlapping more than one gene model were counted as ambiguous with the mode parameter set as “union”. Following Mortazavi et al. [9], transcript abundance estimates were computed as RPKMs (Reads Per Kilobase of exon model per Million mapped reads) with the following modifications: (i) a set of paired-end reads were counted as one in compiling sequence counts to represent a single sampling event and (ii) TMM (trimmed mean of M)-normalized values were used in place of the nominal library size to account for compositional biases [10]. The correction factors for TMM-normalization (i.e., the weighted trimmed mean of M values to the reference) were calculated using the Bioconductor edgeR package [11]. The weights were from the delta method on binomial data, and the library whose upper quartile is closest to the mean upper quartile was used as the reference.
Differential expression analysis was performed in edgeR by fitting a negative binomial model to the sequence count data. Using the quantile-adjusted conditional maximum likelihood method, dispersion parameters were estimated for each gene as a measure of the overall stage-to-stage variability to facilitate between-gene comparisons. All hypothesis testing was carried out using exact test for the negative binomial distribution with a common dispersion term for all genes. P-values less than 0.01 were considered significant. Dispersion parameters were estimated directly from the count data for comparisons contrasting a single stage or two related stages relative to all other stages. For comparisons between pairs of lifecycle stages, a common dispersion value of 0.2 was used, which is equivalent to allowing within-stage variations in expression levels of up to 45%. This value was chosen based on the level of variability observed between the immature and mature MF samples. Because longer transcripts give more statistical power for detecting differential expression between samples [12], Gene Ontology (GO) analysis was performed using the goseq package that adjusts transcript length bias in deep sequencing data [13]. GO annotation was retrieved from the UniProtKB-GOA database [14], and statistically over-represented GO terms in a given gene list were identified using the Wallenius non-central hypergeometric distribution. Hierarchical clustering analysis was performed using GeneSpring GX (Agilent Technologies). RKPM values for each gene were baseline transformed to the median of all samples, and hierarchically clustered with centroid linkage using Pearson's uncentered correlation coefficient as distance metric.
In total, 104 million paired-end reads (2×54 bp) were generated from polyA-tailed mRNA using the Illumina Genome Analyser IIx (Table S1). Sequence reads were aligned to the genome assembly using TopHat [5], and the number of reads aligned to each gene model was summed yielding relative transcript levels for individual genes. Approximately 50% of the sequenced reads were mapped to the reference genome after low quality alignments were removed; 10% of which were aligned to genomic regions outside of the current gene models. Sequencing depth varied between the lifecycle stage libraries, affecting gene model coverage and the distribution of the read counts per gene model for each library (Figure 1 and Figure S1). Overall, in each library, 8,000–10,000 genes (equivalent to 70 to 90% of the currently annotated gene models) had 5 or more mapped reads. Sequence counts were RPKM (Reads Per Kilobase of exon model per Million mapped reads)-transformed and TMM (trimmed mean of M)-normalized to assist in the interpretation of transcript abundance comparisons between stages and genes [9], [10]. For statistical inferences, however, raw read counts were directly used. Further analysis of our sequence data from a genomics perspective, covering issues related to missing, incomplete or incorrect gene models of the 2007 assembly [1] will be published elsewhere (in preparation).
Our sequencing libraries contained reads that map to the Wolbachia genome [GenBank:AE017321]. However, the study was not adequately designed such that one could quantitatively analyze these reads in a biologically meaningful way. Abundance estimates (inferred from read counts) of these transcripts most likely deviate substantially from their true in vivo levels. Poly-A selection directly affects the relative abundance of non-poly-A Wolbachia transcripts with respect to B. malayi transcripts. Moreover, the nature and extent of the biases introduced by oligo-dT method to the relative abundance levels among the non-poly-A species (with respect to each other) is not well understood, and one cannot assume that these biases would remain uniform among different sample preparations. Another layer of uncertainty stems from the possibility that these “Wolbachia” sequences were transcribed from the B. malayi nuclear genome rather than the endosymbiont as a consequence of the past horizontal gene transfer events, leading to a differential capture of (presumably) poly-A tailed “Wolbachia” transcripts of the B. malayi nuclear origin. However, given the incomplete draft nature of the B. malayi genome assembly and the inherent difficulty in mapping short reads originating from multiple loci that are similar in sequences, it remains challenging to rigorously test this hypothesis in silico.
To investigate the global transcriptional differences between stages and between genes during development, a negative binomial (NB) based model [11] was fit to sequence count data. First, the degree of between-stage differences was assessed globally using a multidimensional scaling (MDS) of all-against-all comparisons in the NB model (Figure 2). The resulting sample relations appear consistent with the expected biological differences between the samples. The MDS plot indicates that, in relative terms, the transcriptome profiles of the immature and mature MF are more similar to each other than either is to other stages. Likewise, the eggs & embryos sample is closely related to the adult female sample, part of which consists of the germ-line cells. Interestingly, this plot also shows how different the transcriptome profiles of adult male and female worms are to each other.
Next, we made between-gene comparisons in terms of overall transcriptional variability across stages. It is generally hypothesized that while some genes are expressed constitutively, genes with specific developmental functions are expressed at specific stages. To quantify the level of transcriptional variation for each gene across the seven lifecycle stages, the NB dispersion parameters were estimated for each gene, and used as a measure of the extra-Poisson, stage-to-stage variability. Genome-wide distribution of the dispersion parameter estimates suggests that the level of transcriptional variation is not uniform across all genes (Figure S2). Although the majority of genes show low to moderate levels of variation, certain groups of genes exhibit a significantly greater level of variation. Approximately 25% of genes have NB dispersion parameter values larger than 1. After ranking by dispersion, genes were partitioned into quarters and designated as Q1 through Q4 in the order of decreasing variability.
To examine genes displaying life stage dependent transcriptional patterns in greater detail, the top 25% most variable genes according to the NB dispersion (i.e., Q1) were subjected to an unsupervised hierarchical clustering (Figure 3A). The resulting heatmap and dendrogram suggest that there are four major transcriptional patterns, each of which corresponds to an increased transcript abundance in (i) female and/or eggs & embryos, (ii) male, (iii) microfilariae, or (iv) late larval stages. The transcriptional patterns identified through the clustering analysis largely recapitulate the sample relations revealed in the MDS plot (Figure 2). To classify genes into these broad but distinct co-expression groups in a statistically robust manner, we performed a series of exact tests for the NB distribution using raw read counts for all genes (Figure 3B). Relying solely on the “shape” of expression patterns derived from RPKM values, without considering how many reads contributed to each pattern, may lead to false-positive findings. We first identified genes preferentially transcribed during single stages by performing exact tests contrasting each individual stage relative to the mean of all other stages. The resulting gene lists were augmented by additional exact tests to include genes displaying increased transcript abundance in two (related) stages with respect to all other stages. At the level of p-value<0.01, mutually-exclusive, non-redundant gene lists were compiled for each group. In total, we cataloged 2,430 genes into groups with distinct life stage dependent transcriptional patterns. Comparing the gene lists to the highly variable genes in the Q1 group suggests that members of the four main expression groups account for ∼80% of the top 25% most variable genes (Figure 3C). Genes that are highly variable in transcript abundance, yet are not assigned to any of the four main groups (n = 563) likely display complex transcriptional patterns falling outside of the four categories. In addition, five direct pairwise comparisons were made between relevant stages to gain further insights into the transcriptomic features associated with (1) sex differences, (2) intrauterine reproductive processes, (3) MF maturation, and (4) late larval development (Figure 3D). Cross-referencing with the previously defined coexpression groups (Figure 3B) indicates that stage specificity is not homogeneous within each group of differentially transcribed genes, highlighting the complexity of the relative transcriptome differences among the lifecycle stages examined in the study. The results outlined above are described in further detail in the following sections.
We identified and compared statistically overrepresented GO terms in groups of genes that differ in their level of transcriptional variation over the lifecycle (i.e., Q1 to Q4) to investigate specific gene sets and functional categories distinctly associated with high levels of transcriptional variation (Table S2 and Figure S2). This analysis identified ‘structural constituent of cuticle’ (GO:0042302) as the most significantly overrepresented GO category among Q1 genes that exhibit high levels of between-stage transcriptional variation. Forty-six cuticle collagen genes are annotated with this GO term, and thirty-three of these have distinct lifecycle stage dependent transcriptional patterns (18 late larval, 12 female/eggs, 2 male and 1 microfilarial; Dataset S1). Additional GO terms overrepresented among Q1 genes include those related to serine type endopeptidase inhibitor (serpin), structural molecule, and kinase/phosphatase activity. By contrast, GO categories associated with protein metabolism, such as translation, protein transport and proteasome complex are significantly overrepresented among genes displaying relatively little transcriptional variation over lifecycle stages (i.e., Q2-4).
Although transcript levels of 990 genes are significantly higher during larval stages, 886 and 554 genes display elevated transcript abundance in adult male, and adult female and/or eggs & embryos, respectively (Figure 3B). A direct pairwise comparison of male versus female transcriptome further indentified 1,279 genes with male-biased expression and 651 genes with female-biased expression (Figure 3D). At the level of GO categories, structural molecular activity and those associated with protein phosphorylation and dephosphorylation are prominent among genes preferentially transcribed in adult male. A closer look at individual genes with male-biased expression reveals that major sperm proteins are largely responsible for driving the statistical significance of structural molecular activity (GO:0005198) in these comparisons. By contrast, structural constituents of cuticle (collagens), transcription factor/regulator activity, nuclear receptor activity and serpin activity constitute a main theme of the overrepresented functional categories among genes preferentially transcribed in adult female and/or eggs & embryos.
In an effort to elucidate female germline-enriched transcripts and gain insight into intrauterine reproductive processes, the transcriptome profile of a library enriched for eggs and embryos was compared with that of whole adult female (Figure 3D). However, because the eggs & embryos transcriptome is inherently a subset of the adult female transcriptome, this pairwise comparison is almost subtractive in nature and is likely biased against identifying transcripts enriched in germline tissues. On the contrary, detection of female transcripts either not expressed or expressed at lower levels in eggs and embryos likely remains unaffected by this asymmetric sample relation. For this reason, we used the adult male transcriptome profile as an additional reference point to better identify genes showing a germline-enriched expression pattern. We performed a Venn diagram analysis with three datasets: (1) genes with enriched expression in adult female relative to eggs & embryos, (2) genes with enriched expression in eggs & embryos relative to adult male, and (3) genes with enriched expression in adult female and/or eggs & embryos relative to all other stages (Figure S3). We considered genes belonging to the first set to exhibit somatic tissue-enriched expression pattern, and those belonging to either of the last two sets, but excluded from the first set, to exhibit germline-enriched expression pattern. Based on these criteria, 788 and 239 genes show enriched expression in female germline and somatic tissues, respectively. GO term overrepresentation analysis indicates that functional categories, such as transcription factor activity, DNA binding, regulation of transcription and nuclear receptor activity are more frequently found among genes displaying germline-enriched expression. On the contrary, genes implicated in chloride transport, lipid binding, and proteolysis are overrepresented among those with somatic tissue-enriched expression pattern (Table S3). Interestingly, structural constituents of cuticle (GO:0042302) is overrepresented among both genes with germline-enriched and somatic tissue-enriched expression patterns. A closer look at individual genes reveals that mutually exclusive subsets of collagens are overrepresented in each gene set.
When compared across all stages, transcript levels of 148 genes are distinctly elevated during the MF stage. Overrepresented GO terms in this group include zinc ion binding, nucleic acid binding, chitinase activity, and proteolysis (Figure 3B and Table 1). Most notably, among these are 44 genes that encode proteins with C2H2-type zinc finger domains. There are 195 zinc finger protein genes annotated in the B. malayi draft genome, some of which have high transcript levels in stages other than MF (i.e., 3 late larval, 17 male and 6 female/eggs). In a similarly biased manner, 3 out of 4 endochitinase genes identified in the current B. malayi genome show transcriptional increase during MF stages. Diverse classes of proteases are also represented in this gene set (e.g., cathepsin L-like proteases including Bm-cpl-6, papain cysteine protease family, metalloprotease I, aspartyl protease and trypsin-like protease). Direct comparison of immature and mature MF (IM and MM) indicates that 126 genes show differential transcript abundance between the two samples (Figure 3D). Many different metabolic genes are found in the IM overexpressed gene set, while the endochitinases are overrepresented in the MM.
We identified 842 genes displaying increased transcript abundance during L3 and/or L4 stages relative to other lifecycle stages (Figure 3B). Functional categories overrepresented among these genes include structural components of the cuticle, oxidoreductase activity, serpin activity, chloride transport, hedgehog receptor activity, glycogen biosynthetic process, and proteolysis. As suggested by the last GO category, various proteases (e.g., metalloprotease, papain family peptidase, zinc carboxypeptidase family and cathepsin-like cysteine proteases, including Bm-cpl-1,4 and 5) are prominently represented in this gene set, a pattern similarly found in the MF transcriptome. A pairwise comparison of the transcriptomes of late larval stages indicates that 342 genes have elevated transcript levels in L3s, and 155 in L4s. At the level of functional categories, cysteine-type peptidase activity (e.g., cathepsin-z and -L like proteases) and serpin activity are overrepresented among L3-enriched transcripts, whereas structural constituents of the cuticle and cellular component organization are overrepresented among L4-enriched transcripts (Table S3). In addition, our data indicate that abundant larval transcripts (Alt1.2 and Alt2) show increased abundance in L3s relative to L4s.
Using high-throughput sequencing, we have undertaken a comprehensive genome-wide survey of the developmental transcriptome of the human filarial parasite B. malayi. Although deep sequencing data are highly informative in identifying novel transcribed elements and splice variants that help improve genome annotation [15], the present study aims to characterize transcriptome changes along the progression of the parasite's lifecycle. Transcriptome changes mediating cuticular molting likely represent one of the most notable developmental transitions in RNA expression. Like all nematodes, Brugia spp. have five lifecycle stages that are punctuated by molting of the collagenous cuticle. The tightly regulated process of molting involves cell signaling within the hypodermis to cue secretion of the new collagenous cuticle, shedding of the old cuticle and proteolytic remodeling of the new cuticle [16], [17]. Analysis of overrepresented GO terms highlights structural cuticle components, extracellular matrix components and cysteine-peptidase inhibitors, among others, in genes with high levels of transcriptional variation over the lifecycle (Table S2). In particular, the cuticle collagen gene family displays distinct dynamic transcriptional patterns over the course of the lifecycle, likely reflecting compositional variation in cuticular structure among different life stages. Besides these structural components, genes displaying the most dramatic transcriptional variation in our data set are likely associated with developmental processes that differ between the larval and the adult stages and/or between the genders (e.g., gametogenesis). By contrast, genes constitutively expressed over the developmental period studied frequently have predicted cellular functions related to protein expression, modification and transport, possibly representing core cellular processes that are essential to the survival of cells independent of the lifecycle stage.
The present study indicates that genes exhibiting adult male enriched transcriptional pattern (relative to adult female and/or other stages) show strong statistical bias towards GO categories related to cytoskeleton, structural molecule activity, protein phosphorylation and dephosphorylation (Table 1). Many of these gene sets and functional categories are highly represented among classes of male-enriched transcripts in parasitic nematodes [18], [19], [20], [21] and have been identified in the Caenorhabditis elegans male and hermaphrodite germline as being involved in spermatogenesis [22]. Nematode sperm are unique in that they utilize a nematode-specific cytoskeletal element, major sperm protein, for ameboid motility. It is hypothesized that because mature nematode sperm lack ribosomal elements, the phosphorylation and dephosphorylation of molecules by a host of enzymes within the differentiated cells could promote maturation and pseudopod extension [22]. Seven of the genes found to be differentially expressed in male worms in our study were also found in a microarray comparison of adult male and female worms [23], and were shown by in situ localization to be expressed either in sperm or vas deferens tissue of adult male worms and not in gravid adult female worms [24]. If we compare our RNA-seq data with recent microarray work comparing gene expression in adult male and female B. malayi [19], 515 of our 1,276 (40%) genes with male-biased expression match with male up-regulated genes found in the microarray comparison, and 150 out of the 651 (23%) genes with female-biased expression match the microarray findings.
In filarial nematodes, fertilization is internal and gravid females hold oocytes, sperm, zygotes, developing embryos, and MF in their uteri. Structural constituents of cuticle, transcription factor activity, DNA binding, and regulation of transcription emerged as notable themes in our analysis of overrepresented functional categories among genes with increased transcript levels in adult female and/or eggs & embryos (Table 1). These are likely relevant in the context of embryogenesis. Pairwise comparison of adult female with adult male presents us with a similar but more expanded view on features of genes displaying female-enriched expression (Table S3). Further comparisons with genes displaying germline-enriched expression patterns suggest that many of the female-biased transcripts, and more importantly, the majority of the above mentioned functional categories are attributable to the characteristics of the germline transcriptome. For instance, 33 out of 34 genes annotated with transcription factor activity (e.g., nuclear hormone receptors and homeobox domain containing proteins) that are enriched in female and/or eggs & embryos, have a distinctly germline-enriched expression pattern. Bm-fab-1 (Bm1_33050), an embryonic fatty acid binding protein transcript previously found to be female-associated by differential display PCR and whose protein localizes to embryos [25], [26] also exhibits a germline-enriched expression pattern.
Much of our current information on molecular aspects of filarial reproduction comes from microarray and PCR-based transcriptome comparisons between whole adult male and female worms. These studies were based on the assumption that gender-associated transcripts arise from the reproductive organs and their contents. Our data suggest that such an assumption is not wholly unreasonable but may not always hold true. Out of 651 female-enriched transcripts we identified (in comparison to male), 82 display somatic tissue-enriched expression patterns, and it is likely that some of these transcripts are truly not derived from the germline tissues. Spatial expression patterns have not been confirmed for the majority of gender-associated B. malayi genes, and a growing body of research on nematode neurobiology and extracellular signaling lends support to the idea that some gender-associated genes can be expressed in non-reproductive tissues. For example, free-living and parasitic nematodes use gender-specific receptors to sense environmental signals, as demonstrated by the presence of anterior chemosensors in male worms that specifically bind female pheromones [27], [28]. Nematodes also store fat in intestinal cells, which may act as endocrine organs involved in germline signaling and are triggered by activation of intestinal cell nuclear receptors by lipophilic hormones [29], [30], [31].
On the other hand, these observations are not inconsistent with the possibility that some somatic tissue derived transcripts play an essential role in embryonic development or intrauterine reproductive processes. The current study suggests that components incorporated into the embryonic cuticle and the eggshell membrane may be in some part maternal in origin. This interpretation is supported in at least one case where MF sheath protein transcripts in Brugia are detectable by in situ hybridization only in adult female tissues and not in eggs or embryos [24], while the encoded protein is found on the surface of in utero sheathed MF but not in maternal tissues [32]. Other notable transcripts showing enrichment in female somatic tissues in our study include Juv-p120 excretory/secretory proteins and astacin proteases (Bm1_30065; Bm1_13915). Homologs of the latter in C. elegans, nas-4 and nas-9 are found in pharyngeal marginal cells, and in the hypodermis and reproductive tract, respectively [33]. Their functions are unknown but the localizations suggest roles in cuticle and eggshell remodeling.
After expulsion from females, developmentally arrested Brugia MF must undergo a maturation process within the mammalian host to become infective to the mosquito vector [2], [34], [35]. Brugia MF are sheathed in a remnant of the eggshell membrane that is acellular and insoluble, and is composed of chitin and a variety of cross-linking proteins, lipids and polysaccharides [36], [37], [38], [39]. Our data indicate that a large number of transcripts representing DNA-binding proteins with zinc finger motifs as well as several endochitinase transcripts are significantly elevated in MF. Proteomic analysis also revealed a significant enrichment of zinc finger proteins in this stage of the lifecycle [21]. Although the precise role of these DNA binding proteins is unknown, it is tempting to speculate on their involvement in maintaining the developmentally arrested state of circulating MF. Transcriptional increase in proteases and chitin-associated enzymes in MF could be important in the process of casting off the chitinous sheath during or after mosquito midgut penetration [35], [40], [41]. Immunolocalization studies have shown that in sheathed MF, chitinase is stored in the inner body of the MF and secreted to the surface to degrade the sheath upon mosquito infection [42]. Microfilarial maturation is accompanied by transcriptional transitions and changes in the composition of the microfilarial surface [2], [35]. Despite the remarkable change in infectivity, our data suggest that transcriptional differences between IM and MM are relatively small; it is the least pronounced of all pairwise comparisons made in this study (Figure 2 and 3D). Genes involved in ATP synthase activity, tRNA production and cytoskeleton are overrepresented among those that show transcriptional change between IM and MM (Table S3). Although it is difficult to further characterize the exact nature of these changes due to a high proportion of genes with no functional annotation, we hypothesize that a metabolic shift is likely part of the maturation process in anticipation of the transition from the blood of a homeothermic host to the inhospitable midgut and hemocoel of the poikilothermic mosquito vector. It is important to consider that both populations of MF used in this experiment were derived from the peritoneal cavities of infected gerbils. Although we have previously shown a dramatic difference in mosquito infectivity between peritoneally-derived immature and mature MF [2], [34], [35], it is clear that intraperitoneally-derived MF, regardless of age, are considerably less infective than those found in circulating blood [43]. It is possible that the transcriptional profile of mature circulating MF differs from those that are derived from the peritoneal cavity.
Following the introduction of L3s into the peritoneal cavity of gerbils, the L3 to L4 transition requires no migration and occurs approximately 8 days post infection (unpublished). This particular lifecycle transition is of great interest to researchers trying to identify parasite molecules that mediate interactions with the host immune system, and that could be exploited with vaccines to confer protective immunity, or with drugs to prevent infection. Antigens that historically have been of interest in this regard are the ALT (abundant larval transcript) family of potentially secreted larval acidic proteins found predominately in L2 and L3 stages [44], [45], [46]; the L3 cystatin cysteine protease inhibitor family, Bm-SPN-2, TGF-β homologues, macrophage inhibition factor and Bm-VAL-1 [46]; troponin, tropomyosin and cuticular collagens [47]; Onchocerca volvulus activation associated secreted protein (Ov-ASP-1) [48], onchocystatin (Ov-CPI-2) [49] and Ov-SPI-1 [50], and B. malayi glutathione-s-transferase [51]. One hypothetical protein found to be L3 specific in our experiment, Bm1_38105, was also highly ranked as a potential drug target [52].
In the present study, the transcriptome of developmentally arrested, vector-derived L3s was compared to that of peritoneally-derived L4s at 12–13 days post infection. Comparing our RNA-seq data to a recent microarray experiment [53] that assessed transcriptomes of vector-derived L3s to cultured and irradiated L3s, shows that 29 genes are shared, and likely constitute genes required for L3 survival in mosquitoes. These include Alt-2 and Alt1.2 proteins, cathepsin L precursors, Bm-col-2, cystatin, microfilarial surface associated protein, metabolic proteins and BmSERPIN. The differential expression of cathepsins Bm-cpl-1, 4, and 5 in vector stage L3s is supported by EST sequences and these genes are grouped phylogenetically into a distinct clade (Ia) separate from other nematode cathepsin-like proteases [54]. There is strong evidence that these proteins play important roles in the L3 to L4 molt, because targeting the cpl-1 gene in O. volvulus by RNAi decreased the rate of molting [55], and suppression of the cathepsin L-like cysteine protease transcript by injection of siRNA or dsRNA into infected mosquitoes carrying L2 and L3 stages of B. malayi retarded worm growth, disrupted development and resulted in cuticular sloughing [56]. It is important to point out that the L4s we used were from the peritoneal cavity of gerbils, and did not follow the normal behavioral pathway of intradermal passage and migration to the lymphatics. It is possible that the transcriptional profile of intraperitoneally-derived L4s is different than that of worms found in lymphatics; indeed Chirgwin et al. [57] showed different transcriptional profiles for three L3 genes at 3 days post infection in groups that had been injected intradermally and allowed to migrate naturally to the popliteal lymph node in the gerbil model, and those that were confined to the peritoneum.
In this study we provide a detailed overview of the molecular repertoires that define and differentiate distinct lifecycle stages of the parasite, extending and complementing previously published work on stage-specific gene expression [2], [19], [21], [24], [53], [58], [59]. Inclusion of seven different developmental stage samples uniquely allows us to place specific between-stage transcriptional differences into the broader context of the transcriptomic landscape during the lifecycle of B. malayi. It is important to emphasize, however, that this is just an overview of observations and that these data will be mined by the community to provide specific information on particular gene sets to bring these deep sequencing data into more complete biological context.
Because expression dynamics is an important consideration in the genome-wide assessment of candidate targets for control [52], [60], [61], our comprehensive analysis of transcript abundance over developmental time is a valuable addition to a growing body of genomic and post-genomic resources that guide and support the concerted efforts to develop better intervention strategies.
|
10.1371/journal.ppat.1006582 | Streptococcus pneumoniae in the heart subvert the host response through biofilm-mediated resident macrophage killing | For over 130 years, invasive pneumococcal disease has been associated with the presence of extracellular planktonic pneumococci, i.e. diplococci or short chains in affected tissues. Herein, we show that Streptococcus pneumoniae that invade the myocardium instead replicate within cellular vesicles and transition into non-purulent biofilms. Pneumococci within mature cardiac microlesions exhibited salient biofilm features including intrinsic resistance to antibiotic killing and the presence of an extracellular matrix. Dual RNA-seq and subsequent principal component analyses of heart- and blood-isolated pneumococci confirmed the biofilm phenotype in vivo and revealed stark anatomical site-specific differences in virulence gene expression; the latter having major implications on future vaccine antigen selection. Our RNA-seq approach also identified three genomic islands as exclusively expressed in vivo. Deletion of one such island, Region of Diversity 12, resulted in a biofilm-deficient and highly inflammogenic phenotype within the heart; indicating a possible link between the biofilm phenotype and a dampened host-response. We subsequently determined that biofilm pneumococci released greater amounts of the toxin pneumolysin than did planktonic or RD12 deficient pneumococci. This allowed heart-invaded wildtype pneumococci to kill resident cardiac macrophages and subsequently subvert cytokine/chemokine production and neutrophil infiltration into the myocardium. This is the first report for pneumococcal biofilm formation in an invasive disease setting. We show that biofilm pneumococci actively suppress the host response through pneumolysin-mediated immune cell killing. As such, our findings contradict the emerging notion that biofilm pneumococci are passively immunoquiescent.
| Since its discovery in 1881, invasive disease caused by Streptococcus pneumoniae, the leading cause of community-acquired pneumonia and a prototypical extracellular pathogen, has been tied exclusively to the planktonic phenotype, i.e. individual diplococci or short chains. Herein, we report that heart-invaded pneumococci can instead replicate intracellularly and transition into an immunoquiescent biofilm. Using dual RNA-seq technology we capture the complete gene expression profile of S. pneumoniae within infected hearts as well as the bloodstream. In doing so, we affirm that pneumococci within the heart are indeed forming biofilms and identify virulence determinants that play important roles in vivo. Accordingly, we identify a novel role for the genomic island Region of Diversity 12 in promotion of biofilm formation, virulence, and dampening of the host response. We subsequently show that biofilm pneumococci prevent immune cell infiltration into the heart not by passive means but instead through enhanced fratricide-mediated release of the toxin pneumolysin that kills resident macrophages, pre-empting their response. Collectively our manuscript describes a novel site, i.e. intracellular; previously unreported growth phenotype during invasive disease, i.e. biofilm formation; and counter-intuitive molecular mechanism, i.e. resident macrophage killing; for how pneumococci establish themselves in the heart without inflammation.
| Hospitalization for community-acquired pneumonia (CAP) is an established risk factor for adverse cardiac events that includes heart failure, arrhythmia, and infarction; with as many as one-in-four adults hospitalized for CAP experiencing some form of pneumonia-associated adverse cardiac event (PACE)[1, 2]. Streptococcus pneumoniae (the pneumococcus), the leading cause of CAP [3], has been directly linked to PACE. In 2007, Musher et al. reported that one in five individuals hospitalized for pneumococcal pneumonia experienced PACE and these individuals had four-fold greater mortality than those with pneumococcal pneumonia alone [4]. More recently, Eurich et al. reported that pneumococcal pneumonia was specifically associated with greater incidence of heart failure, even during convalescence, and this persisted for a period of up to 10 years [5]. Thus, clinical studies strongly imply that some form of cardiac damage is incurred during invasive pneumococcal disease (IPD).
During IPD, circulating S. pneumoniae are capable of binding to the vascular endothelium and translocating into the heart [6]. Within the myocardium, invaded pneumococci form what we have termed “cardiac microlesions”, i.e. non-purulent pockets of pneumococci that are typically adjacent to blood vessels. Cardiac microlesions have been shown to disrupt normal electrophysiology and impair contractile function [6–8]. In antibiotic rescued animals, including non-human primates with experimental pneumococcal pneumonia, tissue damage associated with cardiac microlesions resulted in de novo scar formation [6, 9]. Thus, the acute and long-lasting damage caused by heart-invaded S. pneumoniae is one explanation for PACE and the adverse cardiac events that occur thereafter. Pertinent to this study, the cholesterol-dependent pore-forming toxin produced by S. pneumoniae, pneumolysin (Ply), has been shown to play an important role in cardiomyocyte and infiltrated macrophage killing [6–8]. Nonetheless and despite all these advances in knowledge, mechanisms by which pneumococci establish themselves within the myocardium without innate immune cell recognition remains unknown.
In stark contrast to what occurs during pneumonia and IPD, S. pneumoniae forms biofilms in the nasopharynx during colonization and within the middle ear during otitis media [10, 11]. Biofilms are surface-attached communities of bacteria encased within an extracellular matrix (ECM)[12–15]. In the nasopharynx, biofilm growth of S. pneumoniae confers resistance to desiccation and the host immune response [10]. The slower metabolic rate of bacteria in biofilms also confers intrinsic resistance to antibiotic killing, which helps to explain the recalcitrance of otitis media to treatment [11]. Importantly, a growing body of literature demonstrates that biofilm pneumococci elicit a considerably weaker immune response from host cells when compared to their planktonic or biofilm-dispersed counterparts [15, 16]. It has been speculated that this promotes long-term colonization by delaying the onset of adaptive immunity [10, 17]. However, the mechanisms by which biofilms suppress immune cell recognition have not been described. Finally, a role for pneumococcal biofilms during IPD has not been reported. This has instead been attributed to the planktonic phenotype which due to reduced surface area better evade stochastic C3b complement deposition and resultant opsonophagocytic killing [18].
Herein we report that heart-invaded pneumococci present within the myocardium replicate within cellular vesicles and over time transition to a mature biofilm. This is the first report of intracellular replication of S. pneumoniae or pneumococcal biofilm formation in an invasive disease context. Using dual RNA-seq analysis we show that pneumococci within the heart are highly distinct from their circulating counterparts and that virulence gene expression is highly anatomical-site specific; this has key implications on vaccine design. We identify a previously unappreciated genomic island that promotes biofilm formation in vivo and show that heart-invaded biofilm pneumococci establish their non-inflammogenic profile not through passive immunoquiescence as previously speculated, but instead via the rapid killing of cardiac macrophages due to enhanced pneumolysin release. These studies advance our understanding of pneumococcal pathogenesis, shed light on the basis of cardiac damage, and reveal a new role for biofilms and pneumolysin during pneumococcal infection.
To determine the morphogenesis of microlesions, heart sections from mice infected with S. pneumoniae serotype 4, strain TIGR4 were examined by transmission electron microscopy (TEM) (Fig 1A). The smallest assemblages of pneumococci that could be detected between 18 and 42 hours post infection (hpi) were consistently within clear, spherical, and discrete intracellular vesicles 4–8 μm in diameter each containing 5–10 electron dense diplococci spaced 1–2 μm apart. Pneumococci-filled vesicles were frequently adjacent to swollen mitochondria and alongside or within areas of the cardiomyocytes undergoing hydropic degeneration. All vesicles, even those containing hundreds of pneumococci, had equidistantly spaced diplococci with well-defined separation from the host cytoplasm. The largest microlesions appeared to be the result of bacterial replication along with the expansion, budding, and merging of smaller vesicles. Larger microlesions also had remnants of vesicular membrane present throughout and were consistently associated with cellular debris both within the vesicles and on their periphery. In all instances, larger microlesions were not associated with immune cells.
Considerable bacterial heterogeneity was evident within the larger microlesions. We observed an accumulation of dead, i.e. ghost pneumococci (S1 Fig), and differences in regards to the presence of capsular polysaccharide. While pneumococci within the smallest vesicles exhibited uniform capsule distribution (Fig 1B.1), those at the periphery of the larger microlesions had capsule only at one pole (Fig 1B.2), and those at the center of largest microlesions had little to no detectable capsule present (Fig 1B.3). Strikingly, captured TEM images of TIGR4 cardiac microlesions strongly resembled those previously reported for in vitro formed biofilms (Fig 1C)[19]. Similarities included the equidistant spacing of pneumococci, the accumulation of ghost cells, and reduced presence of detectable capsule. Of note, serotype 6A, strain 6A-10 did not form cardiac microlesions following mouse challenge, despite the presence of detectable bacteria in the myocardium (S2A Fig). TEM imaging of 6A-10 infected hearts instead showed pneumococci within macrophages adjacent to the vasculature (S2B Fig).
Given the morphological similarities of TIGR4 within mature cardiac microlesions to those within in vitro biofilms, we tested the former for salient biofilm properties. Heart-isolated pneumococci (HIP) were intrinsically resistant to antimicrobial killing; a phenotype absent in blood-isolated pneumococci (BIP) from the same mouse (Fig 2A). This resistance to antimicrobials was lost following >1 hour of HIP outgrowth in THY. We also detected extracellular DNA, an established biofilm ECM component [13], within cardiac microlesions (Fig 2B, S3 Fig). In the nasopharynx, the absence of glucose and presence of neuraminidase-exposed terminal galactose on mucosal epithelial cells has been shown to promote pneumococcal biofilm formation by de-repressing Streptococcal pyruvate oxidase (SpxB)-mediated metabolism. SpxB has also been demonstrated to be required for S. pneumoniae biofilm formation within the nasopharynx [20]. Staining of control and infected heart sections with a fluorescent lectin specific for terminal galactose showed this carbohydrate was exposed only in areas of the heart immediately adjacent to cardiac microlesions (Fig 2C). What is more, a TIGR4 spxB deficient mutant failed to form biofilms in vitro (Fig 2D) and cardiac microlesions (Fig 2E) in vivo despite causing sustained bacteremia for 42 hours and adhering to rat brain capillary endothelial cells, i.e. RBCEC6 cells, in vitro at normal levels (S4A and S4B Fig). Based upon this collective body of evidence we conclude that cell invaded TIGR4 replicate within small vesicles and transition to a biofilm during mature cardiac microlesion development. Of note, terminal galactose exposure was not evident in mouse hearts having sterile injury due to myocardial infarct (S5 Fig). Thus, the exposure of galactose in areas surrounding microlesions was not due to tissue injury.
Pneumococci are phase-variable and stochastically alternate between opaque and transparent colony phenotypes. During nasopharyngeal colonization and in vitro biofilm formation, the transparent phenotype is selected for due to an enhanced capacity to adhere to surfaces. In the lungs during pneumonia and the bloodstream during IPD, the opaque variant instead dominates due to its enhanced resistance to opsonophagocytic killing [19, 21–23]. Along such lines, the percentage of transparent pneumococci present in blood of TIGR4-infected mice decreased from the inoculum of 84% to 25% during the course of infection (Fig 3A). In contrast and within the same animals, the majority of pneumococci isolated from hearts were transparent over time (Fig 3A). We subsequently hypothesized that translocation into the heart was a selective event and required pneumococci with a tissue-tropic phenotype. If true, then mice infected with HIP should experience greater cardiac invasion than those infected with BIP (modeled in Fig 3B). Strikingly, HIP-infected animals had a 2.5-fold greater number of cardiac microlesions than their BIP-infected counterparts despite equivalent levels of bacteremia (Fig 3C–3E). In vitro HIP and BIP exhibited comparable adhesion to and invasion of RBCEC6 cells, yet HIP had enhanced adhesion to and invasion of HL-1 cardiomyocytes (Fig 3F, S6 Fig). Thus, TIGR4 isolated from the myocardium were not primed for translocation across vascular endothelial cells and into the heart, but were instead better suited for cardiomyocyte-related interactions.
To elucidate how TIGR4 in the heart differed from those in the bloodstream we performed deep-sequencing of cDNA derived from the infected organs. Total RNA from intact hearts of HIP-infected mice (~107−8 CFU per heart) and pooled blood (~107 CFU/mL) from TIGR4-infected neutropenic mice was used to generate cDNA and ≥300 million RNA-seq reads were performed per biological sample. The normalized number of RNA-seq reads mapping to each TIGR4 gene (RNA reads per kilobase of transcripts per million mapped reads, RPKM) is a direct correlate of individual gene expression levels (S1 Table). For pairwise differential expression analyses, RNA-seq reads were first sub-sampled to match the condition with the lowest number of reads; these normalized RPKM values are provided in S2 Table. It is of note that this approach yielded RNA reads that corresponded to ~90% of the TIGR4 genome. This approach also yielded an exhaustive in-depth reading of transcripts corresponding to infected mice. All RNA-seq data generated as part of this study is available through the NCBI Gene Expression Omnibus (GEO) database (see methods).
While the results provided in S1 and S2 Tables allow for detailed genome-wide comparisons between the HIP and BIP transcriptomes, herein we focused only on established virulence determinants that would inform us on the pathogenic process (Fig 4). Virulence determinants with the highest levels of gene expression (i.e. >1000 RPKMs: corresponding to ~10% of encoded genes) in the heart were: Pneumococcal adhesion and virulence protein B (PavB; SP_0082), Pneumococcal surface protein A (PspA, SP_0117), the capsular polysaccharide biosynthesis locus (SP_0346–0360), Zinc metalloprotease B (ZmpB, SP_0664), SpxB (SP_0730), Ply (SP_1923), Autolysin (LytA, SP_1937), and Pneumococcal choline binding protein A (PcpA, SP_2136) (Fig 4A). The high levels of gene expression for these virulence determinants suggest that they play an important role within the heart. In stark contrast, the genes encoding PavB and pneumolysin were negligibly expressed within blood. Notably, the genes encoding Pneumococcal pilus-1 (RlrA pathogenicity islet, SP_0462–0468), Pneumococcal serine-rich repeat protein (PsrP) and its accessory proteins (psrP-secY2A2, SP_1755–1772), and Choline binding protein A (CbpA, SP_2190) had minimal expression in both heart and blood. Thus, unequivocal anatomical site-specific differences in virulence gene expression occurred in vivo, with some previously identified key determinants having surprisingly low levels of both blood and heart-related transcription (Fig 4B and 4C). qRT-PCR of a 69-gene panel validated these RNA-seq results (S7 Fig).
We subsequently performed principal component analysis (PCA) using our in vivo RNA-seq data and transcriptome data obtained from in vitro planktonic- and in vitro biofilm-grown pneumococci (Fig 5A). Replicates of the same condition (same color) clustered most closely together on the PCA plot as expected. The second principal component, PC2 (Y-axis) separated HIP and in vitro biofilm pneumococci (bottom of the plot) from the BIP and in vitro planktonic conditions (top of the plot), confirming that the HIP harbored gene expression features more similar to in vitro biofilm pneumococci than to in vitro planktonic pneumococci. Conversely, pneumococci in the blood were more similar to in vitro planktonic pneumococci than in vitro biofilm pneumococci. In addition, pneumococci in vivo also harbored different sets of gene expression features that clearly distinguished them from their in vitro counterparts by virtue of their separation along PC1 (X-axis, left and right sides of the plot, respectively); circos analyses of the RNA-seq data confirmed these relationships (Fig 5B). The gene expression profiles that governed the separation of biofilm-planktonic (S8 Fig) and in vitro-in vivo populations (S9 Fig) are provided. Cumulatively, these findings show that in vivo gene expression profile is anatomical site-specific and drastically different from in vitro.
Driving the disparity between in vivo and in vitro transcription profiles were genes encoded within Region of Diversity (RD)2 (SP_0163 to SP_0171), the second half of RD6 (SP_1057 to SP_1065), and RD12 (SP_ 1947 to SP_1955), which were only expressed in vivo (Fig 5C, S9 Fig). RDs are horizontally acquired genomic islands not present in all pneumococci, many of which have been shown to contribute to virulence [24, 25]. We did not focus subsequent attention on RD6 as prior studies have shown that this 27-kb pathogenicity island encodes piaABCD, the pneumococcal iron-acquisition operon, and phgABC, the pneumococcal hyperosmotic growth operon; both of which were required for pneumococcal survival in the blood [26–28].
We explored whether RD2 or RD12 impacted the ability of TIGR4 to form biofilms and cardiac microlesions. While deletion of RD2 had no discernible effect, deletion of RD12 abrogated biofilm formation in a two-day polystyrene plate model (Fig 6A). RD12 encodes a two-component class II pneumococcal lantibiotic bacteriocin called pneumococcin A1/A2 and its accessory proteins (S10A Fig)[29]. As such, we speculated that RD12 might promote fratricide which is an important aspect of biofilm ECM formation [30]. Indeed, the RD12 deficient mutant (T4ΩRD12) exhibited a stark absence of propidium iodide stainable DNA, typically accessible in dead pneumococci, following growth on glass coverslips (Fig 6B). This occurred despite the fact that deletion of RD12 did not have any long-term growth defects or changes in autolysis activity as determined by bile solubility assay (S10B and S10C Fig). In vivo, mice infected with the RD2 deficient mutant had equivalent levels of bacteria in the blood compared to wildtype infected controls. T4ΩRD12 was however hyper-virulent with ~10-fold higher bacterial titers in blood of infected mice at 30 hours (Fig 6C). This was not due to changes in capsule levels (S11 Fig). In stark contrast to wildtype TIGR4, mice infected with T4ΩRD12 demonstrated profuse bacterial dissemination in the heart (Fig 6D) accompanied by significantly greater neutrophil infiltration (Fig 6E). T4ΩRD12 did not form biofilms in the heart as evidenced by the loss of intrinsic antibiotic resistance in HIP when compared to paired blood isolates (Fig 6F).
Our results suggest that biofilm formation is in some fashion tied to a muted neutrophil response in the heart. In support of this notion, macrophages exposed to wildtype biofilm TIGR4 produced less chemokine-inducing TNFα and CXCL2 than did macrophages challenged with wildtype planktonic TIGR4 or their biofilm-incapable RD12 deficient counterparts (Fig 6H, S12A Fig). HL-1 cardiomyocytes did not produce meaningful TNFα, CXCL1, or CXCL2 (Fig 6H, S12A Fig) following TIGR4 challenge.
Ply is highly inflammatory [31–33]. Therefore, the observation that TIGR4 biofilms were immunoquiescent conflicted with our results that showed ply was expressed at higher levels in the heart and within in vitro biofilms than blood or in vitro planktonic TIGR4, respectively. Subsequent immunoblot analyses confirmed that biofilm TIGR4 released greater amounts of Ply into the supernatant than did planktonic TIGR4 (Fig 7A). Explaining this discrepancy, we observed that biofilm TIGR4 killed macrophages faster than planktonic TIGR4 (Fig 7B); and that this coincided with a marked reduction in detectable levels of TNFα (Fig 7C) and CXCL2 (S12B Fig) in supernatants. Planktonic T4ΩRD12 was also less cytotoxic to macrophages than were wildtype pneumococci (Fig 6G), consistent with an observed reduction in pneumolysin release (Fig 7A), and previously described greater inflammogenic profile in the heart (Fig 6D and 6E) and in vitro (Fig 6H). Implicating pneumolysin as the principal factor responsible for this phenotype, pneumolysin deficient TIGR4 mutant (T4Δply) grown as biofilm had negligible cytotoxicity (Fig 7B, S12C Fig) and elicited a robust macrophage response (Fig 7C, S12B Fig). What is more, complementation of the pneumolysin deficient biofilm-TIGR4 with exogenous pneumolysin restored cytotoxicity to wildtype levels (Fig 7B) and severely dampened TNFα and CXCL2 production by challenged macrophages (Fig 7C, S12B Fig). Thus in vitro, pneumolysin released by biofilm TIGR4 pre-empted macrophage cytokine and chemokine production by their rapid killing. Of note, 6A-10 which naturally lacks RD12, did not show enhanced release of pneumolysin during biofilm growth, and had a comparatively modest enhancement in capability to kill macrophages and subvert cytokine production as a biofilm (S13 Fig). This perhaps explains its inability to form cardiac microlesions in vivo.
Finally, we examined hearts from TIGR4 and T4Δply infected mice for differences in cardiac macrophage numbers and neutrophil infiltrates. Unfortunately, drastic differences in bacterial burden between cohorts made direct comparisons between these strains and conditions invalid (S14A and S14B Fig). Nonetheless, the rare cardiac microlesion formed within T4Δply infected mice displayed extensive immune cell infiltration when examined by TEM and immunofluorescent microscopy (Fig 7D, S14C Fig). As a work around, we passively immunized mice with neutralizing antibody against Ply and infected mice with wildtype HIP or BIP. Consistent with a prior report [6], antibodies against Ply had no impact on pneumococcal burden in the bloodstream of TIGR4 challenged mice; nor in this instance, within the heart (S14D and S14E Fig). Cardiac sections from HIP-infected mice that received Ply antibody stained positively for macrophages and neutrophils in the area immediately surrounding microlesions. In contrast, cardiac sections from mice not receiving antibody lacked the same, indicating immune cell death and/or less neutrophil infiltration (Fig 7E). Of note, this robust modulation of immune cells was seen only at the interface of biofilm contact with the host cells in the myocardium. At the gross level and using flow cytometry of whole heart extracts as measure, mice that received Ply antibody had more neutrophils in their hearts than did untreated controls following HIP challenge (Fig 7F), whereas the number of cardiac macrophages remained constant (Fig 7G). What is more, untreated mice infected with BIP had the greatest number of cardiac macrophages and neutrophils detected by immunofluorescent microscopy and flow cytometry (Fig 7E–7G, S14F Fig). Thus, biofilm formation in the myocardium by HIP indeed suppressed subsequent immune cell infiltration and this occurred in a highly focal and Ply-dependent manner.
This is the first report to describe intracellular replication of S. pneumoniae and to demonstrate an integral role for biofilms during IPD. Together these are a mechanism by which pneumococci establish themselves within the myocardium and subvert their clearance. Our dual RNA-seq approach, in addition to providing insight in regards to the pathogenic process, revealed anatomical-site specific differences in S. pneumoniae virulence gene expression that may have major implications on antigen selection for future protein-based vaccines. The unexpected observation that biofilm pneumococci in the heart pre-empt immune cell infiltration through rapid resident macrophage killing strongly suggests that the immunoquiescent profile of biofilm attributed to pneumococci at other sites, i.e. nasopharynx, may be through similar means. Thus, this study advances our understanding of pneumococcal pathogenesis and the roles of biofilms and pneumolysin.
TEM examination of infected hearts revealed S. pneumoniae is capable of intracellular replication within the myocardium. This observation was unexpected since S. pneumoniae is prototypical for extracellular Gram-positive bacteria [34]. Although observations of intracellular pneumococci within adenoid biopsy specimens from children with otitis media or rhinosinusitis have been reported [35, 36], and pneumococci are known to translocate across vascular endothelial cells within intracellular compartments [37], this is to our knowledge the first report to suggest intracellular replication as an integral step in pneumococcal pathogenesis. Importantly, the exact cell type(s) affected within an infected heart remains to be elucidated. Most probable candidates based on their abundance include cardiomyocytes, resident macrophages, and/or fibroblasts. Moreover, the origin of the early bacteria-filled vesicle is also unclear. The most likely possibilities are: 1) extracellular uptake in a clathrin-coated vesicle from the cell surface [37], or 2) xenophagy, the process by which a cell directs autophagy against an internalized pathogen [38]. Ongoing studies are focused on identifying the specific cells types involved and molecular basis for cardiomyocyte invasion. We propose that intracellular replication allows heart-invaded pneumococci, initially in a planktonic state, to evade innate and adaptive immune mechanisms as they transition into a biofilm.
Bacteria within biofilms are characterized by intrinsic resistance to antibiotic killing, attachment to a surface, and production of an extracellular matrix [12]; all properties observed for pneumococci within infected hearts. S. pneumoniae biofilms have also been associated with an accumulation of non-viable cells within the biofilm, greater frequency of the transparent phenotype, heterogeneous production of capsule, enhanced adhesiveness to cells, and a requirement for pyruvate oxidase [15, 19, 21, 23]. These features were also observed for pneumococci within infected hearts. Given that pneumococci can be detected within other organs, such as the spleen, following bacteremia [39], it is reasonable to propose that biofilms may be forming in other organs during disseminated infection. Of particular interest are the kidneys, since pneumococcal infection has also been linked to acute kidney injury and long-term dysfunction [40].
Dual RNA-seq is a powerful technique that provided a snapshot of how S. pneumoniae adapts to and interacts with the host in the heart. For sake of brevity, we do not discuss the host response and limit our discussion to the established virulence determinants that provide insight on the pathogenic process. Ply is a pore-forming toxin that kills cells via necroptosis during microlesion formation [6, 7]. As shown by Shak et al and herein, its release by biofilm pneumococci helps the bacterium to form biofilms on the host cell surface [41] and establish residency by killing host tissue-resident macrophages. PavB is a fibronectin-binding MSCRAMM (i.e. microbial surface component recognizing adhesive matrix molecules)[42]. Fibronectin in damaged heart tissue has been conclusively reported and PavB most likely acts as a cardio-adhesin [43]. PspA is the major choline binding protein found on the surface of the pneumococcus and is involved in resistance to complement [44]. Pneumococcal choline binding protein A, PcpA, not to be confused with the adhesin CbpA, has been demonstrated to play a vital role in modulating the host immune response by recruiting myeloid-derived suppressor cells and controlling the inflammatory environment within the lungs during pneumonia [45]. We are currently testing whether this occurs in the heart and is an additional explanation for the lack of immune cells associated with cardiac microlesions. Finally, each of the RDs up-regulated in the heart were associated with a copy of the Tpr/Phr peptide quorum sensing-signaling cassette (SP_0163–0164 in RD2, SP_1057–1058 in RD6, and SP_1946–1947 in RD12)[46]. Other investigators have demonstrated a role for Tpr and its orthologs in biofilm formation and virulence of diverse bacteria [47–49].
The reasons for why an RD12 deficient mutant was hyper-virulent are not fully clear. In the heart, and based upon our in vitro results, we propose that RD12 encoded Pneumococcin A contributes towards biofilm formation by inducing fratricide and it is the resultant release of bacterial products including DNA and Ply that helps to establish the ECM [13, 41]. The latter also contributing to the non-inflammogenic biofilm phenotype that was observed. We and others have previously reported that pneumococci within biofilms elicit a muted host response from nasopharyngeal epithelial cells in vitro and the airways of challenged mice [15, 16]. As such, we initially hypothesized that the absence of immune cell infiltrates within cardiac microlesions was due to some form of passive immune evasion. Our observation that macrophages responded robustly to pneumococci lacking pneumolysin in biofilms disproved the former. Instead, it became evident that the observed muted cytokine response was due to rapid killing of cardiac macrophages by pneumolysin, thereby pre-empting the host response. Importantly, our previous study i.e. Gilley et al. concluded that pneumolysin produced by pneumococci in the heart killed infiltrating monocytes [7]. Herein, we tie the release of pneumolysin to the biofilm phenotype and demonstrate that heart resident macrophages are depleted in a pneumolysin dependent manner. Rapid macrophage death then precludes immune cell infiltration by restricting cytokine and chemokine production. Thus, this study changes our understanding of how biofilms modulate the host response and corrects our prior report by providing a more detailed molecular explanation for the observed immunoquiescent phenotype.
Our cardiac findings are consistent with the known role for Ply in establishing early residency within nasopharynx [50]. They also explain why TEM imaging of nasal septa from colonized mice showed considerably greater mucosal epithelial cell perturbation, yet less cytokine production when compared to mice colonized with a biofilm deficient mutant [15]. Although cardiomyocytes are killed as a result of pneumolysin exposure [6, 8], and the same occurs for respiratory epithelial cells [51], the fact that these cells are not immune cells may explain why their damage or death does not lead to overt cytokine and chemokine production [15]. We therefore propose that biofilm-mediated release of pneumolysin and resident macrophage killing may be an explanation for the low levels of cytokines and chemokines detected during nasopharyngeal colonization. This notion warrants testing.
It is important to note that mice infected with strain 6A-10 did not develop cardiac microlesions and instead had pneumococci entrapped within cardiac macrophages. This indicates not all clinical isolates are capable of causing microlesions as seen with TIGR4. The most likely explanation for this difference is the considerable genetic heterogeneity that occurs between different clinical isolates; ~10% of their genomes [24]. For example, 6A-10 does not encode RD12. In fact, we identified RD12 in only five of the S. pneumoniae genomes publically available through PubMed. While the association between severe S. pneumoniae disease, cardiac damage, and adverse cardiac events in humans is now unequivocal [1, 2, 4–6, 9], the exact cardio-pathological hallmarks and how they vary between clinical isolates remains unknown. These are undoubtedly impacted by the genetic content of the invading strain, and based upon our results herein, the resultant biofilm forming ability and strain-specific dynamics of pneumolysin release. Importantly, the role RD12 plays could presumably be compensated for by the other bacteriocin systems known to be present in S. pneumoniae [52].
One striking observation was the anatomical-site specific differences in pneumococcal virulence gene expression. For example, the genes encoding Ply and PavB were expressed in the heart but not the blood. Similarly striking were the major differences between in vivo and in vitro gene expression, which shows that our in vitro condition is not an adequate mimic of the host. As such, a pressing need to expand on our studies and comprehensively determine the transcriptome of S. pneumoniae in the lungs, nasopharynx, and other body sites is now highly evident. Such a comprehensive in vivo gene atlas would not only elucidate the pathogenic process but would also serve to identify pneumococcal proteins that are consistently expressed across body sites. Presumably, these would be among the most suitable vaccine antigens to protect against all stages of pneumococcal disease. For example, that antibody against Ply does not protect against bacteremia following TIGR4 challenge was shown herein following passive immunization and also reported as part of our original manuscript that describes cardiac microlesion formation [6]. Alternatively, PspA, which was highly expressed in both heart and blood and has already been shown to be a protective antigen [53], may be such a candidate antigen.
In summary, we report that cardiac microlesion formation during IPD involves an intracellular stage and requires that the pneumococci transition into a biofilm. Pneumococci within the heart have a unique transcriptional profile that surprisingly does not include several established virulence determinants but included genes within three RDs that were drastically up-regulated in vivo. This highlights the dynamic nature of pneumococcal gene expression in vivo and this should be taken into consideration when considering vaccine antigens. We show that RD12 is important for biofilm formation in vitro and reduced cardio-pathology in vivo. This possibly implicates an important role for bacteriocin systems in human disease. Finally, we show that biofilm pneumococci subvert host response through the rapid killing of cardiac macrophages in a Ply-dependent manner. Moving forward, studies that examine in vivo bacterial gene expression are critical for a fuller understanding of bacterial pathogenesis, host-pathogen interactions and rational vaccine design.
S. pneumoniae serotype 4, strain TIGR4 was the parent wildtype strain used and its annotated finished (gap-free) genome is available [54]. An isogenic mutant lacking spxB (T4 ΔspxB) was created by allelic exchange using a mutagenic PCR construct consisting of the ermB erythromycin cassette flanked by upstream and downstream fragments of the gene as previously described [25]. Isogenic TIGR4 mutants lacking the Regions of Diversity, RD2 (T4ΩRD2), and RD12 (T4ΩRD12) have been previously described [25]. S. pneumoniae serotype 6A, strain 6A-10 and its isogenic mutant lacking ply (6A-10Δply) have also been previously described [15]. Pneumococci were grown in Todd Hewitt Broth (THB) (Acumedia, Neogen) with 0.5% yeast extract (THY) at 37°C in 5% CO2 for experiments.
All mouse experiments were reviewed and approved by the Institutional Animal Care and Use Committees at The University of Alabama at Birmingham, UAB (Protocol # IACUC-20175) and The University of Texas Health San Antonio (Protocol # 13032-34-01C). At both institutes animal care and experimental protocols adhered to Public Law 89–544 (Animal Welfare Act) and its amendments, Public Health Services guidelines, and the Guide for the care and use of Laboratory Animals (U.S. Department of Health & Human Services).
Female 6-7-week-old BALB/cJ mice were challenged with ~103 CFU of exponential phase pneumococci in 100μL phosphate-buffered saline (PBS) by intraperitoneal injection. For studies with Blood Isolated Pneumococci (BIP)- and Heart Isolated Pneumococci (HIP)-infected mice, pneumococci were obtained as described below. Blood for assessment of bacterial burden was obtained by tail bleeds. At fixed time points or when deemed moribund, mice were euthanized by CO2 asphyxiation and death was confirmed by pneumothorax before heart-collection. For passive immunization, mice were administered 3μg of anti-pneumolysin neutralizing antibody, PLY-4 (# ab71810, Abcam) intraperitoneally in 100μL PBS 1 hour before bacterial challenge and 14 hours post- infection.
Excised hearts were washed with ice cold PBS, fixed with phosphate buffered 4% formaldehyde with 1% glutaraldehyde, and then processed as previously described [55]. Once embedded within resin and sectioned at 1μm in thickness, electron microscopy was performed using a JEOL JEM-1230 transmission electron microscope (Peabody, MA).
Blood was collected from anesthetized mice retro-orbitally and transferred to heparin-coated collection tubes. Following euthanasia, hearts were surgically excised and washed in PBS to remove blood. Isolated hearts were homogenized in 5mL of PBS followed by filtration of the homogenate through a 40μm cell strainer. Paired blood (BIP) and strained heart (HIP) samples were flash-frozen at -80°C in working aliquots with 10% glycerol.
Freshly collected BIP and HIP samples were serially diluted in PBS and plated on tryptic soy agar plates supplemented with 100μL of catalase (MP Biomedicals, LLC), and incubated at 37°C in 5% CO2 for 16 hours. Colonies were examined under oblique transmitted light to determine the frequency of transparent versus opaque colony variant [56]. Our TIGR4 parent wildtype strain was composed of 84% transparent pneumococci and 16% opaque pneumococci.
The effect of antimicrobial agents on BIP and HIP was determined using a modified version of the standard micro-dilution assay [19, 57]. These assays were conducted using samples immediately frozen after their collection to preserve their BIP or HIP phenotypes. Thawed aliquots of BIP and HIP were diluted in Dulbecco’s Modified Eagle’s Medium, DMEM (Corning) containing penicillin (0.125μg/mL) or erythromycin (0.5μg/mL) to a final concentration of 103 CFU/mL. BIP and HIP in DMEM without antibiotic were used as controls. At regular time intervals of 1 hour, 5μL of each bacterial suspension was spotted on tryptic soy blood agar plates (Remel, USA) and incubated at 37°C in 5%CO2 for 16 hours. The percentage fraction of antibiotic tolerant pneumococci in each sample per time point was calculated as: (# recovered CFU in antibiotic / # CFU in non-antibiotic control) x 100.
Hearts collected from infected mice were washed thoroughly with PBS then embedded in cassettes with Optimal Cutting Temperature Compound (Tissue-Tek, 4583). Frozen 7μm thick cardiac sections were fixed with 10% neutral buffered formalin, permeabilized in 0.2% Triton X and blocked with PBS containing 5% serum from species to which the secondary antibody belonged (blocking buffer). Sections were then incubated overnight at 4°C with blocking buffer containing a 1:1000 dilution of primary antibody: rabbit anti-serotype 4 capsular polysaccharide antibody (Statens serum Institut: cat #16747), or rabbit anti-mH2A.1 antibody (EMD Millipore, cat#ABE215). The next day sections were vigorously washed with 0.2%Triton X and then incubated for 1 hour at room temperature (RT) with blocking buffer containing secondary antibody at 1:2000 dilution: FITC labeled goat α-rabbit antibody (Jackson Immuno Research, cat#111-096-144), or rhodamine labeled donkey α-rabbit antibody (EMD Millipore, cat#AP182R). For neutrophil and macrophage staining, cardiac sections were incubated with rat α-mouse Ly-6G primary antibody (clone 1A8; BD Biosciences) and rat α-mouse CD107b primary antibody (clone M3/84; BD Biosciences) diluted at 1:500 in the blocking buffer for 1 hour at RT. After incubation, sections were washed and incubated with Alexa 488- conjugated goat α-rat secondary antibody (Jackson Immuno Research, West Grove, PA) diluted at 1:500 in blocking buffer for 45 minutes at RT. Exposed galactose was stained for with fluorescein labeled Erythrina crystagalli lectin (Vector Laboratories, FL-1141) after blocking with carbo-free blocking solution for 1 hour (Vector Laboratories, SP-5040). All slides were stained with DAPI (Molecular Probes by Life Technologies, R37606) mounting the sections with FluorSave (Calbiochem: 345789) and covering with coverslip for visualization. Images of cardiac sections were captured at The University of Texas Health San Antonio using a Zeiss LSM 710 confocal microscope and at UAB using a Leica LMD6 microscope equipped with DFC3000G monochrome camera. Image stitching of whole IFM stained cardiac sections was performed using the Leica LASX software. When indicated, cardiac microlesions were enumerated by counting foci of pneumococci in three capsule stained heart sections, each section at least >50μm apart. The first section was cut 300 μm into the heart from the surface.
Static pneumococcal biofilms were grown in 6-well polystyrene plates (COSTAR) as previously described [15]. FilmTracer LIVE/DEAD Biofilm viability kit (Invitrogen, L10316) was used to determine viability of 24-hour biofilms grown on 1% BSA coated coverslips as per Manufacturer’s protocols. Biofilms were visualized using a Leica LMD6 microscope with DFC3000G monochrome camera and Z-stacked to construct 3D-images.
Adhesion assays on HL-1 mouse atrial cardiomyocytes (generously provided by Dr. William Claycomb, New Orleans, LA) and RBCEC6 rat brain capillary endothelial cells were performed as described previously [58]. All experiments were performed in triplicates.
For HIP-RNA, hearts were excised from HIP-infected mice (n = 3) when deemed moribund. Hearts were rinsed, diced, fragments washed with ice cold PBS, and the fragments homogenized in RNAprotect bacteria reagent (Qiagen) and stored at -80°C. We empirically determined that high titers of pneumococci were necessary to capture the bacterial transcriptome in the bloodstream when using dual RNA-seq. These levels were not routinely seen following conventional challenge and we resorted to neutrophil depletion to achieve the necessary titers (>107 CFU/mL). Ten mice pre-depleted for neutrophils using anti-Ly6G antibody (BioXCell, clone RB6-8C5) were infected with TIGR4. Blood was collected in RNAprotect bacteria reagent (Qiagen) when the mice were deemed moribund such that blood from 5 mice were pooled as one sample (n = 2 BIP samples) and stored at -80°C. On the day of total RNA isolation from heart homogenates, the samples were thawed and spun down to discard the supernatant. The pellets were further homogenized in 600 μL RLT with B-ME buffer using a motorized mortar for 30 seconds. The re-homogenized samples were then disaggregated in a Qiashredder followed by RNA extraction with the RNeasy Micro Kit (Qiagen) with DNase treatment on column and in solution. The isolated RNA was quantitated using Nanodrop and Bioanalyzer. Samples were then depleted of rRNAs using the Ribo-Zero rRNA Removal Kit for Gram-positive bacteria and human/mouse/rat (Illumina, San Diego, CA).
For the in vitro biofilm and planktonic samples: planktonic mid-log phase (OD620nm = 0.5) TIGR4 grown in THB were used to seed continuous once-flow through biofilm reactors. Biofilms were allowed to grow for 48 hours prior to collection of bacteria. RNA was isolated from the paired planktonic seed cultures (n = 3) and their respective biofilms (n = 3). Total RNA was extracted from each replicate separately using enzymatic lysis of pneumococcal cells (10μL mutanolysin at 25 units/μL, 20μL proteinase K at 20mg/mL, 15μL lysozyme AT 15mg/mL, and 55μL TE) followed by RNA extraction using the same protocol mentioned above. Blood samples were isolated in a similar manner as biofilm and planktonic samples, including enzymatic lysis, except that cells were first disaggregated with the Qiashredder like for the heart samples.
Illumina strand-specific RNA-seq libraries were constructed with the TruSeq RNA Sample Prep kit (Illumina, San Diego, CA) per manufacturer’s protocol. Between 1st and 2nd-strand cDNA synthesis, the primers and nucleotides were removed from the samples with NucAway spin columns (Ambion, Austin, TX). The 2nd strand was synthesized with a dNTP mix containing dUTP. Adapters containing 6 nucleotide indexes were ligated to the double-stranded cDNA. After adapter ligation, the 2nd strand cDNA was digested with 2 units of Uracil-N-Glycosylase (Applied Biosystems, Carlsbad, CA). Size selection of the library was performed with AMPure XT beads (Beckman Coulter Genomics, Danvers, MA). In order to achieve sufficient levels of RNA sequencing for bacterial transcripts in the presence of an abundance of mouse transcripts (dual RNA-seq), libraries were loaded on 150nt paired-end runs of the Illumina HiSeq4000 sequencing platform as follows: half of a channel for each of the BIP and HIP libraries, a quarter of a channel for each of the biofilm and planktonic libraries, and an eighth of a channel for each of the three-control uninfected mouse heart libraries. Raw RNA-seq data and associated in silico analyses for BIP, HIP, biofilm and planktonic TIGR4 are deposited at GEO (accession number GSE86118). Reads from each of the 3 HIP-infected heart samples, 2 pooled BIP samples, 3 biofilm- and 3 planktonic- pneumococci samples were mapped onto the reference S. pneumoniae TIGR4 genome using Bowtie version 0.12.9[59]. The alignment BAM files from Bowtie were used to compute gene expression levels and test each gene for differential expression. The number of reads that mapped to each TIGR4 gene was calculated using the python package HTSeq version 0.4.7 [59]. Differential gene expression analysis was conducted using the DESeq R package version 1.5.24 (available from Bioconductor)[60]. The DESeq analysis resulted in the determination of potential differentially expressed genes when compared between the control planktonic samples and the in vitro biofilm, heart and blood samples, respectively. Read counts for each sample were normalized for sequencing depth (RPKM) and distortion caused by highly differentially expressed genes. Then the negative binomial (NB) model was used to test the significance of differential expression between three pairs of conditions (i.e. in vitro biofilm vs. planktonic, heart vs. planktonic, and blood vs. planktonic). The differentially expressed genes were deemed significant if the False Discovery Rate (FDR) was less than 0.05, the gene expression was above the 10th percentile and showed greater than 2-fold change difference (up-regulated or down-regulated) between the paired conditions.
We performed PCA analysis (R statistical software v2.15.2) on log10(RPKM) gene expression values across all growth conditions and all replicates for each condition, i.e. in vitro planktonic (n = 3), in vitro biofilm (n = 3), heart (n = 3) and blood (n = 2) samples. This analysis was heavily skewed by genes that showed no expression at all in vivo (RPKM = 0). In order to circumvent this problem, we used a cutoff of RPKM>1 across all conditions prior to PCA analysis. This analysis revealed tight clustering of all replicates within a given condition. In addition, principal component 1 (PC1, Fig 5A X-axis) separated in vivo from in vitro conditions, while PC2 (Fig 5A Y-axis) separated biofilm (in vitro and heart) from planktonic (in vitro and blood). It should be noted that the overall depth of coverage of gene expression value interrogation in the in vivo conditions was significantly lower than for in vitro cultures (due to the overwhelming presence of mouse transcripts), explaining the PCA skew when including genes with zero expression. Yet, the distribution of RPKM values across the TIGR4 genome revealed similar average RPKMs of 605 and 520 for in vivo and in vitro samples, respectively. In addition, maximum expression values observed were ~25,000 in vitro in contrast to ~125,000 in vivo. Finally, many of the very highly in vivo expressed genes were part of operons within the RDs described in the results and discussion sections. Therefore, our interrogation of the in vivo gene expression profiles was robust. We then computed correlation values (R statistical software v2.15.2) for all genes to PC1 and PC2 separately. Using a correlation cutoff of 85%, we identified 142 and 105 genes that correlated with PC1 and PC2, respectively (S8 and S9 Figs).
A Circos plot for whole transcriptome comparisons of BIP, HIP, in vitro biofilm- and in vitro planktonic-pneumococci samples for expression levels of 2,105 TIGR4 genes was generated. Read counts were computed for 2,105 genes within each sample using the python package ‘htseq’. Genes with a count of 0 across all samples were excluded resulting in 2090 genes. The read counts were normalized for library size in each sample and a normalized value of counts per million-mapped-reads (CPM) was computed for all genes. Additionally, genes with CPM values < 1 in all samples were excluded resulting in 1969 genes. The mean expression value for each gene was computed within each of the 4 conditions. The average gene expression values were converted to z-scores and were used to rank the genes within each condition. Genes with z-scores ≥ +1 were classified as genes with ‘high’ expression (red color) while genes with z-scores ≤ -1 were classified as genes with ‘low’ expression (green color). The remaining genes were classified as genes with ‘intermediate’ expression. These gene expression values, gene ranks and gene stratification were utilized to generate the circular plots (Fig 5B) using the ‘Circos’ tool version 0.69 [61].
qRT-PCR confirmation of RNA-seq results was conducted using the ABI 7900HT Fast Real-Time PCR System (Applied Biosystems). 69 pneumococcal genes of interest were analyzed by qRT-PCR across all conditions and replicates of the RNA samples used for RNA-seq. Gene expression data was normalized using three genes that were unregulated across all conditions: SP_0378, SP_1489, and SP_1667. The comparative critical threshold (Ct) method was used to determine the ΔCt. The ΔCt was then correlated to the log2(Ratio) of expression from RNA-seq results through a linear regression. The primer sequences used for qRT-PCR are provided in S3 Table.
Freshly excised hearts were minced on ice and digested in serum free Iscoves DMEM supplemented with 2mg/mL of Collagenase Type 2 (Worthington, Cat #LS004176) and 0.02mg/mL of Deoxyribonuclease I from bovine pancreas (Sigma, Cat #DN25-100MG) at 37°C for 30 minutes. These cell suspensions were then neutralized with IDMEM containing 10% FBS and filtered through 0.45μm strainer before blocking with 2.4G2 (BD Pharmingen, Cat #553141) for 30 minutes on ice, and staining with Gr-1-APC (Clone RB6-8C5, eBioscience), CD11b-APC-Cy7 (M1/70, BD Pharmingen), Ly 6C-PerCP-Cy5.5 (HK1.4, eBioscience), Ly6G-FITC (1A8, eBioscience), MHC-II PE-Cy5 (M5/114.15.2, eBioscience), MerTK-PE (DS5MMER, eBIoscience), F4/80-PE-Cy5 (BM8, eBioscience) and CD64-Biotin (X54-5/7.1, Biolegend and used in conjunction with Streptavidin PE-Cy7, eBioscience) antibodies for 30 minutes on ice protected from light. Cells were washed with PBS prior to flow cytometry. The samples were then analyzed on BD LSR-II (UAB Flow Cytometry Core Facility). Neutrophils were identified as Gr-1+CD11b+Ly-6G+F4/80-MHC-II-. Cardiac macrophages were first gated on F4/80+CD11b+ cells and further gated for expression of MerTK and CD64. Flow cytometry data were analyzed using FlowJo software. Percent neutrophils and macrophages were determined as (Percent live gated cells from the heart) x (percent positive).
Percentage cytotoxicity and TNFα, CXCL1, and CXCL2 production by mouse J774A.1 macrophages and HL-1 atrial cardiomyocytes at designated time-points following exposure to an equal biomass of pneumococci (corresponding to multiplicity of infection of ~10 planktonic bacteria in DMEM) were measured by Pierce LDH cytotoxicity assay kit (Thermo Scientific) and ELISA (R&D systems), respectively. To set equal biomass, biofilm-pneumococci were flushed out of the continuous flow-through systems using THY and the optical density of the biofilm suspension adjusted to that of log-phase planktonic bacteria grown in parallel in THY (OD620nm = 0.5).
Samples for western blot quantification for pneumolysin levels were prepared by lysing pellets of planktonic- and flow through biofilm-cultures of S. pneumoniae at equal biomass using pneumococcal lysis buffer (0.01% SDS, 0.1% DOC, and 0.015 M Na-citrate) and concentrating supernatant proteins using acetone precipitation [62]. Samples were frozen with protease inhibitors (Sigma). Equal biomass of pellets and supernatants were loaded after BCA quantification. Isogenic pneumolysin deficient strains were tested as the negative controls. Normalized densitometric quantification of pneumolysin levels in the supernatant was performed using ImageJ processing software.
Statistical analysis of in silico data is provided in the Supplemental Experimental Procedures. For wet lab research, multiple group analyses were performed using One-Way ANOVA Kruskall-Wallis Test with Dunn’s Multiple Comparison Post-test. For all non-parametric data sets, we used a Mann-Whitney test while Student’s t-test was used to analyze parametric data sets. These statistical analyses were performed using Prism 5.0 (GraphPad Software: La Jolla, CA). Data are represented as mean ± SEM. P-value ≤0.05 were deemed significant.
|
10.1371/journal.pntd.0001856 | Serine Protease(s) Secreted by the Nematode Trichuris muris Degrade the Mucus Barrier | The polymeric mucin component of the intestinal mucus barrier changes during nematode infection to provide not only physical protection but also to directly affect pathogenic nematodes and aid expulsion. Despite this, the direct interaction of the nematodes with the mucins and the mucus barrier has not previously been addressed. We used the well-established Trichuris muris nematode model to investigate the effect on mucins of the complex mixture of immunogenic proteins secreted by the nematode called excretory/secretory products (ESPs). Different regimes of T. muris infection were used to simulate chronic (low dose) or acute (high dose) infection. Mucus/mucins isolated from mice and from the human intestinal cell line, LS174T, were treated with ESPs. We demonstrate that serine protease(s) secreted by the nematode have the ability to change the properties of the mucus barrier, making it more porous by degrading the mucin component of the mucus gel. Specifically, the serine protease(s) acted on the N-terminal polymerising domain of the major intestinal mucin Muc2, resulting in depolymerisation of Muc2 polymers. Importantly, the respiratory/gastric mucin Muc5ac, which is induced in the intestine and is critical for worm expulsion, was protected from the depolymerising effect exerted by ESPs. Furthermore, serine protease inhibitors (Serpins) which may protect the mucins, in particular Muc2, from depolymerisation, were highly expressed in mice resistant to chronic infection. Thus, we demonstrate that nematodes secrete serine protease(s) to degrade mucins within the mucus barrier, which may modify the niche of the parasite to prevent clearance from the host or facilitate efficient mating and egg laying from the posterior end of the parasite that is in intimate contact with the mucus barrier. However, during a TH2-mediated worm expulsion response, serpins, Muc5ac and increased levels of Muc2 protect the barrier from degradation by the nematode secreted protease(s).
| Gastrointestinal parasitic worm infections cause significant morbidity, affecting up to a third of the world's populationand their domestic pets and livestock. Mucus, the gel-like material that blankets the surface of the intestine, forms a protective barrier that is an important part of our innate immune system. The whipworm Trichuris is closely associated with the intestinal mucus barrier. The major structural component of this barrier, large glycoproteins known as mucins play a significant role in the expulsion of these worms in a mouse model. Using mice that get longterm chronic infections and others able to expel the worms from the intestine, we uncover a novel role for products secreted by the worms. Enzymes secreted by whipworms can disrupt the mucin network that gives mucus its viscous properties. Moreover, we unravel that worm products are unable to degrade forms of mucins present in the mucus barrier during worm expulsion, suggesting that these enzymes may be released by the worm as part of its regime to improve its niche and survival in the host. However, the host is capable of producing mucins and other protective molecules that protect the mucus barrier from degradation and are detrimental to the viability of the worm.
| Immune mediated elimination of gastrointestinal (GI) parasitic nematodes has been a subject of considerable investigation [1]. Hyperplasia of goblet cells that produce the secreted mucosal barrier is one of the most prominent features of the TH2-type immune response necessary for the expulsion of these pathogens from the intestine [1], [2]. However, until recently, definition of the precise role of goblet cells in host protection remained elusive, especially with regards to the major secreted component of goblet cells, the mucins, which are pivotal to the formation of the mucus layer that overlies the intestinal epithelium. Using established gastrointestinal nematode models Trichuris muris, Trichinella spiralis and Nippostrongylus brasiliensis, we have recently demonstrated that mucins are critical in resolving infection [3], [4], [5]. The major intestinal mucin Muc2 plays a significant role in the concerted protective worm expulsion mechanism and in its absence T. muris expulsion is significantly delayed [4]. Additionally, the Muc5ac mucin, not usually expressed in the murine intestine but induced post-infection during a TH2-type immune response, was demonstrated to be necessary for intestinal worm clearance [3]. Furthermore, Muc5ac was shown to directly affect the viability of the nematode. Bearing in mind that under field conditions GI nematodes can survive for long periods of time, it raises the question of how these parasites interact within the mucosal barrier and subvert the responses against them.
It is well established that GI nematodes secrete a variety of molecules (Excretory Secretory Products, ESPs) into the surrounding niche. These can be highly immunogenic, although, their functions in vivo are not well described [6]. T. muris infection in the mouse provides a unique tractable model that can be used to examine the interaction of parasites with the mucosal barrier during both acute (worm clearance by TH2 immune response) and chronic infection (lack of worm clearance by TH1 immune response) [6]. ESPs are thought to be very well-conserved and similar in terms of their antigenicity in the different Trichuris species. A study by Drake and co-workers has shown that a major 43 kDa protein secreted by the Trichuris nematode has the ability to induce ion-conducting pores in a lipid bilayer [7], whereas other studies have attributed the tunnel formation through the intestinal epithelium to protease activity (zinc metalloproteases, thiol protease and phenol oxidase) present in ESPs [8], [9].
In this study, we aimed to investigate the effect of the T.muris ESPs on the mucus barrier and in particular, on mucins which are an essential part of the co-ordinated TH2-mediated worm expulsion response. Our findings demonstrate for the first time that serine protease(s) secreted by the nematode have the ability to degrade Muc2 and depolymerise the mucin network. In contrast, the mucus barrier formed during worm expulsion is protected from ESP degradation. Particularly, Muc5ac mucin, which is necessary for worm expulsion, was resistant to ESP protease activity. Interestingly, in the mice that expel the nematodes, up-regulation of host serine protease inhibitors (Serpins) is observed which probably provides an additional level of protection of mucins, particularly Muc2, from degradation.
These data, therefore, demonstrate that GI nematode parasites produce protease(s) that degrade the major structural scaffold of the mucus barrier during chronic infection, resulting in a more porous mucus barrier, which in turn can aid establishment and persistence of nematode infection and/or exacerbate inflammation.
BALB/c (Harlan Olac) mice were maintained in the Biological Services Unit at The University of Manchester in a conventional clean Helicobacter hepaticus- and norovirus-free facility. All mice (6–12 wk old) were kept in sterilized, filter-topped cages, and fed autoclaved food in the SPF facility.
The protocols were employed at the University of Manchester and were performed in accordance with guidelines set and approved by the Animal Procedures Committee and Home Office Scientific Procedures Act (1986), United Kingdom under the personal licence issued to SZH (No. PIL: 40/9777).
The techniques used for T. muris maintenance and infection were described previously [10]. Mice were orally infected with approximately 150 eggs for a high dose infection and <15 eggs for a low dose infection. Worm burdens were assessed by counting the number of worms present in the caecum. Worms present in the caecum on day 12 confirmed that infection had established in both groups of mice (high dose and low dose infection). Adult worms present in the caecum of the low dose infected group of mice, on day 35 post infection, confirmed chronic infection (Figure S1). ESPs were isolated ex vivo from adult T. muris nematodes using the method previously described [11]. ESPs were separated into <5 kDa, 5–50 kDa, 50–100 kDa and >100 kDa fractions using size exclusion Amicon columns (Millipore Pty Ltd, Australia).
Human intestinal adenocarcinoma LS174T cells (European Collection of Cell Culture, Salisbury, U.K) were used as a source of glycosylated MUC2. Cells were cultured with complete medium containing DMEM, 2 mM L-glutamine, 100 U/ml penicillin, 100 µg/ml streptomycin and 10% heat inactivated FBS (all from Invitrogen, Paisley, U.K) at 37°C in a humidified incubator (95% air with 5% CO2). When confluent, cells were gently scraped until dislodged from the flasks surface and rinsed with approximately 5 ml of complete media. The harvested cells were dispersed using a 25G needle and resuspended at approximately 1×106 cells/ml [12].
MKN45 gastric cancer cells that express mucin-specific chaperones but no endogenous MUC2 were transfected with the murine Muc2 N-terminal D3 domain (rMuc2-D3) consisting of 3 FLAG tags at the N-terminus and a myc tag at the C-terminus as previously described [13]; rMuc2-D3 was obtained from cell lysates using TRIS-HCl and Triton-X lysis buffer. Conditioned media was collected from transfected MKN45 cells and concentrated to obtain secreted rMuc2-D3.
Immunodetection was carried out using a polyclonal antibody raised against a murine Muc2 (mMuc2) [4] or human MUC2 (hMUC2) [14]. Commercially available 45M1 antibody was used for the detection of mouse Muc5ac [3] and, mouse monoclonal FLAG antibody-clone 2 (Sigma-Aldrich, UK) and c-Myc antibody (9E10) was used to detect rMuc2-D3 [13].
To isolate the mucus from mice, the caecum was gently flushed with PBS to remove the faecal matter, subsequently scraped lightly with a cell scraper into equal volumes of PBS and stored at −80°C until required.
To isolate the mucus from LS174T cells, media was removed and cells were washed vigorously with ice cold PBS, mucus was stored at −80°C until required.
Mucus, mucins and rMuc2-D3 were incubated at 37°C with the ESPs for various time points (as specified) at final concentration of 50 µg/ml or above. Control samples were not treated with the ESPs, but were incubated at 37°C for the maximum time point. ESPs were heat inactivated at 100°C before incubation or incubated at 4°C with mucus as negative controls. ESP activity was quenched using the protease inhibitors: ethylenediaminetetraacetic acid (EDTA), N-ethylmaleimidide (NEM), Lupeptin, Chymostatin and Antipain at 50–150 µg/ml (Sigma-Aldrich, UK).
Caecal tissue was cut longitudinally, washed with PBS and kept hydrated with PBS in a 6 well plate. ESPs were applied topically at 50 µg/ml concentration for 2 h prior to analysis. 0.1 µm blue fluorescently labelled polymer microspheres (Dukes Scientific, UK) were placed on top of the luminal surface of the caecum (set as a reference) and their position analysed using a Nikon C1 Upright confocal microscope [4]. Five measurements were taken per caecum. 3D optical stacks were taken every 5 µm and combined to obtain a Z-stack at the time points stated. All images were analysed using the EZ-C1 freeviewer software (version 3.9).
Isolated mucus samples were solubilised in 6 M Urea and subsequently reduced using 50 mM dithiothreitol (DTT) and carboxylmethylated using 125 mM iodoacetamide prior to electrophoresis on a 1% (w/v) agarose gel. Whole mucus samples were separated by 0.7% agarose gel electrophoresis for 22–24 h before reduction in SSC (0.6 M NaCl, 60 mM sodium citrate) containing 0.1 M DTT for 30 min. The fractions were taken from the top of the tubes, analysed by slot blotting or agarose gel electrophoresis followed by western blotting on to a nitrocellulose membrane [15]. Mucins were detected by PAS staining or mucin-specific antisera. Staining intensity was measured using a GS-800 calibrated densitometer (Bio-Rad Laboratories, U.K).
6–8 M guanidinium chloride (GuCl) gradients were formed in centrifuge tubes using an MSE gradient maker connected to a Gilson Minipuls 2 peristaltic pump. Mucin samples (in 4 M GuCl) were loaded onto the tops of the gradients and centrifuged in a Beckman Optima L-90K Ultracentrifuge (Beckman SW40 rotor) at 40,000RPM for 2.75 h for mucus and 3.50 h for purified mucins at 15°C. The refractive index of each fraction was measured using a refractometer; all gradients were comparable (data not shown). Data from 3–5 mice is presented as the percentage of the area under the curve (AUC) of fractions (Fr) 1–8, 9–14 and 15–18, with and without treatment.
Mucins were purified using isopycnic density gradient centrifugation as previously described [16]. In brief, solubilised mucus was purified using caesium chloride (CsCl)/4M GuCl density gradient at a starting density of 1.4 g/ml, centrifuged in a Beckman Optima L-90K Ultracentrifuge (Beckman Ti70 rotor) at 40,000RPM for 65 h at 15°C. Periodic Acid Schiff's (PAS) rich fractions were pooled, dialysed into 0.2 M GuCl before being subjected to a second CsCl density gradient centrifugation (at a starting density of 1.5 g/ml) [16].
The PAS-rich fractions pooled after the second CsCl density gradient was subjected to anion exchange chromatography as described previously using a Resource Q column [17]. Samples were eluted with the starting buffer (20 mM Tris-HCl at pH8) for 15 min (0.5 ml/min), followed by a linear gradient (60 min) up to 0.4 M Lithium perchlorate-10 mM piperazine at pH5 in 6 M Urea containing 0.02% 3-[(chloamidopropyl) dimethylammonio]-1-propanosulphonate [17].
The secreted intestinal mucus barrier was isolated using the method previously described [5]. Samples were digested with trypsin and analysed using a Q-Tof micro mass spectrometer. MS-MS data were subsequently analysed using the Uni-Prot/Swiss-Prot databases [18].
RNA was isolated from caecal epithelial cells as previously described. cDNA was generated using an IMPROM-RT kit (Promega) and SYBR Green PCR MasterMix (ABgene) was used for quantitative PCR using the ABI HT7900 PCR machine. Primer efficiencies were determined using cDNA dilutions and genes of interest were normalised against the housekeeping gene, β-actin, and expressed as a fold difference to uninfected naïve message levels. The following primer sequences were used to determine the levels of Serpinb6a (TGGACAAGATGGACGAAGAAGA, CCAACTTGTAGAGGGCATCGTT); Serpina3k (GAGCAAAGGGCAAGACCA, CAGCCACATCCAGCACAG) and Serpinb1a (TTCGCCTTGGAGCTGTTC, ATGCCTCCTGTAGCAGCG).
Results are expressed as the mean ± SEM. Statistical analysis was performed using Prism v 3.2 (GraphPad Software). The statistical significance of different groups was assessed by using non-parametric tests. P<0.05 was considered to be significant.
ESPs isolated from T. muris are a complex mixture of components (Figure S2A) that have been shown to contain a variety of enzymes [7]. We sought to determine whether the ESPs released from the T. muris nematode have a direct effect on the mucus barrier and its mucin components that are essential for the expulsion of this nematode from the host. To this end, intestinal mucus isolated from uninfected mice was incubated with ESPs for 24 h at 37°C, subsequently, treated and untreated (control) mucus samples were subjected to rate zonal centrifugation. This analysis revealed a significant change in sedimentation behaviour of the mucins after exposure to ESPs and indicated a degradative effect on the mucins that resulted in a reduction in their size; demonstrated by their lower sedimentation rate (Figure 1A–B).
The PAS staining (carbohydrate assay) profile was comparable to the mMuc2 antibody reactivity as expected since Muc2 has been shown to be the predominant intestinal mucin [19]. In addition to the major PAS and mMuc2 staining peak (fractions 9–14), a more minor PAS and mMuc2 staining species had sedimented to the bottom of the gradient (fractions 15–18). These high density mucins most likely represent the ‘insoluble’ Muc2 component of the mucus gel, as characterised previously by Carlstedt et al. [20]. ESPs did not have a degradative effect on Muc2 if heat inactivated prior to incubation or if the incubation was performed at 4°C (Figure 1C), suggesting that enzymes within the ESPs, most likely with proteolytic activity, were responsible for the degradation of mucin polymers. Additionally, the degradative effect was dose-dependent; with a significantly slower sedimentation rate observed with increasing concentration of ESPs (Figure S2B–C).
Mucins present within the mucus barrier are large disulphide bond-mediated polymers that determine the rheological properties of the mucus layer. When polymeric mucins are reduced to monomers, the mucus layer loses its viscous properties, thus highlighting the importance of mucin polymeric structure for the barrier properties of mucus. The significantly slower sedimentation rate of mucins after ESP treatment indicated that the mucins had a decreased size (Figure 1B), which would be predicted to alter the network properties of the mucus barrier and decrease mucus viscosity. To assess alterations in the mucin network, we measured the movement of fluorescently labelled beads within the mucus barrier. This demonstrated that the diffusion rate of the beads was significantly higher when the mucus layer was treated with ESPs (Figure 1D). Furthermore, this alteration in the mucin network was prevented when the mucus layer was concomitantly treated with ESP and protease inhibitors (PI). Collectively, these experiments suggested that ESPs were capable of degrading Muc2, resulting in a more porous mucus network.
Previous studies from our laboratory have shown that the mucin composition of the mucus barrier is altered during T. muris worm expulsion (day 21 pi). Specifically, the mucus barrier is composed of Muc5ac in addition to Muc2 [3]; there are increased amounts of Muc2; and the mucins are differentially glycosylated [5]. Overall, this results in a less porous network [4]. Therefore, we sought to determine whether these changes in the mucus barrier during worm rejection affected the ability of the ESPs to degrade the mucins. To this end, mucus isolated during the course of acute (high dose infection in BALB/c mice) and chronic (low dose infection in BALB/c mice) infection, was treated with ESPs (Figure 2A–D). Untreated and ESP-treated mucus was subjected to rate zonal centrifugation prior to fractionation and analysis.
ESP treatment altered the sedimentation profiles of mucins from mice with chronic and acute infection on day 14 pi. (Figure S3A–D), as previously observed with ESP-treated mucus from uninfected naïve mice (Figure 1A–B). However, ESPs were unable to depolymerise/degrade the mucins present within the mucus isolated from mice with acute infection on day 21 pi. (Figure 2A–B) This correlates with the changes mentioned above that occur within the mucus barrier during the worm expulsion process. Interestingly, the sedimentation profile of mucins in mucus on day 21 from mice with acute infection remained unaltered even when treated with up to 300 µg/ml concentrations of T. muris ESPs (data not shown). In contrast, mucins in the isolated mucus from chronically infected mice on day 21 pi. were readily degraded by 50 µg/ml of T. muris ESPs (Figure 2C). It is important to note that a small proportion of the mucins present within the mucus isolated from the mice with chronic infection on day 21 appeared degraded before in vitro treatment (Figure 2C–D). However, more strikingly, mucins present within the mucus isolated from mice on day 42 pi. with chronic infection were almost completely degraded without ESP treatment (Figure 2E–F) as compared to those isolated from the mice with an acute infection. These data suggest that mucins present within the mucus barrier during long term chronic infection are degraded in vivo and, additionally, may be more prone to degradation by parasite proteases than those produced in mice able to expel the worms.
Surprisingly, mucins present within the mucus isolated on day 56 pi. of acute infection, which is 35 days after the expulsion of the nematode, did not fully degrade when treated with ESPs (Figure S3E–F). This suggested that the changes in the barrier [3], [4] that occur during the co-ordinated expulsion response are maintained for some time after expulsion.
A clear difference was observed in the ESPs ability to alter the mucin component of the mucus barrier produced during worm expulsion (acute infection; day 21 pi) compared to the chronically infected mice (Figure 2). As we have previously reported, in addition to goblet cell hyperplasia and elevated levels of Muc2, IL-13 induced de novo expression of the Muc5ac mucin is also observed during T. muris expulsion [3]. Immunohistochemistry (using Muc2 and Muc5ac-specific antibodies) of caecal tissue and mass spectrometry analysis of caecal mucus confirms the increased levels of Muc2 production in acute infection and demonstrates that even during acute infection Muc2 is the predominant mucin (Figure 3A & B). Therefore, to investigate the effect of the ESPs specifically on mucins in more detail, mucins were purified from the mucus isolated on day 21 pi. of mice with acute infection. In brief, isolated mucus (pooled from 5 mice) was reduced and carboxymethylated and, the resultant mucin monomers were purified from the mucus by two isopycnic density gradient centrifugation steps. Firstly centrifugation in CsCl/4M GuCl was used to separate mucins from lower buoyant density proteins (Figure S4A) [16], followed by a second centrifugation in CsCl/0.2M GuCl to separate mucins from nucleic acids (Figure S4B) and then the distinct Muc2- and Muc5ac-rich fractions were further purified using anion exchange chromatography (Figure S4C).
Fractions enriched in monomeric Muc2 and Muc5ac were pooled from the anion-exchange column (Figure S4C), treated with 50 µg/ml of ESPs for 24 h and, analysed by rate zonal centrifugation (Figure 3C–D). Fractions analysed by slot blotting revealed that ESPs degraded Muc2, since a significant amount of Muc2 was present in the fractions at the top of the gradient after ESP-treatment (Figure 3C). In contrast, the sedimentation profile of the Muc5ac-rich fraction was unaltered after treatment with ESPs with up to 300 µg/ml (Figure 3D). This implied that ESPs specifically degrade the intestinal mucin, Muc2, but are unable to degrade the Muc5ac mucin, which we have previously shown to aid the co-ordinated worm expulsion process by reducing the nematode's vitality [3].
To further explore the effect of ESPs on the MUC2 glycoprotein and in particular its degree of polymerisation, the human intestinal LS174T cell line, which synthesises and secretes mature glycosylated MUC2, was utilised as a source of MUC2. MUC2 isolated was treated with 50 µg/ml of ESPs for 4, 6, 24, 48 and 72 h were subjected to rate zonal centrifugation, which showed a time dependent shift in MUC2 distribution to the top of the gradient and a loss of staining intensity with ESP treatment (Figure 4A). Interestingly, the putative ‘insoluble’ mucin content, present at the bottom of the gradient in the control samples also gradually decreased after ESP-treatment and no ‘insoluble’ fraction was observed after 72 h of treatment (Figure 4A). Untreated and ESP treated MUC2 was subjected to agarose gel electrophoresis (Figure 4B), transferred onto a nitrocellulose membrane and probed with hMUC2 antibody. The unreduced MUC2 can be seen as at least three bands in the control samples that likely represent different multimeric forms of MUC2 (Figure 4B). Importantly, after ESP treatment the intensity of the slower migrating multimeric (i, ii) MUC2 bands were decreased significantly (Figure 4B–C). After 72 h of ESP-treatment the intensity of the fastest migrating MUC2 band (iii) decreased further (Figure 4C). An overall loss of MUC2 antibody staining was also noted, suggesting ESPs initially affect the polymerisation of MUC2 and over time degrade/cleave the MUC2 protein core.
Proteolytic activity of ESPs was probably responsible for degrading/depolymerising the mucin network as it was demonstrated to be temperature dependent (Figure 1C) and blocked by protease inhibitors (Figure 1D). ESPs secreted by the T. muris nematode contain several different proteases including cysteine, serine and metalloproteases [21]. Therefore, to determine which specific protease(s) were responsible for degrading and altering the polymerisation of MUC2, ESPs were incubated in the presence of specific protease inhibitors (Table S1), prior to the treatment of MUC2. Note that the control samples contained a mixture of all the stated protease inhibitors. MUC2 was significantly degraded despite the presence of aprotinin, ethylenediaminetetraacetic acid (EDTA) and N-ethylmaleimidide (NEM) (Figure 4D–E), which inhibit chymotrypsin/trypsin, metallo- and cysteine proteases, respectively. However, it was noted that NEM, may have a partial effect on the ESP enzymatic activity as the fastest migrating band of MUC2 (iii) was intact and a low reactivity with the slower migrating MUC2 band (ii) was also observed (Figure 4D–E). More strikingly, after treatment with the serine protease inhibitors chymostatin and antipain, the ESPs were unable to alter the abundance of all forms of MUC2 (Figure 4D–E). There was a slight difference in the electrophorectic migration of the MUC2 bands noted after ESPs treatment that could not be prevented with protease inhibitor treatment, suggestive of a potential role of ESP in de-glycosylating mucins or activity of another protease in the ESPs. Therefore, overall, these data demonstrate that the depolymerisation activity of ESPs is due to serine proteases.
The serine protease with the enzymatic activity against mucins was fractionated from the ESPs by performing size fractionation into the following fractions <5 kDa, 5–50 kDa, 50–100 kDa and >100 kDa. MUC2 was treated with the ESP-fractions and analysed using rate zonal centrifugation. This demonstrated that 50–100 kDa ESP-fraction had activity against MUC2; whereas, no other ESP-fraction affected MUC2 (Figure 5A).
To ascertain whether the active 50–100 kDa fraction cleaves and depolymerises murine Muc2, the N-terminal D3 polymerisation domain (rMuc2-D3) of Muc2 was expressed in human MKN45 gastric cells which express mucin-specific chaperones and no endogenous MUC2; rMuc2-D3 is present intracellularly as a monomer and dimer and secreted mainly in its dimeric/multimeric forms [13]. The active 50–100 kDa ESP-fraction depolymerised the dimeric form of rMuc2-D3 protein (Figure 5C), similar to that observed when rMuc3-D3 was analysed after treatment with a reducing agent (Figure 5B). rMuc2-D3 isolated from cell lysates and secretions in conditioned media was then treated with the active ESP-fraction (Figure 5D) and analysed by SDS-PAGE/Western blot (detected using the anti-Myc antibody). In addition to the multimeric forms of rMuc2-D3 being reduced to monomers, smaller cleavage products were also observed (highlighted with arrows; Figure 5D), confirming that ESPs not only depolymerised the murine Muc2 mucin but also degraded/cleaved it into smaller fragments. Importantly, the other ESP-fractions did not have this effect on rMuc2-D3 protein (Figure 5C), and treatment with the serine protease inhibitor antipain inhibited this activity (data not shown), confirming that serine protease(s) present in the 50–100 kDa fraction were responsible for depolymerising/degrading Muc2.
Worm expulsion is near complete by day 21 pi. [6] in the mice with the high dose infection (Figure S1), and our data clearly show that at this stage the mucins, and in particular Muc2, present within the mucus barrier were somehow protected from the degradative effects exerted by ESPs. This raised the possibility of the presence of protective non-mucin components such as protease inhibitors which may be secreted by the host within the mucus barrier to hinder the ESP activity. To this end, we isolated the secreted mucus from mice during the course of acute and chronic infection using a method previously described [5]. The proteins present within mucus were identified using trypsin digestion followed by tandem mass spectrometry (MS) analysis. Tandem MS analysis identified three members of the serine protease inhibitor family (Serpins), Serpinb6a, Serpina3k and Serpinb1a, to be present in the mucus barrier during infection. Interestingly, Serpins were not detected in the uninfected and chronically infected mice on day 14 pi., with 1–2 unique peptides identified on day 21 pi. of chronic infection (Figure 6A, B). In contrast, the mucus barrier isolated during acute infection, contained higher levels of serpins on day 14 and day 21 pi. RT-PCR analysis demonstrated that mRNA levels encoding Serpinb1a, Serpinb6a and Serpina3k (Figure 6B), identified in the mucus barrier with MS analysis, were elevated in the caecal epithelium of mice during the TH2-mediated immune response; however no major changes were observed during chronic infection. The host therefore potentially up-regulates serine protease inhibitors during the worm expulsion process in order to prevent mucin degradation in vivo, and subsequent inflammation and aid the rejection of the nematode from the intestine.
Mucins are an essential part of the TH2-mediated immune response that enforces intestinal expulsion of the nematode, Trichuris muris [1], [3], [4]. In this study, we demonstrate how the Trichuris nematode has evolved ways to promote its survival within its intestinal niche by degrading mucins; the molecular framework of the host protective mucus barrier. This is the first study to describe that serine protease(s) released by the nematode can cleave the polymerising domain of Muc2, the major intestinal mucin, leading to a more porous mucus barrier network. Here, we clearly demonstrate that the Muc2 mucin, but not Muc5ac, was degraded by secretions from T. muris nematodes; and serine protease inhibitors are present in the mucus barrier during worm expulsion, which may hinder the degradative effects of ESPs. The degradation of Muc2, but not Muc5ac, by parasite proteases provides a plausible explanation for our previous observation that, whilst Muc2 aids worm expulsion [4] only Muc5ac is essential for expulsion [3]. Therefore, components of the ESPs are produced by the nematode to alter the properties of the mucus barrier and thus facilitate its own survival and/or improve the conditions within its niche.
The Trichuris nematodes are extremely successful within the host, because they not only have the ability to survive an immune attack but are thought to actively subvert the immune responses generated by the host by exerting immunomodulatory effects [6]. These nematodes have been shown to produce complex secretions containing proteases and proteins that have immunogenic properties, thus described as ‘excretory secretory antigen’, but which are likely to have a protective function [7], in particular in maintaining infection within its mucosal niche. Mucins are responsible for the physical properties of the mucus barrier. These large heavily glycosylated molecules polymerise, mediated by disulphide bond formation involving the cysteine-rich domains at the N- and C-termini of the mucin polypeptide, which gives the barrier its viscous properties [22]. The parasite proteases have evolved the ability to act on the polymerising D3 domain of the intestinal mucin, Muc2, which is thought to be largely protease resistant [23], due to the high level of intra-molecular disulphide bonding. The dimeric form of a recombinant mucin protein was depolymerised into monomers after ESP-treatment and this was prevented if serine proteases inhibitors were used and a smaller fragment than the monomeric form was also observed. Clearly, the depolymerisation/degradation of mucins impacts on the properties of the mucus barrier, as the barrier was more porous after ESP treatment, which also corroborates our previous finding of a more porous mucin network in mice with chronic infection [4].
The host has developed ways to counteract the depolymerising ability of ESPs. We observed a significant difference in the ESPs ability to degrade mucins isolated during acute and chronic infection. As described previously, under the TH2-mediated immune response the mucin component of the mucus barrier changes during worm expulsion with a prominent increase in Muc2 [4], de novo expression of the mucin Muc5ac [3] and change in glycosylation observed [5]. ESPs were unable to degrade the mucins in mucus isolations on day 21 of acute infection, which is the peak of the immune response [6]. Interestingly, when the mucins were purified we could show that the ESPs have the ability to degrade mouse and human Muc2/MUC2, but not the Muc5ac mucin, which is not usually expressed in the intestinal epithelium. It is plausible that these nematodes have evolved a capability to degrade the major mucin in the intestine, Muc2, whilst being unable to degrade Muc5ac. This suggests that the Trichuris proteases are specifically acting on a peptide sequence within the MUC2/Muc2 protein core that appears conserved between mouse and human. The presence of the IL-13 induced protease-resistant Muc5ac during nematode expulsion will maintain mucus viscosity and retention of anti-helminthic factors in the mucus and subsequently be detrimental for nematode viability as previously demonstrated [3].
Interestingly, in addition to Muc5ac and increased levels of Muc2, we show that Serpins were upregulated and were present within the mucus barrier of mice resistant to chronic infection. The presence of Serpins, and possibly other protease inhibitors in the mucus barrier during worm expulsion could explain why mucins, in particular Muc2, were protected from degradation when treated with ESPs but Muc2 when purified, was susceptible to degradation. Another possibility for the lack of degradation of mucus could be due to the increased concentration of mucins and other proteins within the mucus barrier [4]. The increased levels of proteins could result in more competition for ESPs to cleave sites and, therefore, make the mucins less susceptible to degradation as illustrated previously in respiratory mucus [24]. The changes to the properties and composition of the mucus barrier could hinder ESPs activity, which could also explain the decrease in vitality of the nematode during worm expulsion [3], [4].
This is not the first time pathogen exo-products have been shown to degrade mucins, several other pathogens have adopted a similar mechanism to survive within the mucosal layer. Helicobacter pylori secretes ‘mucinases’ which allow its corkscrew-like motion through the mucus layer [25]. Protozoan parasites such as Entamoeba histolytica [26], Trichomonas vaginalis [27] and Naegleria fowleri [28] all release cysteine proteases which have the ability to degrade mucins. We demonstrated that cysteine protease inhibitors (NEM and aprotinin) only very partially limited the activity of Trichuris ESPs to degrade MUC2. Treatment with chymostatin and antipain inhibited the depolymerisation of MUC2 implicating trypsin and/or serine protease activity. However, since degradation was not inhibited by aprotinin (cysteine and trypsin protease inhibitor), it is most likely that serine protease activity is responsible for degrading Muc2/MUC2. ESPs contained serine protease(s) of molecular weight range 50–100 kDa with the depolymerising activity against the gel-forming mucins. Interestingly, serine proteases of molecular weight of 85 and 105 kDa have been reported to be isolated from T. muris ESPs previously [21]. Whilst serine proteases appeared to be the major Muc2 protease, our data imply that cysteine proteases present in the ESPs were in part responsible for the affects observed on the insoluble Muc2/MUC2-gel. Serine and cysteine proteases, therefore, may act in concert to disrupt the polymeric mucin network. Supporting, this hypothesis further is the presence of Serpins within the mucus barrier prior to and during worm expulsion, which may hinder the ability of ESPs to break down the mucus barrier.
There is an added level of unique complexity in the assembly of the intestinal MUC2: an uncharacterised ‘non-reducible linkage’ which results in an ‘insoluble’ gel enabling MUC2 to form a barrier resistant to the harsh environment of the intestine [20]. In addition, other proteins such as Fc Ig binding protein (Fcgbp) have been shown to associate with Muc2/MUC2 and could potentially act as a cross-linkers [19]. Interestingly, for the first time we demonstrate that Trichuris ESPs can degrade the Muc2/MUC2 polymers into smaller subunits and may have a further effect on the MUC2 protein as there was a change in the electrophorectic migration suggesting ESPs may be involved in de-glycosylating/degrading the MUC2 protein. Interestingly, as the mucins produced in vivo during acute infection are protected, the mucins present in the mucus barrier of susceptible mice also have reduced glycosylation [5], which may make them more prone to the effects of ESPs. Taken together, the data suggests that the serine protease activity of the ESPs cleaves the mucin polymerising domain (Figure 7) resulting in mucin monomers [29] which can be cleaved/degraded into smaller fragments (Figure 7). This will subsequently result in a mucus barrier that is more porous and, will therefore, exacerbate inflammation and aid persistence of infection.
Although the stability and turnover of the ESPs is not known in vivo, ESPs can clearly increase the porosity of the mucus layer by depolymerisation of mucins, which would hinder the retention of host protective factors within the mucus barrier. Furthermore, bearing in mind the niche in which Trichuris lives there will be a major interface between the adult parasite and secreted mucins within the mucus layer via the posterior half of the worm, which protrudes out of the epithelium into the caecal lumen. The posterior section of adult worms is involved in mating and ultimately egg deposition, and it is tempting to speculate that modification of the mucus barrier properties, perhaps via the proteolytic activity described here would and, allow optimal mobility of the posterior end of the worm, facilitating efficient mating and egg laying during chronic infection. Many questions remain unanswered including identification of the specific protease(s) and the cleavage site on the mucin, the site of protease(s) production and the details of the host anti-protease response. Answering these will deepen our understanding of the host-parasite relationship of this group of ubiquitous and important gastrointestinal dwelling nematodes.
|
10.1371/journal.pbio.1001300 | Social Transfer of Pathogenic Fungus Promotes Active Immunisation in Ant Colonies | Due to the omnipresent risk of epidemics, insect societies have evolved sophisticated disease defences at the individual and colony level. An intriguing yet little understood phenomenon is that social contact to pathogen-exposed individuals reduces susceptibility of previously naive nestmates to this pathogen. We tested whether such social immunisation in Lasius ants against the entomopathogenic fungus Metarhizium anisopliae is based on active upregulation of the immune system of nestmates following contact to an infectious individual or passive protection via transfer of immune effectors among group members—that is, active versus passive immunisation. We found no evidence for involvement of passive immunisation via transfer of antimicrobials among colony members. Instead, intensive allogrooming behaviour between naive and pathogen-exposed ants before fungal conidia firmly attached to their cuticle suggested passage of the pathogen from the exposed individuals to their nestmates. By tracing fluorescence-labelled conidia we indeed detected frequent pathogen transfer to the nestmates, where they caused low-level infections as revealed by growth of small numbers of fungal colony forming units from their dissected body content. These infections rarely led to death, but instead promoted an enhanced ability to inhibit fungal growth and an active upregulation of immune genes involved in antifungal defences (defensin and prophenoloxidase, PPO). Contrarily, there was no upregulation of the gene cathepsin L, which is associated with antibacterial and antiviral defences, and we found no increased antibacterial activity of nestmates of fungus-exposed ants. This indicates that social immunisation after fungal exposure is specific, similar to recent findings for individual-level immune priming in invertebrates. Epidemiological modeling further suggests that active social immunisation is adaptive, as it leads to faster elimination of the disease and lower death rates than passive immunisation. Interestingly, humans have also utilised the protective effect of low-level infections to fight smallpox by intentional transfer of low pathogen doses (“variolation” or “inoculation”).
| Close social contact facilitates pathogen transmission in societies, often causing epidemics. In contrast to this, we show that limited transmission of a fungal pathogen in ant colonies can be beneficial for the host, because it promotes “social immunisation” of healthy group members. We found that ants exposed to the fungus are heavily groomed by their healthy nestmates. Grooming removes a significant number of fungal conidiospores from the body surface of exposed ants and reduces their risk of falling sick. At the same time, previously healthy nestmates are themselves exposed to a small number of conidiospores, triggering low-level infections. These micro-infections are not deadly, but result in upregulated expression of a specific set of immune genes and pathogen-specific protective immune stimulation. Pathogen transfer by social interactions is therefore the underlying mechanism of social immunisation against fungal infections in ant societies. There is a similarity between such natural social immunisation and human efforts to induce immunity against deadly diseases, such as smallpox. Before vaccination with dead or attenuated strains was invented, immunity in human societies was induced by actively transferring low-level infections (“variolation”), just like in ants.
| The first encounter of a host with a particular pathogen often leads to the outbreak of the disease, yet a secondary exposure rarely causes illness, due to the immunological memory of the host. Whereas immune memory in vertebrates is well appreciated [1], the phenomenon of an individual developing specific immunity against a subsequent pathogen exposure—referred to as immune priming—has only recently been described in invertebrates, both within the lifetime of an individual [2]–[8] and in transgenerational protection of offspring ([8]–[12], but see [13]). In contrast to vertebrates, the underlying mechanisms are not yet understood in invertebrates [14],[15]. In addition to this immunological memory at the level of individuals, a similar phenomenon occurs at the colony level in insect societies [16]–[18]. Society members act collectively, similar to cells in a body, and work as a superorganism [19],[20] in multiple aspects, including anti-pathogen defence [21]. For instance, an initial pathogen contact of a colony due to the presence of exposed individuals has been shown to lower the susceptibility of their nestmates to infection when they are later exposed to the same pathogen [16]–[18]. In addition to this physiological “social immunisation,” the collectively performed hygiene behaviour that complements individual defences in social insects [22]–[24] is also affected. Allogrooming of exposed individuals by their nestmates occurs more frequently in colonies with previous experience with this pathogen than in naive colonies [25],[26]. In contrast to individual immune priming, social immunisation thus refers to a protection of naive individuals of a colony after social contact to exposed individuals.
The phenomenon of social immunisation occurs broadly in insect societies—in unrelated social host species (ants and termites) and against divergent pathogen taxa (fungi [17],[18] and bacteria [16])—yet the mechanisms underlying this effect are largely elusive (but see [16]) and have only been hypothesised upon for fungal pathogens [3],[17],[18],[27]. In this study, we therefore aimed to determine the underlying causes of social immunisation in colonies of the ant Lasius neglectus after exposure of single individuals to the entomopathogenic fungus Metarhizium anisopliae, a common natural pathogen of ants [28],[29]. In this system, we have previously described that 5 d of social contact to an individual exposed to fungal conidia (conidiospores; [30]) led to a lower susceptibility of nestmate ants when challenged with a high fungal dose after this period [18]. It remained open, however, which social interactions may trigger this effect and how they elicit changes in nestmate immunity.
The observed protection in nestmates of exposed ants may be caused by the active upregulation of their own immune systems following social contact to the fungus-exposed individual. Alternatively, social transfer of immune mediators produced by colony members may lead to passive protection of nestmates without requiring the activation of their own immune systems (as outlined by [3],[17],[27]). The active and passive route to social immunisation may also act in concert.
Active upregulation of the nestmates' immune system may be caused by perception of a trigger signal elicited from the exposed individual, possibly of behavioural or chemical nature. In humans, mere visual perception of sick individuals was recently shown to cause preventive stimulation of the immune system [31]. Similarly, in plants, herbivory defence was promoted by perception of volatile chemical cues elicited by an attacked neighbouring plant [32]. Active stimulation of the immune system can also be caused by low-level infections [3],[8],[33],[34], which may result from social transfer of the pathogen from the exposed individual to its nestmates (as suggested by [3]), occurring during “normal” social interactions, or as a byproduct of collective sanitary behaviour such as allogrooming of the exposed individual by its nestmates [22],[35].
Passive immunisation may result from a social exchange of antimicrobials produced by the exposed individuals and transferred to their nestmates. Possible transfer pathways include the “external route” over the body surface or the “internal route” by exchange of body fluids [16]. The external body surface (cuticle) of ants is covered with antimicrobial substances produced in an ant-specific gland (metapleural gland [36],[37]) and nestmates could easily pick up these substances and apply them on their own bodies by allo- and self-grooming. Immune effectors produced inside the body of infected individuals may be exchanged during the common social feeding behaviour of regurgitation and feeding of trophallactic droplets [16],[38], as has recently been suggested as a mechanism for social immunisation of ant colonies after bacterial exposure [16]. Whereas bacterial infections are typically orally transmitted [39], entomopathogenic fungi are externally transmitted, making distinct disease dynamics of these pathogen taxa likely.
In this study, we applied a multi-level approach to determine the functional mechanism of social immunisation of ant colonies against a fungal pathogen. We analysed the behavioural interaction rates between group members and determined whether social contact may lead to exchange of the pathogen or immune effectors, or whether social immunisation may be triggered by social signals. We determined both the physiological immunity of fungus-exposed individuals and their nestmates, as well as their immune gene expression. Lastly, we developed an epidemiological model to explore long-term colony-level effects of social immunisation depending on the underlying mechanisms.
We have previously shown that social contact to a Lasius worker exposed to conidia (dispersal form, conidiospores; [30]) of the entomopathogenic fungus M. anisopliae, but not to control-treated ants, increased the survival of previously naive nestmates when challenged with the same M. anisopliae strain 5 d later [18]. We now directly assessed the immune function of nestmates with a novel and sensitive “antifungal activity assay.” We incubated ant tissue with blastospores (within-host infection form; [30]) of the fungus to measure the ability of ants to inhibit fungal growth. We found a significantly higher antifungal activity in nestmates of fungus-exposed as compared to nestmates of control-treated individuals (Figure 1). This was true not only after 5 d of social contact to an exposed individual, but already after 3 d (GLM, F = 3.859, df = 3, p = 0.017; treatment type [fungus treatment versus sham control]: F = 10.634, df = 1, p = 0.002; time [3 versus 5 d post-treatment]: F = 0.001, df = 1, p = 0.973; interaction [Treatment Type×Time]: F = 0.942, df = 1, p = 0.338).
To understand the mechanism behind increased antifungal defence in nestmates of exposed ants, it is important to study the behaviour of group members. First, behavioural changes of individuals after fungal exposure may be a signal to their nestmates to upregulate their immune system. Second, the social interactions define the routes and opportunities for potential exchange of immune effectors [40],[41] or the pathogen itself [42].
Compared to control-treated ants, which did not elicit social immunisation in their nestmates, fungus-exposed ants did not show significantly changed rates of either brood care behaviour [18] or self-grooming activity (LVU, unpublished data). Similarly, other studies found that pathogen exposure had no effect on self-grooming [26] or only when doses present in the colony were very high [25]. This makes it unlikely that nestmates may have perceived a trigger signal by social interaction or potential observation of the individual behaviour of exposed ants.
To obtain information on possible pathways for transfer of the pathogen or immune mediators, we analysed the social interactions between colony members in more detail. As in our original experimental setup we grouped five naive nestmates with a single treated Lasius worker that had either received infectious M. anisopliae conidia (fungus treatment) or the same treatment without the pathogen (sham control). We observed three types of social interactions between group members. Antennation behaviour—that is, nestmate recognition behaviour by antennal contact [43]—occurred extremely rarely (6.6% of all interactions). Moreover, rates did not differ between treated and nestmate ants or among nestmates, for both fungus treatment and sham control (Generalised Linear Model [GLM] with negative binomial errors, LR χ2 = 1.969, df = 3, p = 0.579; data not shown). All other social interactions observed between group members consisted of (a) allogrooming (i.e., cleaning the body surface of another ant) and (b) trophallaxis behaviour (i.e., exchange of regurgitated liquid food droplets) [43]. Both may be important pathways for social immunisation [3],[16],[17],[27].
It is well known that nestmates actively contact exposed individuals and remove infectious material with their mouth by allogrooming, which is a very efficient social sanitary behaviour [43],[44] increasing survival of pathogen-exposed individuals, but typically not compromising the survival of the nestmates [25],[35],[45],[46]. Still, the grooming ant may contract the pathogen if it is not able to kill all infectious material in its mouth (infrabuccal pockets; [47],[48]) or gut [49], or if it unintentionally rubs off conidia with other body parts than the mouth during this intimate social interaction. In addition, allogrooming may lead to uptake of antimicrobial substances from the body surface of an exposed individual similar to exchanges of cuticular waxes important for nestmate recognition [50].
In our experiment, allogrooming rates between treated individuals and their nestmates were higher than among nestmates, but independent of the treatment type (fungus versus sham control; Figure 2A; GLM with negative binomial errors, LR χ2 = 15.134, df = 3, p = 0.002; ant pairing [treated-nestmate versus nestmate-nestmate]: Wald χ2 = 14.501, df = 1, p<0.001; treatment type [fungus versus sham control]: Wald χ2 = 0.006, df = 1, p = 0.939). Upregulation of grooming frequency not only against individuals treated with infectious material but also with sham control solutions is known from previous studies [29],[51] and indicates that ants are very sensitive to applications on the bodies of their group members.
Despite the lack of difference between the two treatment types, intensive grooming towards treated individuals provides a potential route for transfer of either the pathogen itself or external immune effectors. One important factor is the timing of allogrooming expression during the infection course of M. anisopliae. Entomopathogenic fungi like M. anisopliae infect their hosts by external adhesion onto and active penetration of the cuticle [52]. After contact to the insect cuticle, the conidia first adhere loosely to the body surface within several hours and then germinate and form a penetration plug to actively enter the host body within approximately 24 to 48 h [46],[53]. Infection of the host and onset of an active immune response therefore occurs with a time delay of 2 to 3 d after exposure [54],[55]. Allogrooming in the first 1 to 2 d would therefore allow for pathogen transfer, whereas after this time exposed ants lose their infectiousness [26]. Intensified allogrooming 3 or 4 d after exposure would instead indicate exchange of external antimicrobial substances.
We analysed the time course of allogrooming frequency between treated individuals and their nestmates and found no change over time in the control treatment (GLM with repeated measures, time: F = 0.973, dfHuynh-Feldt = 3.648, p = 0.416). Allogrooming between nestmates and fungus-exposed individuals, however, was significantly higher in the first 2 d compared to later phases of the experiment (Figure 2B; time: F = 4.006, dfHuynh-Feldt = 3.306, p = 0.006 [day1 versus day2: p = 0.178; day1 versus day3: p = 0.041; day1 versus day4: p = 0.001; day1 versus day5: p = 0.014]). Based on these data we suggest that if a transfer between group members occurs via allogrooming, it more likely involves a transfer of conidia, detachable early after exposure, than immune effectors, which can only be upregulated and transferred to the cuticle after infection of the individual 24–48 h after exposure.
Social feeding via regurgitation and transfer of a trophallactic droplet may promote transfer of internal antimicrobial substances [16]. However, we found no differences in the rates of trophallaxis among all four groups, that is, neither between treated ants and their nestmates nor among the nestmates in either the fungus treatment or the control group (Figure 2C; GLM with negative binomial errors, LR χ2 = 2.555, df = 3, p = 0.465). Our data show that fungal exposure does not alter trophallaxis rates between exposed individuals and their nestmates, making passive immunisation by transfer of internally produced antimicrobial substances rather unlikely in our model system. Our findings after fungal exposure contrast with observations that trophallaxis rates between individuals injected with dead bacteria or bacterial cell wall components (but also wounding controls) were increased compared to trophallaxis rates among untreated individuals ([16],[56], but see [57]).
Taken together, our behavioural observations strongly suggest exchange of the fungal pathogen between the fungus-exposed ant and its nestmates during intensified, early grooming as the most likely mechanism for the observed anti-fungal protection in the nestmates. We therefore determined if fungal conidia indeed were transferred from the exposed individual to its untreated nestmates by direct tracing of fluorescently labelled conidia.
We applied conidia of M. anisopliae labelled with red fluorescent protein (RFP) onto the exposed ant and determined their presence or absence on the cuticle of all group members after 2 d of social contact. We expected maximum pathogen transfer to have occurred at this time as (a) grooming activity between exposed ants and their nestmates is most intense in the first 30 h (Figure 2B) and (b) conidia are no longer transferable after this time [26],[53].
As expected we found high amounts of conidia on all directly exposed individuals (15/15) and furthermore detected low numbers of conidia on the cuticles of 37% (17/45) of nestmates (Figure S1; for negative controls see Materials and Methods). Interestingly, not only the quantity but also the location of conidia differed: whereas directly exposed individuals carried them mostly in areas likely difficult to reach by grooming such as joints and the antennal grooves, conidia on nestmates were rather attached to antennae and legs (Figure S1), suggesting that nestmates pick up the pathogen from the fungus-exposed individual during grooming. We can thus confirm pathogen transfer to the nestmates. In a next step we determined if the transferred conidia successfully established an infection in the nestmates.
To quantitatively determine the infection load of directly fungus-exposed individuals and their nestmates over the course of the experiment, we sterilised their body surface to destroy all remaining conidia, dissected the ants, and plated their body contents on agar plates to count emerging fungal colony forming units (CFUs). We used morphological determination, as well as PCR [58], to confirm that outgrowing CFUs were indeed M. anisopliae, which was the case for all CFUs (see Figure S2 as an example). None of the 30 negative controls (see Materials and Methods) and none of the individuals measured within 24 h after exposure (0/10 fungus-treated, 0/14 nestmates; Figure S3) showed fungal growth, confirming that we effectively sterilised the ants and measured only live fungus from inside the body.
Three as well as five days after exposure, CFUs grew from the body content of nearly all directly exposed ants (80% [8/10] and 90% [9/10]) and a similarly high number of nestmates (64% and 64% [each 9/14]; Figures 3, S3; Fisher's exact test; day 3, p = 0.653; day 5, p = 0.341). These data show that fungal infections in nestmates were more common than estimated from external pathogen transfer using labelled conidia. This may either indicate that we did not detect all conidia or that an additional infection route via the infrabuccal pocket in the mouth or the gut system occurred, for instance if groomed-off conidia were not completely prevented from germinating [47]–[49]. Fungal infection load in nestmates revealed that their infections were “low-level infections.” The number of CFUs growing out of their bodies when infected was significantly lower than those growing from directly exposed ants at both day 3 (Figures 3A, S3; Mann-Whitney U-test: n1 = 8, n2 = 9, U = 4.0, p = 0.002) and day 5 (Figures 3B, S3: n1 = 9, n2 = 9, U = 0.0, p<0.001). On average, the infection load of infected nestmates was 8 (4.4 versus 36.0) and 12 (8.1 versus 102.4) times smaller than that of directly exposed individuals on days 3 or 5, respectively.
Even if low-level infections occurred in the majority of nestmates, only 2% (3/150) died from a M. anisopliae infection after 5 d of social contact with the exposed individuals (who showed death rates of approximately 50% due to application of an LD50). This confirms that the effects of M. anisopliae infections are highly dosage dependent ([35] and MKo and STr, unpublished data).
To determine if the observed increase in antifungal activity of nestmates was a direct cause of these low-level infections, we established low-level infections in individuals in the absence of social interactions. To this end, we exposed isolated ants with a conidia dose that led to the same death rate (LD2) and infection level as observed in the socially exposed nestmates. We found that low-dose, directly exposed ants had a significantly increased antifungal activity 3 d after exposure compared to control-treated ants (Figure 4). Interestingly, directly exposed individuals with a high dose (LD50; as used for exposure of the single ants in our experiment above) showed a significantly decreased capacity to inhibit fungal growth (Figure 4; ANOVA: F = 10.361, df = 2, p<0.001; post hoc Protected Fisher's LSD tests all pairwise: sham control versus LD2: p = 0.046, sham control versus LD50: p = 0.021; LD2 versus LD50: p<0.001). This immune-suppressive effect of a high-dose infection is likely caused by the immune-interference and toxicity of M. anisopliae or by the fact that the immune responses had been depleted [41],[59]–[61]. Immune stimulation of low-level infections has previously been described for both vertebrates and invertebrates [3],[8],[33],[34], and its protective effect yielded clinical application in humans [62],[63] and poultry health management [64].
We have established that low-level infections, caused by social contact or direct low-dose exposure, lead to increased antifungal activity. Yet this does not exclude that nestmates with social contact to an exposed individual may also obtain signals that could actively trigger their antifungal immunity (similar to [31],[32]). To test this, we performed a “spatial-separation experiment” in which body contact and pathogen transfer to the exposed individual were prevented, whereas exchange of visual signals or volatile chemicals was still possible. The antifungal activity of nestmates of fungus-exposed individuals did not differ from that of nestmates of control-treated ants after 3 d of this constrained contact (t test: t = −0.376, df = 18, p = 0.711). These data suggest that a visual or volatile signal alone—at least one that acts over distance—is not sufficient to promote antifungal activity in the nestmates. Non-volatile chemical signals, such as cuticular hydrocarbons [65] that are part of the ants' cuticle, may in theory still play an additional role. However, their perception would always require body contact, which promotes pathogen transfer at the same time. We conclude that low-level infections alone provide a sufficient explanation for an active social immunisation of nestmates. We then tested if it may be complemented by a passive transfer of antimicrobial substances among nestmates.
We performed a “temporal-separation experiment” and allowed the exposed ant to interact with its nestmates for 48 h. In this period, the pathogen (a) lost its ability to be transferred (for confirmation see Materials and Methods) and (b) established an infection in the ants, likely triggering an immune response [53]–[55]. After this time, we separated the treated individual and its “early nestmates” and added five “new nestmates” to both (see Figure 5A,B). Three days later, we measured the antifungal activity of the new nestmates. We found no difference between new nestmates of control-treated versus fungus-exposed ants (Figure 5A; t test: t = −0.159, df = 18, p = 0.876) or between new nestmates of early nestmates to a control-treated versus exposed individual (Figure 5B; t test: t = −1.273, df = 18, p = 0.219). This reveals that nestmates do not show an increase in antifungal activity if pathogen transfer is excluded.
Passive transfer of antimicrobials among the group members thus seems very unlikely as an explanation for social immunisation. However, such transferable substances might be upregulated in infected individuals and simply failed to elicit immunisation of nestmates in our experiment. We therefore also analysed both the fungus-exposed ant and its nestmates directly for the presence of potentially transferable antimicrobials 3 d after treatment. Although allogrooming rates among nestmates were low in both sham control and fungus-treated groups (Figure 2A), and trophallaxis rates were completely independent of treatment (Figure 2C), infected nestmates may be important in transferring antimicrobial substances, as their antifungal activity is higher than that of directly exposed ants, which suffer a much higher infection level (Figure 4).
We tested whether transferable substances of fungus-exposed individuals or their nestmates had higher antifungal activity than those of control-treated individuals and their respective nestmates. For externally transferable substances via allogrooming, we measured the antifungal activity of (a) the cuticle and (b) the thorax containing the metapleural gland content, which is known to have antimicrobial function and to be secreted onto the cuticle [36]. We also measured the antifungal activity of (c) the trophallactic droplet that is produced in the ant's body and is transferred via social feeding. We found that neither the cuticles nor the thoraxes containing the metapleural gland nor the trophallactic droplets of fungus-exposed individuals showed a different antifungal activity than the respective body parts of control-treated individuals (Figure 5C; t tests; cuticle: t = 1.064, df = 10, p = 0.312; thorax: t = 0.224, df = 10, p = 0.828; trophallactic droplets: t = −0.594, df = 18, p = 0.560). The same was true for the nestmates (Figure 5D; t tests; cuticle: t = 0.107, df = 18, p = 0.916; thorax: t = 0.894, df = 18, p = 0.383; trophallactic droplets: t = −0.717, df = 18, p = 0.482). This result was not an artifact caused by a potential effect of the control treatment, as the antifungal activity in these individuals was not different from completely untreated ants (Materials and Methods).
Taken together, we found no evidence for (a) a potential protective effect of nestmates in the absence of pathogen transfer and (b) potential upregulation of socially transferable antimicrobials in exposed colonies. This contrasts observations that trophallactic droplets obtained from bacteria-exposed ants had higher antibacterial activity than that of controls [16], making passive immunisation a likely mechanism involved in social immunisation of ant colonies after bacterial exposure [16], but not after fungal exposure. Instead, we documented that social interaction, most likely allogrooming, leads to pathogen transfer and sublethal low-level infections in the majority of nestmates of fungus-exposed individuals and that low-level infections are necessary and sufficient to induce an increased antifungal activity.
To directly assess the effect of low-level infections on the immune response, we measured immune gene expression in nestmates using quantitative real-time PCR. We chose three immune genes known to be involved in the humoral and cellular defences of ants: (1) the antimicrobial peptide (AMP) defensin [66],[67], a soluble mediator that most closely resembles termicin, an antifungal peptide in termites [68],[69]; (2) prophenoloxidase (PPO), a key mediator of immune function in ants [70],[71] that is essential for the process of melanization upon infection by a variety of pathogens, including entomopathogenic fungi [72],[73]; and (3) cathepsin L, a lysosomal protease expressed in hemocytes [74], which has both antibacterial [75] and antiviral activity [76], but has not been implicated in antifungal responses. In Camponotus pennsylvanicus, another cathepsin (cathepsin D) was found to occur in higher amounts in the trophallactic droplets of ants after injection of heat-killed bacteria or LPS [16], suggesting the involvement of cathepsins in antibacterial responses in ants. We confirmed that our host ant, L. neglectus, also responds to bacterial infection with cathepsin upregulation. Septic injury with Bacillus thuringiensis led to upregulation of cathepsin L gene expression, but not PPO, or defensin expression, compared to pricked controls (Figure S4; defensin: t test; t = 0.186, df = 4, p = 0.862; PPO: t test; t = −1.448, df = 4, p = 0.221; cathepsin L: t test; t = −3.695, df = 4, p = 0.021; gene expression standardised to the housekeeping gene 18s rRNA). The choice of these three immune genes in this study therefore allowed us to examine the specific effects of social immunisation against the fungus M. anisopliae on immune pathways involved in insect defences.
We compared mRNA levels of the three genes in nestmates of fungus-exposed individuals versus nestmates of control-treated individuals on day 3—that is, the first day that we observed an increase in their antifungal activity (Figure 1). After normalising to a housekeeping gene (18s rRNA), elevated expression was observed in nestmates of fungus-exposed individuals relative to nestmates of control-treated individuals for both defensin and PPO (Figure 6; defensin: Welch's t test; Welch t = −2.348, df = 26, p = 0.032; PPO: t test; t = −2.923, df = 26, p = 0.007), whereas cathepsin L showed no difference (t test; t = −0.094, df = 26, p = 0.926). This reveals an active upregulation of immune gene expression in nestmates of fungus-exposed ants and suggests the induction of a specific immune response distinct from immune responses to bacteria (Figure S4; [16]). Similar specific immune upregulation after fungal infection is known to occur in Drosophila [77].
To determine if the observed specificity in our candidate gene approach, which is limited to a small set of genes, reflects specificity at the functional level, we tested the nestmates' capacity to inhibit growth of the bacterium Arthrobacter globiformis in an “antibacterial activity assay.” We found that nestmates exhibited similar antibacterial activity for fungus and control treatment (Figure 7; t test: t = −0.644, df = 18, p = 0.528), revealing that social immunisation after fungal exposure of the colony is specific and does not lead to a protective effect against bacteria.
We developed an epidemiological model to explore the adaptive value and colony-level long-term effects of social immunisation. We compared the effect of active versus passive immunisation in our ant-fungus system by extending classical SIS and SIR (Susceptible-Infectious-Recovered/Removed) models, which describe the progress of epidemics over time using the simplification that the diversity in the population can be reduced to a few states. Possible states in SIR models include individuals susceptible to the disease outbreak (S), infectious individuals (I), and recovered or dead individuals (R; [78],[79]). We included an active or passive immunisation mechanism by constructing a SIRM (Susceptible-Infectious-Removed-iMmune) model, in which ants can take five different states. Healthy nestmates are defined as susceptible (S) individuals, pathogen-exposed individuals as infectious (I) ones, and individuals dying from the disease are removed (R) from the model. Successful immunisation (by active or passive immunisation) leads to initially immune (Mi) individuals that may persist to create late-stage immune individuals (Ml; Figure 8). We describe the mean number of ants in each state by ordinary differential equations (ODEs; for details, see Text S2). We have thereby chosen a simple approach focusing on the comparison of active versus passive immunisation, but not taking into account spatial effects on epidemiology in societies that have been modelled elsewhere by cellular automata [27],[80],[81] or pair-wise approximations models [82].
Ants can change their state by social interactions with each other and depending on their infection state (Figure 8A,B). Allogrooming reduces the fungus load of infectious (I), changing them to susceptible (S), but at the same time can increase the fungus load of the susceptible individuals (S), changing them to infectious (I). Active immunisation can occur when individuals receive a low-level infection and actively build up immunity, changing from infectious (I) to immune (Mi) with a given active immunisation rate. Under passive immunisation, susceptible (S) individuals change directly to the immune state (Mi) with a passive immunisation rate when receiving antimicrobial substances from infectious (I) individuals. Under the active immunisation scenario, initially immune ants (Mi) may then either die (R) if infection levels are too high and lead to the disease or enter into the later stage of immunity (Ml). Under passive immunisation, all initially immunised individuals become late-stage immune. Late-stage immune ants (Ml) can then lose their immunisation and become susceptible individuals (S; see Figure 8A,B and Text S2). Each transition is governed by a transition rate, which in total were fixed to similar ranges in order to allow easy model comparison. The following qualitative results did not depend on the precise rate values, so that we report only representative outcomes of our simulations in Figure 8C,D.
We found that more individuals typically reach the immune state (Mi, and turn into Ml) after passive immunisation (Figure 8C), as a single infectious individual may immunise multiple susceptible nestmates, whereas actively immunised ants need to first be in the infectious state themselves. Yet we found that infections die out (I becomes 0) more quickly under active immunisation (Figure 8D), leaving only a very small reservoir for individuals to become immunised. Moreover, active immunisation leads to a lower number of dead individuals (R). This is despite the fact that contraction of disease through pathogen transfer can only occur in the active route (with a risk of dying similar to our experimental outcome). Increasing this risk leads to higher death rates and lower immunisation in a linear relationship (simulations not shown). Taken together, active immunisation via pathogen transfer seems beneficial, as it allows more rapid disease elimination and produces lower death rates in colonies, except if the pathogen requires only a very low exposure dose to establish lethal infections in its host.
In this study, we identified active immunisation as the underlying mode of social group-level immunisation in ant societies after fungal exposure of single individuals. Social contact to a fungus-exposed individual led to low-level infections in the majority of previously naive nestmates (Figures 3, S1, S3) and to a higher capacity to inhibit fungal growth (Figure 1). We found that these low-level infections per se, even in the absence of social contact, are necessary and sufficient to explain the increased antifungal activity of nestmates (Figure 4). We found no evidence for visual or volatile chemical cues acting as additional trigger signals for the immune stimulation of the nestmates. Furthermore, neither ant behaviour (Figure 2) nor physiology (Figure 5C,D) gave an indication for passive nestmate immunisation via transfer of antimicrobials from either exposed ants or their nestmates to the other group members. Finally, experimental elimination of the active route resulted in the absence of protective antifungal activity in nestmates (Figure 5A,B). The increased immune activity of nestmates of fungus-exposed individuals correlates with an increased expression of immune genes such as the antimicrobial peptide defensin and the enzyme, prophenoloxidase (PPO, Figure 6A,B), which both have known antifungal properties [55],[83]. Cathepsin L, a lysosomal protease rather involved in antibacterial and antiviral responses ([75],[76]; Figure S4), was not expressed at higher levels in nestmates of fungus-exposed compared to control-treated ants (Figure 6C). In addition to the specific immune gene upregulation revealed by our candidate gene approach, we also found in a functional assay that nestmate immunity is not generally increased, but acts against Metarhizium fungus (Figure 1) and not Arthrobacter bacteria (Figure 7). Precisely how specific social immunisation is at both the functional and gene expression levels remains to be addressed, and will be facilitated by the emerging genomic information on ants and other social insects [84]–[87].
To our knowledge, our study provides the first mechanistic explanation for the phenomenon of reduced susceptibility of nestmates after social contact to a fungus-exposed individual, that is, social immunisation, described for both ants [18] and termites [17]. Whether group-level immunisation in termite societies follows the same principle as in Lasius ants remains to be shown. Interestingly, our study on fungal exposure contrasts with findings of the suggested mechanisms of social immunisation of ants after bacterial exposure, where transfer of antimicrobial substances from the exposed individual via social feeding seems to elicit protection of nestmates [16]. We suggest that distinct infection modes of bacterial and fungal pathogens underlie these differences. Bacterial infections typically occur via oral uptake [39], so that bacteria-exposed individuals do not carry socially transferable spores on their cuticle, as is the case with entomopathogenic fungi. Moreover, the long delay between exposure and infection is not common in bacterial infections, allowing for faster production of immune effectors in the exposed individuals and an earlier potential onset of immunisation.
Social immunisation may not be limited to the highly eusocial insect societies but could similarly occur in other societies or at the family level. If also detected in vertebrates, the underlying mechanisms may be very different, as vertebrates have the additional adaptive/acquired immune component and do not rely solely on the innate immune system that characterises invertebrate immunity [1],[21]. Humans have used the intentional transfer of low-level infections—referred to as “variolation” or “inoculation”—in an attempt to fight smallpox and frequently succeeded in creating long-term protection against this otherwise often deadly disease [62],[63]. In humans, the technique was later replaced by less risky immunisation with attenuated strains as soon as these became available [88], but variolation is still used for, for example, poultry disease management [64]. It is still unclear whether acquiring the protective low-level infections in ants is also an active strategy or, rather, an unintentional byproduct of social contact similar to “contact immunity” occurring in human societies, for example, after live strain polio or smallpox vaccination, where vaccinated individuals became spreaders and vaccinated their family members [89],[90]. It is interesting that allogrooming by the ants is not restricted to single individuals, which would be a good strategy to avoid infecting the whole colony, but is rather performed by many colony members, all of which pick up a low-level infection. This may hint at social immunisation by low-level infections being an adaptive evolutionary strategy.
Our epidemiological modeling indeed suggests that active immunisation is a beneficial strategy for ant colonies, as it allows for faster disease elimination and therefore leads to lower death rates than passive immunisation would. This is particularly true if exposure to low pathogen levels confers a low risk of mortality, as is the case with Metarhizium fungus, which requires relatively large doses to elicit a deadly course of disease. We therefore predict that social transfer of pathogens with higher infectivity [91] would not be an advantageous strategy for societies. A comparative analysis of mechanisms employed by social insects against pathogen types differing in their virulence and transmission would thus be highly interesting. Moreover, it seems likely that active immune stimulation following low-level infections may induce individual immune priming and, thereby, a longer lasting protection of colony members than if they simply received immune effectors. The long-lived societies of social insects [43] are at especially high risk of re-encountering the same pathogens multiple times during their lifespans [21], and could greatly benefit from a persistent, rather than transient, social immunisation, particularly against common pathogens such as the fungus Metarhizium. To fully understand long-term epidemiological dynamics at the society level it will be indispensable to learn more about the mechanisms involved at the individual level—for example, to better understand if immune priming plays a role in social immunisation.
The unicolonial ant species Lasius neglectus [92],[93] was sampled from four populations (Jena, Germany; Volterra, Italy; Seva and Bellaterra, both Spain; for details on sample locations, see [94]) and reared in the laboratory as described in Ugelvig and Cremer (2007) [18]. Behavioural observations were performed on workers collected in 2006 from all four populations, whereas all further experiments used L. neglectus workers collected in 2008 from Jena, Germany. Ants were kept at a constant temperature of 23°C with 75% humidity and a day/night cycle of 14 h light/10 h dark during the experiments. Experiments were performed in petri dishes with a plastered floor and 10% sucrose solution as food.
We used the entomopathogenic fungus Metarhizium anisopliae var. anisopliae (strain Ma 275, KVL 03-143; obtained from Prof. J. Eilenberg, Faculty of Life Sciences, University of Copenhagen, Denmark) to expose the ants in our experiments. To determine inhibition of fungal growth by ant material (antifungal activity assay, see below) and the transfer of conidia to the cuticle of nestmates traced by fluorescence microscopy, we used the RFP (Red Fluorescent Protein) labelled strain 2575 ([95]; obtained from Prof. M. Bidochka, Brock University, Canada). For exposure of ants, we applied the fungal conidia (conidiospores)—that is, the dispersal form that is produced in a natural infection cycle from dead insect cadavers [30]—on the ants, whereas we used blastospores—that is, a single cell spore stage produced inside the body of the infected host [30],[52]—for measuring the antifungal activity. Multiple aliquots of conidia of each strain were kept at −80°C and were grown on malt extract agar at 23°C for 2–4 wk prior to each experiment. Conidia were harvested by suspending them in 0.05% Triton X-100 (Sigma) and stored at 4°C for a maximum of 3–4 wk. All conidia suspensions had a germination rate of >98% as determined directly before each experiment. We produced liquid cultures of blastospores following an adjusted protocol by Kleespies and Zimmermann (1994) [96], though growing the spores at 23°C. Blastospores were harvested by sieving them through a sterile 41 µm nylon net filter (Merck Millipore).
We exposed individual ant workers by applying a 0.3 µl droplet of a suspension of 109 conidia/ml in 0.05% Triton X solution (fungus treatment), which corresponds to the lethal dose (LD) 50 for isolated ants. To obtain low-level infections in the same order as those picked up by the nestmates during social contact (as confirmed by comparison of internal infection load of the socially transferred and directly applied group), we exposed the ants to 0.3 µl of a 105 conidia/ml suspension (LD2) and kept them isolated. For the sham control, we treated the ants with a 0.3 µl droplet of a 0.05% Triton X solution only. Subsequently, the ants were dried on a piece of filter paper for several minutes.
We grouped six workers (1 treated individual and 5 naive nestmates, to be distinguished by colour marking [Edding 780]) and three larvae of L. neglectus in a petri dish (Ø = 5.5 cm) with a dampened plaster floor and a piece of filter paper (1×1 cm) moistened with 10% sucrose solution as food supply. The treated individual received either a sham control or a fungus treatment as described above. Our experimental setup is equivalent to the experiment described in more detail in Ugelvig and Cremer (2007) [18], which either led to a social immunisation of nestmates (fungus treatment) or not (sham control) after 5 d of social contact. We used this setup for observations of ant-ant interactions, obtaining physiological immune measures and conidia transmission analysis, yet made some measurements already after 1, 2, or 3 d of social contact.
We changed this general setup for two experiments. First, to determine if signal transfer alone may be sufficient to elicit social immunisation in nestmates, we prevented direct social contact between the treated ant (n = 10 for sham control and fungus treatment, respectively) and its nestmates. This was done by keeping the treated individual in a plastic tube (200 µl, Ø of opening = 0.7 cm, containing cotton wool moistened with 10% sucrose solution), attached to the main petri dish, but separated by a double-layered nylon mesh (mesh size 20 µm). The setup prevented direct physical contact yet allowed exchange of visual or volatile chemical signals. After 3 d, nestmates were frozen and subjected to the antifungal activity assay as described below. In a second setup, we excluded both signal and pathogen transfer from the exposed individual to its nestmates occurring in the first 2 experimental days, only allowing for potential later exchange of antimicrobial substances. To this end, we removed the exposed individual 2 d after fungal exposure from its “early nestmates” and placed it with “new nestmates” (Figure 5A), the latter being tested for their antifungal activity after 3 d with the treated individual (n = 10 replicates for sham control and fungus treatment, respectively). The new nestmates therefore only had contact to an exposed nestmate after conidia had firmly attached to the host's cuticle, and no longer could be transferred to nestmates (as experimentally confirmed by absence of colony forming units [CFUs] in the new nestmates, see below). When removing the treated individual, we added five new nestmates to the five early nestmates (Figure 5B) to test if early nestmates may transfer immunity to the new nestmates in the form of antimicrobial substances. New nestmates were frozen after 3 d of social contact to the early nestmates of either the control-treated or fungus-exposed individual, and their antifungal activity measured as described below.
All workers in the observed ant groups were individually colour marked. We then conducted 10 daily behavioural scan samples for each individual in each of six ant nests (replicates) from each of the four study populations (total n = 24 ant groups per treatment, i.e. 288 ants) over the 5 d of the experiment (as described in [18]). We were interested in the behavioural interactions between different individuals, which we analysed separately for interactions between the treated individual (total interactions n = 240 per treatment) and its nestmates and among nestmates only (total interactions n = 480 per treatment). The following types of interactive behaviours could be recorded: antennation (recognition behaviour), allogrooming (mutual cleaning of the body surface), and trophallaxis (exchange of regurgitated liquid food; [38]). For statistical analysis of the behavioural data, see the statistics section below.
We developed a sensitive antifungal and antibacterial assay (MS, unpublished) that reveals the antimicrobial activity of ant tissue via the growth inhibition of a pathogen culture (as reduced absorbance in a spectrophotometer) compared to a pathogen growth control without an ant sample. For each assay, we first determined the required ratio of pathogen, ant sample, and buffer to be in the linear range of the growth curve in which antimicrobial activity could be detected. We measured growth inhibition against blastospores of M. anisopliae by using either complete ants (n = 10 replicate samples for each group), specific ant body parts (gaster cuticle and thorax; n = 6 replicate samples for each group), or the trophallactic droplet (n = 10 replicate samples for each group) of treated ants (sham control and fungus treatment) and their respective nestmates. Most measurements were taken 3 d (i.e., 72 h) after treatment of the single individual. Nestmates of control and exposed ants were also analysed on day 5 (i.e., 120 h) after treatment. Bacterial growth inhibition against vegetative cells of A. globiformis was determined for the nestmates of fungus-exposed and control-treated individuals (n = 10 replicates each). In all cases, the body parts or exudates from five individuals were pooled to obtain a single replicate sample. Both antifungal and antibacterial activity was determined as the reduction of either M. anisopliae fungal blastospore or A. globiformis bacterial vegetative cell growth, measured as absorbance in a spectrophotometer (SpectraMax M2e, Molecular Devices, similar to [97],[98]), after incubation of ant samples with the fungal or bacterial suspension. For detailed information, see Text S1, and for statistical analyses, see below.
We set up 15 experimental groups each consisting of five nestmates and one individual exposed to RFP-labelled conidia. After 2 d of social contact all ants were removed and frozen at −20°C. The cuticles of three random nestmates per group—that is, 45 nestmates in total—and cuticles from the 15 directly exposed individuals were examined for the presence of RFP-labelled conidia using a fluorescence microscope (Leica MZ16 FA; Software: Leica Application Suite Advanced Fluorescence 2.3.0; Filter Cube: ET DsRed). Each ant was screened for the presence of conidia for a maximum duration of 30 min. In addition we checked the cuticle of 15 naive ants as negative control using the same method. We did not detect any structures resembling RFP-labelled conidia on any of the naive ants.
We exposed 30 ants, kept them in individual petri dishes, and randomly assigned them to either of the three groups (n = 10 ants each): ants that were frozen (−20°C) after 1, 3, or 5 d post-exposure. On day 1 post-exposure 10 of 10 ants were alive, 3 d post-exposure 8 of 10 ants survived, and 5 d post-exposure 4 of 10 ants survived. In addition, we set up 21 experimental groups, each consisting of five nestmates and one fungus-exposed individual, which were also frozen (in equal numbers) 1, 3, or 5 d post-exposure. None of the nestmates had died at this time point.
All individually kept, directly exposed ants (i.e., 10 per day) and two randomly chosen nestmates per experimental group (i.e., 14 per day) were surface-sterilised in ethanol and sodium hypochlorite (as described in [18]) to destroy all fungal material on the cuticle prior to dissection under a stereomicroscope (Leica S6E). For each ant, all contents of the gaster (abdomen) without the cuticle were removed and dissolved in 30 µl of Triton X. The body contents were then plated on selective medium agar plates (containing: chloramphenicol 100 mg/l, streptomycin 50 mg/l, dodin 110 mg/l) and kept at 23°C. After 2 wk of cultivation, the number of colony forming units (CFUs) per plate was determined. We identified CFUs as pure M. anisopliae cultures by morphological fungal determination and amplification of specific M. anisopliae genes by PCR (see Text S1). For statistical analysis, we used both presence/absence of CFUs for each individual and the number of CFUs growing out of infected ants (for details, see statistical analysis section below).
For method development, we performed the following negative controls: (a) 15 completely untreated ants and (b) 15 ants that were exposed to conidia but were surface-sterilised after 3 h (i.e., before the fungus could penetrate the cuticle and reach the inside of the ant). We did not detect any fungal growth from these 30 ants. Moreover, we could confirm that pathogen transfer did not occur towards the new nestmates of either directly exposed ants or early nestmates (n = 14 replicates each).
We set up 30 experimental groups consisting of five nestmates and one fungus-exposed individual each. After the 5 d of social contact to the exposed individuals, each nestmate was isolated in a single petri dish for another 12 d. During the whole experimental period of 17 d, the survival of nestmates was checked daily. Dead nestmates were surface-sterilised as above and put on moist filter paper in a petri dish at constant temperature, 23°C. Cadavers were checked for a period of 3 wk for the growth of M. anisopliae.
The bacterium Bacillus thuringiensis (strain NRRL B-18765, obtained from the permanent strain collection of the Northern Research Laboratory, U.S. Department of Agriculture, Peoria, Illinois, USA) was precultured in LB medium and grown to an OD600 of 0.1. We centrifuged 1 ml of the suspension at a speed of 3,000× g for 5 min and discarded the supernatant to obtain a concentrated bacterial pellet as in [99]. Ants were immobilized and pricked ventrally between the 2nd and 3rd gaster sternite with a sterilized needle (minutien needles, Sphinx V2A 0.1×12 mm, bioform) dipped in either LB medium (sham control) or the concentrated bacterial pellet (n = 10 ants per treatment, replicated three times; i.e., total n = 30 ants per treatment). The ants were frozen for gene expression analysis 12 h after pricking.
Ants were analysed either individually (nestmates of Metarhizium-exposed ants) or in pools of 10 ants (bacterial septic injury) by qPCR for gene expression of three immune genes and the housekeeping gene, 18s rRNA. For immune genes, we chose the antimicrobial peptide defensin [68],[69], the enzyme prophenoloxidase (PPO [72],[73]), and the lysosomal protease cathepsin L [74],[76]. For details of the procedures on RNA extraction, cDNA preparation, and qPCR, please see Text S1 and the statistical analysis section below.
We always tested the distributions underlying our data and chose the corresponding tests. If data were not normally distributed even after transformation, we applied models with specified error structures or non-parametric tests. Reported p values are two-sided. All statistical analyses were carried out in IBM SPSS Statistics version 19.0 or Sigma Stat 3.5 (Systat Software Inc.). All figures are based on raw data.
For the behavioural observations, we first analysed all behaviours overall over the 5 experimental days. Due to the nature of the data (overdispersed count data), generalised linear models (GLM) with negative binomial errors and a log link function were employed using the following factors: treatment type (fungus treatment versus sham control), ant pairing (treated-nestmate versus nestmate-nestmate), and the interaction between them. As neither nests within populations nor populations behaved differently, they were not included in the final models. We give the likelihood ratio (LR) χ2 to test if our overall model explains the data better than a model with only the intercept. As we detected significant differences for allogrooming, we performed a second test to analyse the effect of time in the interactions between treated individuals and their nestmates for the two treatment types separately (n = 240) using a GLM with repeated measures. Simple contrasts with day 1 as reference were employed to test the differences between day 1 and the succeeding days (Figure 2B).
For statistical analysis of the antifungal and antibacterial activity, the absorbance values (optical density) of the different treatment groups were compared by one-way ANOVAs or t tests as data were normally distributed or could be transformed to obtain normality. For the antifungal activity of nestmates of exposed versus control nestmates, we applied a GLM to analyse the effects of treatment type (fungus treatment versus sham control) and time (day 3 versus day 5 post-treatment), as well as their interaction (Figure 1).
For analysis of pathogen load, we compared directly exposed and nestmate ants for (a) the proportion of individuals that were infected (i.e., showed at least a single CFU; Fisher exact test) and (b) the number of CFUs in the individuals that showed an infection (Mann Whitney U test; Figure 3). As the experimental grouping did not influence the number of CFUs found in nestmates from the same ant group, this factor could be excluded from statistical analysis comparing treated individuals and nestmates (GLM with negative binomial errors, LR χ2 = 112.362, df = 34, p = 0.000; Replicate, Wald: χ2 = 21.273, df = 17, p = 0.214).
Gene expression analyses were run in two to three technical replicates. Normalised gene expression values (the average of technical replicates, standardised to the housekeeping gene) were either a priori normally distributed or could be normalised by transformation and were analysed using t test or—in the case of unequal variances between groups (defensin, Figure 6A)—Welch's t test for unequal variances [100].
We applied ordinary differential equations (ODE) to extend classical SIR modeling (Susceptible-Infectious-Removed) with an immunised state to a SIRM model (Susceptible-Infectious-Removed-iMmune), in which the immune individuals were further separated into an initial and a late phase of immunity. See Figure 8A,B for the model and how we calculated state changes and Text S2 for model construction and simulations.
|
10.1371/journal.pgen.1006235 | Accelerating Gene Discovery by Phenotyping Whole-Genome Sequenced Multi-mutation Strains and Using the Sequence Kernel Association Test (SKAT) | Forward genetic screens represent powerful, unbiased approaches to uncover novel components in any biological process. Such screens suffer from a major bottleneck, however, namely the cloning of corresponding genes causing the phenotypic variation. Reverse genetic screens have been employed as a way to circumvent this issue, but can often be limited in scope. Here we demonstrate an innovative approach to gene discovery. Using C. elegans as a model system, we used a whole-genome sequenced multi-mutation library, from the Million Mutation Project, together with the Sequence Kernel Association Test (SKAT), to rapidly screen for and identify genes associated with a phenotype of interest, namely defects in dye-filling of ciliated sensory neurons. Such anomalies in dye-filling are often associated with the disruption of cilia, organelles which in humans are implicated in sensory physiology (including vision, smell and hearing), development and disease. Beyond identifying several well characterised dye-filling genes, our approach uncovered three genes not previously linked to ciliated sensory neuron development or function. From these putative novel dye-filling genes, we confirmed the involvement of BGNT-1.1 in ciliated sensory neuron function and morphogenesis. BGNT-1.1 functions at the trans-Golgi network of sheath cells (glia) to influence dye-filling and cilium length, in a cell non-autonomous manner. Notably, BGNT-1.1 is the orthologue of human B3GNT1/B4GAT1, a glycosyltransferase associated with Walker-Warburg syndrome (WWS). WWS is a multigenic disorder characterised by muscular dystrophy as well as brain and eye anomalies. Together, our work unveils an effective and innovative approach to gene discovery, and provides the first evidence that B3GNT1-associated Walker-Warburg syndrome may be considered a ciliopathy.
| Model organisms are useful tools for uncovering new genes involved in a biological process via genetic screens. Such an approach is powerful, but suffers from drawbacks that can slow down gene discovery. In forward genetics screens, difficult-to-map phenotypes present daunting challenges, and whole-genome coverage can be equally challenging for reverse genetic screens where typically only a single gene’s function is assayed per strain. Here, we show a different approach which includes positive aspects of forward (high-coverage, randomly-induced mutations) and reverse genetics (prior knowledge of gene disruption) to accelerate gene discovery. We paired a whole-genome sequenced multi-mutation C. elegans library with a rare-variant associated test to rapidly identify genes associated with a phenotype of interest: defects in sensory neurons bearing sensory organelles called cilia, via a simple dye-filling assay to probe the form and function of these cells. We found two well characterised dye-filling genes and three genes, not previously linked to ciliated sensory neuron development or function, that were associated with dye-filling defects. We reveal that disruption of one of these (BGNT-1.1), whose human orthologue is associated with Walker-Warburg syndrome, results in abrogated uptake of dye and cilia length defects. We believe that our novel approach is useful for any organism with a small genome that can be quickly sequenced and where many mutant strains can be easily isolated and phenotyped, such as Drosophila and Arabidopsis.
| A powerful, tried and true approach to identify which genes function in a particular biological process is to create collections of organisms harbouring multiple mutations via random mutagenesis, followed by screening the mutant library for organisms that exhibit the desired altered phenotypes. Although such forward genetics strategies have produced numerous fundamental discoveries, a significant limitation of this approach in metazoans is the prolonged time required to identify the causative mutations. The bottleneck typically arises from the required genetic mapping, complementation tests to exclude known genes, and sequencing of candidates genes.
To circumvent the major disadvantage of forward genetics, reverse genetic approaches have been employed. Various strategies for disrupting a collection of known genes (e.g., RNAi, homologous recombination, transposon mutagenesis, etc.) are combined with phenotypic screening to identify candidates. Reverse genetics approaches also have drawbacks, however, including the need to handle and process tens of thousands of strains to assay the entire genome, off-target effects in the case of RNAi, and omission of essential genes.
We hypothesised that we could use whole-genome sequencing in combination with statistical genetics to inaugurate a novel gene discovery approach which retains the advantages of both forward and reverse genetics, and yet minimises their downsides. To do this, we employed the Million Mutation Project (MMP) [1], a collection of 2007 Caenorhabditis elegans strains harbouring randomly-induced mutations whose genomes are fully sequenced (data is publicly available: http://genome.sfu.ca/mmp/about.html). This mutant library represents an unprecedented genetic resource for any multicellular organism, wherein the strains collectively contain one or more potentially disruptive alleles affecting nearly all C. elegans coding regions. On average, each strain contains ~ 400 non-synonymous mutations affecting protein coding sequences.
We postulated that this whole-genome sequence information would allow an “eyes wide open” approach when performing a genetic screen, such that pairing this resource with a high-throughput assay would enable rapid discovery of genes not previously associated with our biological process of interest. Here, we demonstrate that testing for association between variants from the MMP library and phenotype data with the Sequence Kernel Association test (SKAT) [2] allows us to effectively and efficiently predict novel genes important for our chosen biological process: the development and function of the amphid and phasmid sensillum, which includes both ciliated sensory neurons as well as glial-like neuronal support cells.
Primary (non-motile) cilia arise from a modified centriole (basal body) and act as 'cellular antennae' that transduce environmental cues to the cell [3]. They enable sensory physiology (such as olfaction/chemosensation, mechanosensation, vision) and are central to signalling pathways essential for metazoan development [4]. Dysfunction of cilia is implicated in a number of human diseases, including polycystic kidney disease, congenital heart disease, and an emerging group of genetic disorders termed ciliopathies (e.g., Bardet-Biedl, Meckel-Gruber and Joubert Syndromes). In these ciliopathies, disruption of many, if not all, cilia in the human body results in a plethora of defects, including retinal degeneration, organ cyst formation, obesity, brain malformations, and various other ailments [5][6].
In C. elegans, the uptake of a fluorescent lipophilic dye, DiI, from the environment is used to probe the integrity of the amphid and phasmid sensillum, which includes cilia and ciliated sensory neurons, as well as glial-like sheath cells. DiI is selectively incorporated into six pairs of ciliated amphid channel sensory neurons in the head (ADF, ADL, ASH, ASI, ASJ, and ASK), and two pairs of ciliated phasmid channel sensory neurons in the tail (PHA and PHB), via environmentally-exposed cilia present at the tips of dendrites (S1 Fig) [7,8]. Many dye-filling (dyf) mutants known from genetic screens [8,9] harbour mutations in genes influencing ciliated sensory neuron development and function, including ciliogenesis [10], cilia maintenance [11], axon guidance [9], dendrite anchoring/formation [10], as well as cell fate [12]. Importantly, non-cell autonomous effects from disruption of neural support (glial) cells can also result in dye-filling defects [10,13].
When we applied SKAT to the phenotype data we collected from screening the MMP strains for dye-filling, we found that a previously uncharacterised C. elegans gene, bgnt-1.1/F01D4.9, plays an essential role in this process. We found that the ciliated sensory neurons of bgnt-1.1 mutants fail to fill with a lipophilic dye, a phenotype indicative of their dysfunction, and that BGNT-1.1 localises specifically to the trans-Golgi network of the amphid and phasmid sheath cells. These are glial-like neuronal support cells, which are critical for the development and function of ciliated sensory neurons. Interestingly, bgnt-1.1 is the orthologue of human B3GNT1/B4GAT1, a gene implicated in Walker-Warburg syndrome [14,15], a disorder with clinical ailments resembling a ciliary disease (ciliopathy).
We screened 480 randomly-chosen whole-genome sequenced multi-mutation MMP strains, ~25% of the library, for defects in DiI uptake in amphid and phasmid ciliated sensory neurons (Fig 1). We found 40 MMP strains which exhibit significant amphid dye-filling defects and 40 MMP strains which exhibit significant phasmid dye-filling defects; the strains with amphid and phasmid dye-filling defects are not necessarily identical (Fig 1C, Table 1, S1 Table).
We identified 11 completely dye-fill defective strains, where all worms sampled failed to take up dye. A preliminary look at the data indicates that of these, 10 contained deleterious (“knockout”) mutations in previously identified dye-filling genes (e.g., nonsense and frameshift-inducing deletions; S2 Table). Additionally, we uncovered 47 partially dye-fill defective strains, where a proportion of the population failed to fill with dye significantly more often than wild-type worms. Of these partially dye-fill defective strains, 1 harbours a nonsense mutation and 10 display missense mutations in known dye-filling genes (S2 Table).
Despite the fact that we can identify some strains with mutations in genes previously shown to cause dye-filling defects, it is not clear that it is the mutations in these genes which are necessarily the cause of the dye-filling defects in these strains. Furthermore, there are 38 strains where we cannot generate a hypothesis as to what genetic variation is responsible for the dye-filling defect. To facilitate identification of genes responsible for the observed dye-fill defects, we hypothesised that a recently developed statistical genetics approach commonly used in human genetics, but underutilised in model organisms, would allow for the rapid prioritisation of candidate genes. Specifically, we chose to employ the sequence kernel association test (SKAT) to uncover genes associated with the dye-filling phenotype. SKAT is a regression method to test for association between rare and/or common genetic variants in a region and a continuous or dichotomous trait [2].
We chose SKAT over other statistical analyses for several reasons. For our dataset, it was imperative that we chose an association test that deals effectively with rare variants, as 800,000/850,000 of the non-synonymous variants in the MMP library are unique—meaning that they are present in only a single isogenic strain in the library. Hence, genome-wide association study (GWAS) approaches, which typically test for an association between common variants (generally defined as a minor allele frequency > 5%) and a trait of interest, would be unsuitable for analysis of phenotype datasets derived from the MMP library. We also viewed SKAT as an optimal method to use for our dataset because it permits the use of prior information to assign weights to genetic variants. For example, nonsense mutations might be expected to be more deleterious than other variants which may cause more modest changes to the protein, such as missense mutations and in-frame deletions. The C-alpha test [16], which is quite similar to SKAT in the absence of covariants (e.g., age, sex, etc.), could have also been used for our dataset, but we chose to employ SKAT because it facilitates implementing and assigning biologically relevant weights to variants. Finally, SKAT was chosen over other related burden tests, such as the cohort allelic sums test (CAST) [17] and the combined multivariate and collapsing (CMC) method [18], because unlike these tests, SKAT does not assume that all (common) variants will affect the trait in the same direction.
Given that the groups of worms which have amphid dye-filling and phasmid dye-filling defects do not necessarily overlap, we performed SKAT separately for each dataset. We chose to perform the linear regression version of SKAT in combination with log transformation of the response (phenotype) variable, as opposed to a logistic regression version of SKAT because in its current implementation, the logistic regression version of SKAT does not work with proportion data, and takes only dichotomous traits coded as 0 or 1. Quantile-quantile (QQ)-plots were used to choose the appropriate constant to add to the response (phenotype) variable before log transformation (S2 & S3 Figs). Finally, we performed SKAT with biologically relevant weights assigned to the variants. We assigned mutations which would likely result in the creation of a null mutation (nonsense and splicing mutations, as well as frameshift causing deletions) a weight of 1, mutations which would result in truncation of the protein (in-frame deletions) a weight of 0.75, and mutations which would result in a change in amino acid sequence (missense mutation) a weight of 0.25. We hypothesised these were reasonable weights to assign to each class of mutation based on the current knowledge in the field of genetics.
Genome-wide SKAT analyses using biologically relevant weights on the amphid dye-filling dataset reveal 5 genes that reach significance when we adjust for multiple testing using a false discovery rate (FDR; Benjamini-Hochberg procedure) of 5% (FDR adjusted p-value was < 0.05, Table 2, S3 Table). SKAT analyses using biologically relevant weights on the phasmid dye-filling dataset uncovered 3 genes which reached significance, again using a FDR of 5% (FDR adjusted p-value was < 0.05, Table 3, S4 Table). Dye-filling defects of both amphid and phasmid ciliated neurons is significantly associated with genes encoding intraflagellar transport proteins (OSM-1 and CHE-3), and a glycosyltransferase (BGNT-1.1; Tables 2 and 3). Amphid-specific dye-filling defects are found to be associated with genes encoding an Arf-GAP related protein, CNT-1, as well as a mitotic spindle assembly checkpoint protein, MDF-1 (Table 2). No gene was found to be significantly associated with only phasmid dye-filling defects (Table 2 and Table 3).
Of the three genes associated with both amphid and phasmid dye-filling defects, namely osm-1, che-3, and bgnt-1.1, the first two are well characterised genes whose dye-filling defective phenotypes are ascribed to their key roles in intraflagellar transport (IFT). OSM-1 is the orthologue of mammalian IFT172, an IFT-B subcomplex component which functions as an adaptor to link ciliary cargo (e.g., tubulin, receptors and signaling molecules) to the anterograde IFT kinesin motors, and is necessary for ciliogenesis [10]. CHE-3, the orthologue of mammalian DYNC2H1, is a cytoplasmic dynein heavy chain which powers the retrograde IFT-dynein motor. This molecular motor recycles IFT machinery from the growing ciliary tip back to the ciliary base and is also necessary for proper cilium formation/maintenance [19,20]. These two known dye-fill/cilia genes represent excellent positive controls for our screen, and indicate that other genes found to be significantly associated with these phenotypes may be novel dye-fill genes that influence cilia function.
Interestingly, one of the other amphid dye-filling gene hits, cnt-1, encodes a protein that play roles in membrane trafficking/dynamics by influencing small GTPase function, via GTPase-activating protein (GAP) activity. The general involvement of small GTPases of the Arf, Arf-like (Arl) and Rab families in cilium formation/development is well established [3]. cnt-1 encodes the orthologue of human ACAP2, which interacts with both Rab35 [21] and Arf6 [22] to mediate crosstalk between these two proteins, at least in the context of PC12 cell neurite outgrowth, and potentially through endocytic recycling [23]. Another amphid dye-filling gene hit, mdf-1, is homologous to Mad1, and encodes a mitotic spindle assembly checkpoint protein [24]. To the best of our knowledge, our findings are the first to directly implicate mdf-1/Mad1 as being important for cilia development and/or function but other mitotic spindle assembly checkpoint proteins have previously been linked to cilia, including BUBR1 [25] and APC-Cdc20 [26].
The third, putative novel dye-filling gene significantly associated with both amphid and phasmid dye-fill phenotypes is bgnt-1.1 (Tables 2 & 3, S3 and S4 Tables). bgnt-1.1 encodes an unstudied C. elegans glycosyltransferase 49 family member homologous to human B3GNT1/B4GAT1 (S4 Fig). B3GNT1 catalyses the addition of β1–3 linked N-acetylglucosamine to galactose [27]. In HeLa cells, its subcellular localisation is concentrated at the trans-Golgi [28]. B3gnt1 knockout mice exhibit axon guidance phenotypes [29,30] and deficient behavioural responses to estrous females [31]. In humans, mutations in B3GNT1 are associated with a congenital muscular dystrophy with brain and eye anomalies, Walker-Warburg syndrome (WWS) [14,15]. WWS is a suspected, but unconfirmed ciliopathy; it exhibits 6 core features common to ciliopathies, including Dandy-Walker malformation, hypoplasia of the corpus callosum, mental retardation, posterior encephalocele, retinitis pigmentosa and situs inversus [5]. Additionally, one patient is reported to exhibit dysplastic kidneys [14], a developmental disruption which leads to cyst formation, illuminating a potential 7th core ciliopathy feature to this disorder, renal cystic disease. To divulge a potential connection between B3GNT1 and cilia and/or ciliated sensory neuron function, we sought to confirm the role of C. elegans BGNT-1.1 in dye-filling, and analyse its involvement in ciliated sensory neuron development.
Of the eight MMP strains harbouring mutations in bgnt-1.1, three (VC20615, VC20628 and VC20326) exhibit severe dye-fill phenotypes (S1 Table). The C -> T missense mutation in bgnt-1.1 in VC20615 corresponds to P194S alteration in the protein sequence, while VC20628 and VC20326 each harbour an identical G -> A missense mutation in bgnt-1.1, which leads to a G205E amino acid change in the protein sequence. Both of these mutations alter conserved amino acid residues (S5 Fig). In the screen we encountered 5 additional strains that harbour missense mutations in bgnt-1.1 but did not exhibit dye-filling defects. Close examination of the predicted effects of these missense mutations on the amino acid sequence of the protein indicates that these alleles do not lead to amino acid changes in conserved residues (S5 Fig), and thus it is not surprising that these strains do not exhibit dye-fill defects.
To confirm that the mutations in bgnt-1.1 is responsible for the dye-filling phenotypes in bgnt-1.1 mutants, we rescued the dye-fill defects by expressing a fosmid containing a wild-type copy of bgnt-1.1 in an extrachromosomal array (Fig 2). Another way to confirm that disruption of bgnt-1.1 causes dye-fill defects would be to observe this phenotype in a strain harbouring a knock-out mutation in bgnt-1.1. Although there are 49 bgnt-1.1 alleles available, a knock-out allele of bgnt-1.1 did not yet exist. There are two insertion/deletion alleles available, gk1221 and tm4314, but both fall within introns and thus unlikely to affect protein function. Thus, we also tested the causality of bgnt-1.1 via a relatively efficient SNP mapping approach. We established that the dye-fill phenotypes from VC20615 and VC20628 strains map to the bgnt-1.1 locus, on chromosome IV between -5 cM and 8 cM (S6 Fig). Notably, in both VC20615 and VC20628 strains, bgnt-1.1 is the only gene in this region harbouring a mutation which is common to both of these strains.
Finally, we used CRISPR-Cas9 genome engineering [32] to independently generate three bgnt-1.1 knockout alleles. All three alleles delete the first and second exon of bgnt-1.1 and insert a selectable marker, Pmyo-2::GFP, in their place. When tested for dye-filling defects, we observe an identical dye-filling phenotype as found in the 6X outcrossed bgnt-1.1 (gk210889) G205E allele from the Million Mutation Project (Fig 3). In all of these mutants, we observe a great decrease in the amount of dye that enters the amphid ciliated sensory neurons, which is often undetectable, as well as a complete absence of dye-filling of the phasmid ciliated sensory neurons. Together, these findings indicate that loss of bgnt-1.1 function results in dye-filling defects.
To shed light on how bgnt-1.1 affects dye-filling, we studied expression of GFP-tagged BGNT-1.1 constructed via fosmid recombineering (https://transgeneome.mpi-cbg.de/transgeneomics/index.html), and thus containing all of this gene’s endogenous regulatory elements. We find that in C. elegans, the protein localises to discrete structures in the cell body of the amphid and phasmid glial-like sheath cells (AMsh and PHsh, respectively) in the head and tail of the animal (Fig 4A). These cells are intimately associated with the ciliated sensory neurons in the pore region where cilia are exposed to the external environment (S1 Fig).
Mammalian B3GNT1 is found at the trans-Golgi network in HeLa cells [28]. To assess whether this is also where C. elegans BGNT-1.1 localises, we performed antibody staining for SQL-1, an established cis-Golgi marker [33], in the strain expressing BGNT-1.1::GFP. We observe that in both the head and tail, the localisation of BGNT-1.1::GFP is always proximal to the discrete SQL-1 puncta, indicating that C. elegans BGNT-1.1 is concentrated at the trans-Golgi, as expected (Fig 4B).
Next, we queried whether loss of bgnt-1.1 function in the amphid and phasmid sheath cells leads to any gross ciliary morphology defects by expressing a ciliary marker in bgnt-1.1 mutants, namely the GFP-tagged IFT-B subcomplex protein, CHE-2 (IFT80). This experiment indicates that although the cilia of bgnt-1.1 mutants fail to fill with dye, their ciliary structures appear superficially wild-type (S7A Fig). Since modest cilia structure defects may be more difficult to observe using pan-cilia markers, due to overlapping ciliary signals, we also characterised the phenotype of cilia and dendrites in bgnt-1.1 mutants within a single ciliated amphid cell, the ADL neuron. For this purpose, we used the primarily cell-specific ADL promoter, Psrh-220, to drive expression of another cilia marker, IFT-20 (IFT20) tagged with tdTomato. In this strain, we also expressed cytoplasmic GFP in the amphid socket cells so that we could evaluate whether or not the ADL cilia were correctly associated with the surrounding glial support cells and the pore where DiI has access to the amphid ciliated sensory neurons from the environment. Similar to the experiment with the CHE-2::GFP pan-cilia marker, the Psrh-220::IFT-20::tdTomato marker revealed that the ADL cilia and amphid socket (Amso) cell morphology also appear superficially wild-type in bgnt-1.1 mutants (S7B Fig). We then sought to assay for potential phenotypes involving ADL cilia length (Fig 4C); length of socket cell penetration by ADL (proxied by the distance from the distal tip of ADL cilia to the distal end of the socket cell tip; S7C Fig); ADL guidance (proportion of double rod cilia/amphid; S7D Fig); and finally, ADL dendrite blebbing (structural alteration where dendrites take bead on a string appearance; S7E Fig). Our analyses reveal that ADL cilia in bgnt-1.1 mutants are wild-type in most aspects except for a modest cilia length defect. Specifically, bgnt-1.1 mutants were observed to have significantly longer cilia compared to wild-type worms (Fig 4C; p < 0.01, Kruskal-Wallis test).
BGNT-1.1 therefore influences amphid and phasmid neuron development and function, as well all modestly affects cilium length, without overtly affecting the gross structure of neurons or cilium formation. The localisation of BGNT-1.1 at the trans-Golgi network of sheath cells signifies that its effect on ciliated sensory neurons is non-cell autonomous. Interestingly, when the bgnt-1.1 amphid ciliated sensory neurons do fill with dye, we observe bright accumulations of dye along and/or beside the dendrites (Fig 3C). These are often brighter than the staining of the cell bodies. Accumulations of dye have been observed in wild-type worms and have been attributed to the dye-filling of the amphid sheath cells [34], but these are qualitatively much smaller than what we observed in the bgnt-1.1 mutants. This suggests a potential alteration in the ability of the sheath cell to take up, or intracellularly distribute dye when BGNT-1.1 is disrupted.
In humans, mutations in B3GNT1 cause Walker-Warburg syndrome. Given that mutations in B3GNT1 lead to WWS and that it is classified as a dystroglycanopathy [14,15], a group of muscular disorders whose etiology is hypothesised to be caused by aberrant glycosylation of dystroglycan [35], we tested whether or not the C. elegans dystroglycan homologs, dgn-1, dgn-2 and dgn-3, exhibited dye-filling phenotypes. We find that all dgn mutants exhibit dye-filling indistinguishable from wild-type worms (S8 Fig), indicating that BGNT-1.1 function in dye-filling is likely independent of dystroglycan. Interestingly, as highlighted earlier, the WWS congenital muscular dystrophy exhibits 6 features beyond muscle structure/function disruption which are core ciliary disorder (ciliopathy) features [5]. Our findings that C. elegans bgnt-1.1 is expressed in glial cells directly associated with, and necessary for the function of ciliated sensory neurons, is consistent with its role in cilium-dependent dye-filling.
Here we demonstrate that rare-variant association analysis (e.g., SKAT) is an efficient way to rapidly uncover novel genes for a phenotype of interest (e.g., ciliated sensory neuron function) in whole-genome sequenced strains harbouring multiple mutations induced via random mutagenesis. We found that three cilia-related genes, osm-1, che-3 and bgnt-1.1 were significantly associated with dye-filling defects, suggesting that the remaining unstudied genes, cnt-1 and mdf-1, and potentially additional genes we found to have dye-fill defects, likely represent genes important for ciliary/sensory neuron development and/or function.
We confirmed that bgnt-1.1, a gene identified by SKAT as being associated with the dye-filling phenotypes but not previously implicated in cilia or amphid-sensillum function, is a bona fide dye-filling gene. We observed that: (1) two missense mutations in bgnt-1.1 result in severe dye-fill defects in three MMP strains, and three CRISP-Cas9-mediated bgnt-1.1 gene disruptions also cause dye-fill defects; (2) a fosmid containing full-length wild-type bgnt-1.1 rescues the dye-filling phenotype in bgnt-1.1 mutants; (3) the dye-filling phenotypes in the MMP strains with mutations in bgnt-1.1 map to the bgnt-1.1 locus; (4) BGTN-1 is expressed in sheath cells, which are directly implicated in dye-filling; and finally; and (5) mutations in bgnt-1.1 result in a small but statistically significant ciliary length defect. Together, these data strongly indicate that BGNT-1.1, which we find localises as expected to the trans-Golgi network, functions in sheath (glia-like) cells to influence dye-filling.
The power of genome-wide rare-variant association analysis (e.g., SKAT) augments as the number of strains increases (the probability of additional mutations in specific genes is increased), and thus, screening the entire MMP library would likely uncover many additional genes associated with dye-filling defects. To try to assess the minimal number of strains that should be assayed with this approach we performed a power analysis. Raw amphid dye-filling phenotype and genotype data was randomly sub-sampled (without replacement) and analysis was performed via SKAT with, and without, biologically relevant weights. This was done 100 times for each sample size (50, 100, 200, 300, 400). For each sample size, power was calculated as the proportion of times the analysis found at least one gene to be significantly associated with the phenotype. We find that there is 40% power to detect a single gene as being associated with the amphid dye-filling phenotype at N = 400 for both SKAT with and without biologically relevant weights (S9 Fig). Thus, we recommend that future studies using this method should use a sample size close to what was used in this present study (~ 500) to maximize the probability that one or more gene(s) will be found that is significantly associated with the phenotype of interest.
We performed SKAT analyses via two methods, 1) while applying biologically relevant weights to the variants (S3 & S4 Tables), and 2) while weighting all variants equally (S5 & S6 Tables). SKAT analysis of the 480 strains without weights was less powerful, and resulted in identifying only 3 genes as being significantly associated with the amphid dye-filling phenotype and 1 gene as being significantly associated with the phasmid dye-filling phenotype; compared to 5 genes and 3 genes, respectively, when biologically relevant weights were used. However, there appears to be no difference in power when SKAT is performed with or without weights at smaller sample sizes (S9 Fig). Thus, to maximize the ability to detect genes associated with the phenotype of interest, in addition to recommending a minimum sample size of ~ 500, we also recommend assigning biologically relevant weights when using SKAT with the MMP library. The weight assignment could be simple, as done here, or more complex, calculating, for example, the SIFT [36] or Polyphen-2 [37] scores for assessing the severity of each variant in the MMP library.
The genome-wide statistical genetic approach presented here has several advantages over traditional screening approaches. It generates a prioritised list of candidate genes likely responsible for the phenotype of interest. After this list is generated via screening and SKAT analysis, candidates can be tested for their causality of the phenotype through several standard genetic approaches in C. elegans. Candidates could be confirmed, for example, by (i) testing for the phenotype in knock-out mutants or RNAi; (ii) genetic rescue experiments; (iii) performing a genetic complementation test between two loss of function alleles; or, (iv) mapping the mutation to the gene locus. This strategy may work for phenotypes where the traditional polymorphic SNP-mapping strain, CB4856, diverges from the reference wild-type strain, N2, from which the MMP library was generated [1], as well as partially-penetrant or other difficult-to-score phenotypes. In the case of bgnt-1.1, we performed genetic rescue experiments, SNP mapping and created CRISPR-Cas9 knockout strains to support our the SKAT findings, which together confirm that bgnt-1.1 mutations cause dye-filling defects.
Another potential extension and utility of this approach that could work for some (non-neural) phenotypes would be pairing the screening of the MMP strains with RNAi to look for enhancing, suppressing or synthetic phenotypes, and then using SKAT to prioritise a list of candidate genes. Furthermore, as more data from multiple phenotypes are collected on the MMP strains, these could be combined to perform multi-variate genome-wide statistical analysis on whole-genome sequence data. Such approaches are more powerful than univariate approaches in the case of SNP array data [38–40], and such tests can also indicate which variants are pleiotropic, or specific to a single phenotype. How to perform this multivariate phenotype analysis on whole-genome sequences is currently an active area of research and tools to make this possible are being developed, with [41] looking promising.
There are also challenges and limitations to the statistical genetic approach presented here. First, this approach of performing a “medium”-scale screen of the MMP strains is limited to assays that can be done without genetic manipulation of the strains. For example, introducing a functional ‘reporter’ (transgene) into 480 strains would require a prohibitive amount of work, although this has been done for 90 MMP strains [42]. Second, the statistical analysis presented here is only possible for genes which have > 1 variant in the population of worms screened. In practice, we found it works optimally for genes with at least 7 variants. This is due to the distribution of p-values when attempting to control for multiple testing; in our dataset, fewer than 7 variants led to a skewed p-value distribution and an inflation of False-discovery rate adjusted p-values. This strict rule demanding high-coverage for our SKAT analysis leads to only 1150 genes in the 480 MMP strains being considered here. This is due to the distribution of variant counts per gene in the MMP strains (S10 Fig), which exponentially decreases from 1 to N.
How disrupting BGNT-1.1 abrogates dye filling remains uncertain. One possibility is that the glycosyltransferase regulates the association of cilia with the sheath and socket glial-like cells which envelop them (S1 Fig). Specifically, we hypothesise that BGNT-1.1 functions in the trans-golgi network of the amphid and phasmid sheath to glycosylate key unidentified protein(s) important for the association of this sensillum organ. This defect will not be visible at the level of light microscopy, and could perhaps result from changes to the lamellar membrane that surround the amphid/phasmid cilia, or the secreted extracellular material lining these channels. Which substrate(s) the β1,3-N-acetylglucosaminyltransferase, BGNT-1.1 (B3GNT1), glycosylates, and how this influences sensory neuron/glial cell development and function, remains to be determined in a future, detailed study of the gene.
In conclusion, we demonstrated the utility and efficiency of using deep-sequenced multi-mutant strains in combination with SKAT to rapidly uncover novel genes required for a biological process of interest—here, ciliated sensory neuron development and/or function. The role of BGNT-1.1 in this process, seemingly independent of dystroglycan, supports the notion that B3GNT1/B4GAT1-associated Walker-Warburg syndrome may result at least in part from ciliary dysfunction, and thus could be considered a novel ciliopathy. Our findings also underscore the importance of identifying novel dye-filling genes, some of which might be implicated in human ciliopathies. For all new putative dye-filling genes highlighted in this study, we had no prior knowledge of their importance in ciliated sensory neuron function, and may not have (easily) uncovered them using alternative methods. Our approach therefore reduces the hurdle of traditional forward genetic methods, namely identifying the causative allele, and improves upon reverse genetics by allowing high gene/mutation coverage in a relatively small number of strains. Lastly, we propose that our approach is applicable not only for C. elegans, but any organism with a small genome that can be quickly sequenced and where numerous mutant strains can be isolated and phenotyped with relative ease, including Drosophila and Arabidopsis.
Worms were cultured on Nematode Growth Medium (NGM) seeded with Escherichia coli (OP50) at 20°C as described previously [43]. The following strains were obtained from the Caenorhabditis Genetics Center (University of Minnesota, Minneapolis, MN): N2 Bristol, CB4856, CH1869, CH1878 and PR813. VC2010, the wild-type reference strain used during the dye-filling screen, was derived from N2 [44]. The Million Mutation Project strains were isolated and their genomes’ sequenced by Thompson et al. [1]. The 480 Million Mutation Project strains used in this study are listed in S1 Table.
For native rescue of VC20628 bgnt-1.1(gk361915), 25 ng/μl of fosmid WRM065bB05 containing bgnt-1.1 was injected into bgnt-1.1 mutants along with 80 ng/μl of pRF4 rol-6(su1006dm) as a co-injection marker. bgnt-1.1(gk361915); Ex[CHE-2::GFP; pRF4] was created by crossing bgnt-1.1(gk361915) with wild-type worms expressing Ex[CHE-2::GFP; pRF4]. The translational Psrh-220::IFT-20::tdTomato fusion was generated as described in [11], except that tdTomato was used in place of GFP. 1 μl of the PCR product was microinjected into germline of gravid worms along with a co-injection markers (pRF4 rol-6(su1006dm), final concentration of 100 ng/μl). Stable lines expressing this extrachromosomal array were crossed into DM13283 dpy-5(e907); sIs12964[Pgrd-15::GFP; pCeh361] to create the strain MX1924 dpy-5(e907); Ex[Psrh-220::IFT-20::tdTomato; pRF4]; sIs12964[Pgrd-15::GFP; pCeh361]. bgnt-1.1(gk361915) was also introduced to this line via genetic crossing to create MX2236 bgnt-1.1(gk361915); dpy-5(e907); Ex[Psrh-220::IFT-20::tdTomato; pRF4]; sIs12964[Pgrd-15::GFP; pCeh361]. The BGNT-1.1::GFP recombineered fosmid construct (Construct # 6821068113870966 H08) was obtained from the TransgeneOme (https://transgeneome.mpi-cbg.de/transgeneomics/index.html)). To generate a strain expressing this construct, 25 ng/μl of the BGNT-1.1::GFP recombineered fosmid was injected into N2 worms along with 4 ng/ul of Posm-5::XBX-1::tdTomato as a cilia-marker, and 80 ng/μl of pRF4 rol-6(su1006dm) as a co-injection marker.
VC3671 (gk3637), VC3674 (gk3638), and VC3675 (gk3639) for bgnt-1.1/F01D4.9 were generated using the CRISPR-Cas9 system as described by [32] in an N2 VC2010 background [1]. The 20 bp guide sequence for bgnt-1.1 was designed to include a 3’GG motif, as guides with GG at the 3’ end are purported to give higher integration efficiency [45]. 500 bp homology arms (ordered as gBlocks from IDT) were designed to flank exons 1 and 2 of bgnt-1.1. The homology arms were inserted into a Pmyo-2::GFP-neoR-loxP disruption/deletion vector (provided by the Calarco Lab) using Gibson Assembly. The guide sequence and homology arms sequences are available in S7 Table and S8 Table, respectively.
Of the three null mutations, gk3637 was generated using purified Cas9 protein, while gk3638 and gk3639 were generated using a plasmid-encoded version of the protein. The Cas9 protein was prepared according to the procedure described in [46] Paix et al. (2015). The protein injection mix was assembled as described in [46] and used tracrRNA and bgnt-1.1 crRNA ordered from IDT.
Putative integrants were validated by generating PCR amplicons spanning the junction between genomic DNA and the inserted cassette. Primer F01D4.9-1-L was used in conjunction with primer pMyo-2-SEC to validate the region just upstream of the putative insertion. This generated a 1075 bp product that covers genomic DNA as well as a region within the insertion. Primer F01D4.9-1-R was used in conjunction with primer NeoR-SEC to validate the region just downstream of the putative insertion. This generated a 1632 bp product that covers genomic DNA as well as a region within the insertion. Sanger sequencing of the PCR amplicons was conducted by the Nucleic Acid Protein Service Unit (NAPS, UBC).
To rough-map the dye-filling defects of Million Mutation Project strains to an arm of a chromosome we used the high-throughput SNP mapping approach created by Davis et al. The following SNPs used by Davis et al. [47] were omitted from our analysis because the whole genome sequence data from Thompson et al. [1] could not safely deduce that the SNPs from parental strain subjected to mutagenesis, VC2010 (from which the Million Mutation Project strains were generated), matched those of Bristol N2 but not Hawaiin CB4856 (mapping strain): W03D8, F58D5, T01D1, Y6D1A, Y38E10A, T12B5, R10D12, F11A1, and T24C2.
Dye-filling assays were performed using the fluorescent dye DiI (Molecular Probes; DiIC18 Vybrant DiI cell-labelling solution, diluted 1:1000 with M9 buffer). Mixed stage C. elegans cultures were stained for 30 minutes, and Dil uptake into the amphid and phasmid neurons was visualised using either a Zeiss fluorescent dissection scope (dye-filling screen) or spinning disc confocal microscope (WaveFX spinning disc confocal system from Quorum Technologies) using a 25X oil (N.A 0.8) objective and Hammamatsu 9100 EMCCD camera. Volocity software (PerkinElmer) was used for acquisition. The completely dye-filling defective (dyf) mutant strain PR813 osm-5(p813) was used as a positive control for the dye-filling phenotype.
For the dye-filling screen, two plates of mixed-stage C. elegans were dye-filled for each Million Mutation Project strain, and defects were quantified by counting the number of worms exhibiting amphid and/or phasmid dye-filling defects. A worm was classified to have a dye-filling defect if: i) no fluorescence was observed, ii) fluorescence was observed to be greatly reduced (minimum of an estimated 3x fluorescence reduction compared to wild-type staining from the experiment at the same magnification and laser intensity) and/or iii) fluorescence staining pattern was abrogated (e.g., accumulations of fluorescence at tips of dendrites with little to no staining in cell bodies). Fifteen worms were scored from each plate. If the dye-filling of a Million Mutation Project strain appeared qualitatively dimmer than wild-type worms across both plates or if ≥ 25% of the population exhibited a dye-filling defect the assay was repeated for that strain. A Fisher’s exact test followed by p-value adjustment using false discovery rate of 5% (Benjamini–Hochberg procedure) was used to if they exhibited a significant dye-fill defect compared to wild-type (N2). This was done separately for both amphids and phasmids.
For visualisation of fluorescent-tagged proteins, worms were immobilised in 1μl of 25mM levamisole and 1μl of 0.1μm diameter polystyrene microspheres (Polysciences 00876–15, 2.5% w/v suspension) on 10% agarose pads and visualised under a spinning disc confocal microscope (WaveFX spinning disc confocal system from Quorum Technologies) using a 100X oil (N.A 1.4) objective and Hammamatsu 9100 EMCCD camera. Volocity 6.3 was used to deconvolve images as well as measure ADL cilia length and distal tip of ADL cilia to distal end of amphid socket cell length. The researcher was blind while performing the quantisation of ADL cilia/dendrite phenotypes.
Worms were permeabilised, fixed and stained according to standard methods [48]. To mark the cis-Golgi, two anti-SQL-1 antibodies, one directed against the N terminus of SQL-1 and one affinity purified antibody against the C terminus of SQL-1, were used. These antibodies have been characterised previously [33]. Both were visualised with secondary goat-anti rabbit Alexa 594 (Molecular Probes, Eugene, OR; 1:800). Localisation of BGNT-1::GFP and SQL-1 was imaged using a SpinD1454 Roper/Nikon spinning disk microscope with a 100x objective.
We performed SKAT using the SKAT package (version 1.0.9) [2] in R (version 3.2.4). No covariates were used. Given that the MMP library was created via random mutagenesis of the same isogenic parental strain [1] we did not have to control for population stratification. We chose to perform SKAT using a linear regression framework to take full advantage of the proportion data we had collected, as the logistic regression framework for SKAT only allows for a dichotomous response variable. To apply a linear regression framework to our proportion data we added a small constant to all the data points for the response variable, and then log transformed them. We used probability plots to choose the best constant (S2 & S3 Figs), and thus used a constant of 0.005 for the amphid phenotype data, and a constant of 0.05 for the phasmid phenotype data.
Custom, biologically relevant weights were assigned to the variants. Nonsense, splicing mutations and frameshift causing deletions were assigned a weight of 1, in-frame deletions were assigned a weight of 0.75, and missense mutations were assigned a weight of 0.25. Gene-based tests for all genes with a minor allele count > 6 were performed. A false discovery rate (Benjamini-Hochberg procedure) of 5% was used to determine genes which were significantly associated with the phenotype. Make, Bash, Perl and R scripts used to perform the analysis can be found at: https://github.com/ttimbers/Million-Mutation-Project-dye-filling-SKAT.git
To estimate power and recommend a minimum sample size for future experiments we performed a bootstrap power analysis using the amphid dataset. To do this, we randomly sampled (resampling = FALSE) N strains from the dataset we collected, and performed the SKAT analysis presented in this paper. We did this 100 times for N = 50, 100, 200, 300 and 400. We then estimated power as the proportion of times we observed a gene to be significantly associated with the phenotype. This was also done for two, three, four and five genes. The code used to perform this analysis can also be found in the Github repository for this study: https://github.com/ttimbers/Million-Mutation-Project-dye-filling-SKAT.git
Protein sequences (obtained from: http://www.cazy.org/) were aligned using MUSCLE 3.7 [49]. The phylogenetic tree was built using PhyML 3.0 aLRT [50] and viewed using FigTree version 1.3.1 (http://tree.bio.ed.ac.uk/software/figtree/).
|
10.1371/journal.pbio.2000779 | Alteration of protein function by a silent polymorphism linked to tRNA abundance | Synonymous single nucleotide polymorphisms (sSNPs) are considered neutral for protein function, as by definition they exchange only codons, not amino acids. We identified an sSNP that modifies the local translation speed of the cystic fibrosis transmembrane conductance regulator (CFTR), leading to detrimental changes to protein stability and function. This sSNP introduces a codon pairing to a low-abundance tRNA that is particularly rare in human bronchial epithelia, but not in other human tissues, suggesting tissue-specific effects of this sSNP. Up-regulation of the tRNA cognate to the mutated codon counteracts the effects of the sSNP and rescues protein conformation and function. Our results highlight the wide-ranging impact of sSNPs, which invert the programmed local speed of mRNA translation and provide direct evidence for the central role of cellular tRNA levels in mediating the actions of sSNPs in a tissue-specific manner.
| Synonymous single nucleotide polymorphisms (sSNPs) occur at high frequency in the human genome and are associated with ~50 diseases in humans; the responsible molecular mechanisms remain enigmatic. Here, we investigate the impact of the common sSNP, T2562G, on cystic fibrosis transmembrane conductance regulator (CFTR). Although this sSNP, by itself, does not cause cystic fibrosis (CF), it is prevalent in patients with CFTR-related disorders. T2562G sSNP modifies the local translation speed at the Thr854 codon, leading to changes in CFTR stability and channel function. This sSNP introduces a codon pairing to a low-abundance tRNA, which is particularly rare in human bronchial epithelia, but not in other human tissues, suggesting a tissue-specific effect of this sSNP. Enhancement of the cellular concentration of the tRNA cognate to the mutant ACG codon rescues the stability and conduction defects of T2562G-CFTR. These findings reveal an unanticipated mechanism—inverting the programmed local speed of mRNA translation in a tRNA-dependent manner—for sSNP-associated diseases.
| Synonymous single nucleotide polymorphisms (sSNPs) in protein-coding regions occur at much higher frequency in the human genome [1] than initially assumed. Owing to the degeneracy of the genetic code (that is more than 1 codon specifying 1 amino acid), sSNPs are considered silent or invariant for protein folding and function as they synonymously exchange only codons, but not the encoded amino acids. As a corollary of this view, sSNPs have been rationalized as neutral for selection and fitness of an organism [2]. However, synonymous codons of an amino acid are not equally used and the bias in codon usage suggests that synonymous codons have been under evolutionary pressure [3]. Natural selection of rarely used codons conditions circadian rhythm–dependent gene expression [4, 5] and synchronizes mRNA translation with downsteam processes, including protein folding and translocation [6–8]. Thus, sSNPs that alter codon usage might affect cotranslational protein folding and protein conformation. Furthermore, sSNPs might alter local mRNA secondary structure [9], binding sites for RNA-binding proteins or regulatory miRNAs, thus also impacting physiological function [10]. So far, the little experimental evidence in eukaryotic systems linking sSNPs with conformational changes in proteins [11, 12] lacks mechanistic explanation.
Genome-wide association studies link sSNPs with ~50 diseases in humans, highlighting the wide-ranging impact of SNPs that exchange synonymous codons; however, their association with alterations in protein conformation and function remains elusive [10]. Conceptually, codon usage is considered as a proxy for the speed each codon is translated: rare codons are more slowly translated than abundant codons [13]. However, global analysis of ribosome progression along mRNAs in mammalian cells argues that rare codons do not always have an effect on translation speed [14, 15]. A major determinant of elongation speed for a codon is the concentration of its cognate tRNA [16] and the ratio of cognate to near-cognate tRNA [17]. In prokaryotes and unicellular eukaryotes, tRNA concentration correlates well with codon usage [18]. By contrast, in mammalian systems, tRNA concentration varies between proliferating and differentiating cells [19], among organs [20], and strikingly, some tRNAs are uniquely expressed in defined subregions of one organ [21]. These data argue that the rate of translation of a codon cannot be determined simply from genome usage. They also suggest that in multicellular eukaryotes, codon translation speed might vary among different cells and tissues despite uniform genomic codon usage. Hence, the effects of sSNPs might be restricted to specific tissues and are not predictable from codon usage. Consequently, some sSNPs might have so far escaped detection, and their precise effects on protein function remain unresolved, with potential implications for human health.
Here, we investigated the impact of sSNPs on the cystic fibrosis transmembrane conductance regulator (CFTR). CFTR is an ATP-binding cassette (ABC) transporter that functions as a ligand-gated anion channel [22]. Dysfunction of CFTR causes the common, life-shortening disease cystic fibrosis (CF) [23]. Although the ΔF508 mutation is by far the most prevalent CF mutation, to date, more than 2,000 different mutations have been found in the CFTR gene, which vary in disease severity and penetrance [23]. Despite major advances in classifying mutations within the CFTR gene [24], the mechanisms of dysfunction of the large majority, including almost 270 SNPs (comprising both synonymous and nonsynonymous nucleotide substitutions), remain enigmatic. We integrated global analyses of tRNA concentration and translation with (i) thermal stability and proteolytic susceptibility assays as a reporter of CFTR conformation and (ii) single-molecule activity measurements as a readout of CFTR function to comprehensively determine the effects of sSNPs on CFTR expression and function. We identified an sSNP that inverts the programmed local speed of mRNA translation in a tissue-specific, tRNA-dependent manner to alter the global conformational dynamics and physiological function of CFTR.
From all ~270 polymorphic mutations in the CFTR gene (http://www.genet.sickkids.on.ca/app), we selected only synonymous SNPs in the CFTR coding sequence (S1A Fig and S1 Table). The sSNPs were broadly examined for their influence on steady-state protein and mRNA levels in the CF bronchial epithelial cell line (CFBE41o-) and HeLa cells. The mutations G1584A, G2280A, T3339C, and A3870G frequently showed reduced mRNA levels in both HeLa and CFBE41o- cells (Fig 1A), yet the total protein level demonstrated a cell-specific pattern and compared to wild-type CFTR was reduced mostly in HeLa cells, but not in CFBE41o- cells (Fig 1B and S1B Fig). Interestingly, a cell-specific pattern of mRNA steady-state expression was observed for ΔF508 (Fig 1A), but the protein level remained equally low in both cell lines (Fig 1B). Of note, for the T2562G mutation, we detected unusual behavior. It significantly reduced the total protein expression (the sum of B and C bands) by 25%–30% in both HeLa and CFBE41o- cells compared to that of wild-type CFTR (Fig 1B), while the mRNA level remained unchanged (Fig 1A). Like wild-type CFTR [25], T2562G-CFTR consisted of both core-glycosylated (immature, endoplasmic reticulum (ER)-resident, band B) and complex-glycosylated (mature, Golgi processed, band C) forms (S1B Fig). However, the ratio of C to B bands for T2562G-CFTR remained the same as that of wild-type CFTR (S1B Fig), implying a proportional reduction of both the mature form (band C) and the immature ER-resident form (band B). In well-differentiated human CF airway epithelia, T2562G-CFTR localized either at the apical membrane or intracellularly, whereas wild-type CFTR was localized only at the apical membrane and ΔF508-CFTR was retained intracellularly (Fig 1C).
Some sSNPs may alter CFTR mRNA splicing by changing the regulatory motifs of exonic splice enhancers [26]. Although we work exclusively with cDNA, which considers only full-length CFTR transcripts, we sought to address whether in the pre-mRNA of native CFTR the T2562G sSNP might alter splicing to reduce mRNA levels. Hence, we analyzed the effects of T2562G on mRNA splicing by using the minigene approach that was established to assess alternative splicing variants [27, 28]. The T2562G mutation is located in exon 15 of CFTR. Using highly sensitive on-chip capillary electrophoresis detection, which sensitively detects even background splicing in wild-type CFTR, we studied a minigene spanning exon 15 and parts of its flanking introns (S1C–S1E Fig). Of note, the T2562G sSNP did not induce any detectable alternative splicing product above the background of wild-type CFTR (S1C–S1E Fig). This result is consistent with previous observations, which report no changes in the mRNA splicing pattern of CF patients [29], and alternatively spliced products are only detected when T2562G occurs with other (silent) mutations [26, 28].
The reduced protein expression and altered localization of T2562G-CFTR when compared to wild-type CFTR raised the intriguing possibility that T2562G might trigger an alternative channel conformation, which renders T2562G-CFTR a better client for quality control machinery. We therefore investigated the degradation of T2562G-CFTR protein by the proteasome. T2562G-CFTR showed modest, but significantly enhanced ubiquitination and higher susceptibility to proteasomal degradation compared to wild-type CFTR (S2A–S2D Fig). Two E3-ubiquitin–ligating enzymes, the cytosolic C-terminus of Hsc70-interacting protein (CHIP) and the ER-membrane–bound RING domain protein RMA1 (also known as RING Finger Protein 5 [RNF5]), participate in the surveillance of CFTR biogenesis [30]. While RMA1 recognizes folding defects during or soon after translation of both wild-type and ΔF508-CFTR, CHIP inspects their folding status at a later time point, at least after synthesis of nucleotide-binding domain 2 (NBD2) [31]. RMA1 and CHIP both displayed increased binding to T2562G-CFTR than to wild-type CFTR (S2E–S2G Fig), corroborating the observed reduction of the amount of T2562G-CFTR protein (Fig 1B). Using pulse-chase experiments in HeLa cells, we compared the maturation efficiency of wild-type and T2562G-CFTR by monitoring the conversion of newly synthesized immature core-glycosylated band B to mature complex-glycosylated band C. The conversion of the band B to band C was not significantly different between T2562G-CFTR and wild-type CFTR (S2H and S2I Fig). To test whether lower steady-state protein levels of T2562G-CFTR (Fig 1B) might result also from plasma membrane instability, we measured the stability of membrane-localized CFTR by using biotinylation of nonpermeabilized cells [30]. The stability of plasma membrane–localized wild-type and T2562G-CFTR was comparable at early time points (S2J Fig). However, some enhancement of T2562G-CFTR plasma membrane stability was apparent at 24 h (S2J Fig), which might, in part, offset its proteasomal degradation.
Together, the enhanced ubiquitination and binding of T2562G-CFTR to CHIP and RMA1, albeit modest in magnitude, suggest that T2562G induces subtle alterations in CFTR structure that are detected by quality-control machinery. These slight changes escape detection in kinetic experiments (i.e., pulse-chase assays). However, over time, these small differences in CFTR structure and cell-surface stability might accumulate and be detected in steady-state experiments. Thus, to further evaluate structural rearrangements in T2562G-CFTR, we used limited proteolysis and thermal aggregation assays [32]. Lysates of cells expressing CFTR variants were exposed to different temperatures and the fraction of membrane-bound, nonaggregated CFTR was determined (Fig 2A and 2B). As a measure of the thermal aggregation propensity of the CFTR protein, the aggregation temperature (Ta) was defined as the temperature at which 50% of CFTR protein remained membrane-soluble [32]. T2562G-CFTR displayed a higher Ta (73.1 ± 1.1°C; mean ± SEM) than wild-type CFTR (66.6 ± 2.2°C) (Fig 2A and 2B). Interestingly, the magnitude of thermal stabilization achieved by the T2562G mutation (~6°C) was equivalent to the destabilizing effect of the ΔF508 mutation (Ta = 60.9 ± 1.4°C) (Fig 2A and 2B). When compared with wild-type CFTR, T2562G-CFTR reproducibly displayed subtle structural differences: the T2562G sSNP decreased slightly the susceptibility of CFTR to limited proteolysis (Fig 2C and 2D and S3 Fig). Of note, the similarity of the proteolytic patterns of wild-type and T2562G-CFTR (Fig 2C and S3A Fig) argues against large structural rearrangements and instead suggests that the T2562G sSNP causes local conformational changes. Consistent with this idea, we detected no discernible differences in the proteolytic susceptibility and thermal stability of band B, only of band C (Fig 2B, bottom panel and Fig 2D). This suggests that the effects of T2562G are on the overall topology of CFTR.
To address the impact of the T2562G sSNP on CFTR function as a regulated Cl- channel, we studied individual CFTR Cl- channels in excised inside-out membrane patches with the patch-clamp technique. For 2 reasons, we did not study macroscopic CFTR Cl- currents to learn how the T2562G sSNP influences CFTR function: First, studies of macroscopic CFTR Cl- currents do not distinguish between mutation effects on CFTR expression (i.e., channel number) and CFTR function (i.e., conductance and gating). Second, we reasoned that the effects of an sSNP on CFTR function would likely be slight and therefore not easily resolved when studying macroscopic CFTR Cl- currents. We anticipated that high-resolution single-channel recording would be required to discern the impact of the T2562G sSNP on CFTR function.
Once phosphorylated by protein kinase A (PKA), wild-type human CFTR forms a low-conductance, Cl-selective channel regulated by cycles of ATP binding and hydrolysis [22, 33]. In single-channel recordings, wild-type CFTR exhibits a bursting pattern of channel gating with channel openings interrupted by brief, flickery closures, separated by longer closures between bursts (Fig 3A). Strikingly, 2 populations of T2562G-CFTR channels were distinguished following PKA-dependent phosphorylation (Fig 3A, S4 Fig and S2 Table). The first population of T2562G-CFTR channels (wild-type–like [wtl] population) exhibited characteristics identical to those of wild-type CFTR (Fig 3 and S5 Fig). Wild-type–like openings of T2562G-CFTR had the same current amplitude and conductance as wild-type CFTR (Fig 3B and S5 Fig), demonstrating that the channel pore was unaltered from wild-type CFTR. Similarly, wtl openings of T2562G-CFTR exhibited a gating pattern with an identical mean burst duration (MBD) (the average duration of channel openings), interburst interval (IBI) (the average duration of long channel closures) and open probability (Po) (a measure of single-channel activity) as wild-type CFTR (Fig 3C–3E). These data argue that the regulation of wtl openings of T2562G-CFTR by intracellular ATP is the same as wild-type CFTR. By contrast, the second population of T2562G-CFTR channels (small-conductance [sc] population) was distinguished by the small size of channel openings, with a single-channel conductance about half that of wild-type CFTR and a small, but significant reduction in Po (Fig 3 and S5 Fig). We interpret these results to suggest that the sc population of T2562G-CFTR Cl- channels has a constricted pore for Cl- flow and ATP is slightly less efficacious in stimulating channel gating. Of note, sc openings of wild-type CFTR are very rare events and they are distinct from those of T2562G-CFTR (S2 Table); they are also different to the brief sojourns to subconductance states observed with wild-type human CFTR [34].
Several lines of evidence suggest that the sc and wtl channels of T2562G-CFTR are independent channel populations. First, both wtl and sc T2562G-CFTR channels were observed together (Fig 3A, S4–S6 Figs and S2 Table) or by themselves (S4B and S4C Fig and S2 Table). Second, the occurrence of membrane patches with only 1 type of active channel or with dissimilar numbers of wtl and sc channels argues against the occurrence of multiple conductance states [35], mode switching [36], or coupled transitions [37] (S4B and S4C Fig and S2 Table). Third, binomial analysis of membrane patches with 1 sc and 1 wtl channel demonstrated that experimentally measured values of the 3 conductance levels—P(0), P(1), and P(2)—differed from their predicted values (S6 Fig), implying that sc and wtl channels are either dependent on each other or are nonidentical channels. To distinguish between these possibilities, we measured the cooperativity ratio (CR) [38]. In membrane patches with 1 sc and 1 wtl channel, the CR was 2 ± 0.1 (n = 3), implying that sc and wtl channels are independent, nonidentical channels. Thus, the subtle structural changes caused by the T2562G sSNP, which increase CFTR stability, are detrimental to channel function.
How might a synonymous mutation alter protein conformation and function? The T2562G sSNP exchanges the Thr854-ACT codon for a Thr854-ACG triplet (i.e., exchanging a T for G nucleotide) (Fig 4A). A/T nucleotides have lower propensity to partition in secondary interactions than G/C nucleotides. The unaltered mRNA levels of the T2562G-CFTR (Fig 1A) argue against any significant global changes in the mRNA secondary structure and stability through the T2562G sSNP; nevertheless, subtle local changes might escape detection. Thus, we also exchanged the Thr854ACT triplet for its 2 other synonymous alternatives, ACC and ACA (Fig 4A). Both T2562C- and T2562A-CFTR were indistinguishable from wild-type CFTR in their mRNA and protein expression levels, thermal stability (S7 Fig), and channel function (Fig 3 and S5 Fig), suggesting that the observed effect of the T2562G sSNP is specific for the ACG codon.
All 4 Thr codons are of moderate genome usage, differing only by 2–3-fold (Fig 4A), while the difference between high- and low-abundance codons is 7.5–8.7-fold (S3 Table). Next, we used ribosome profiling to assess the relative average speed of translation of each Thr codon in CFBE41o- cells. The residence frequency of a codon in the ribosomal A site (that is, the site accepting the aminoacyl-tRNA) correlates with the ribosome dwell time at that particular codon and is proportional to the codon’s translational speed [21, 39]. We calibrated the ribosome-protected fragments on the ribosomal A site by using the 5′ ends of the sequencing reads as previously described [40]. Strikingly, the ACG codon was among the codons with the highest ribosomal occupancy, implying that it is among the most slowly translated codons in CFBE41o- cells (Fig 4B). By contrast, the ACT codon less frequently dwelt in the ribosomal A site and was translated with much higher velocity (Fig 4B).
The concentration of the cognate tRNA is one major determinant of ribosome speed at a codon. To address whether ribosome occupancy at Thr codons correlates with the concentration of their cognate tRNAs, we determined the absolute tRNA concentration in HeLa cells (Fig 4C) and related the abundance of each single tRNA to that of CFBE41o- cells and CF patient–derived primary human bronchial epithelial cells (HBEs) by using tRNA-tailored microarrays (Fig 4D, S8A Fig and S1 Text). tRNA isoacceptors (that are different tRNA species carrying the same amino acids, but with different anticodons) varied greatly in their concentrations, spanning up to an order of magnitude (Fig 4C). For example, compare tRNAPro that pairs to Pro (CCT/C/A/G) codons and tRNAThr that reads ACG codon (Fig 4C). While tRNAThr(UGU) that pairs to ACT codons is among the moderately abundant tRNAs, tRNAThr(CGU) decoding the ACG codon is one of the rarest tRNAs in HeLa (Fig 4C, both designated with arrows). Comparative microarray analysis demonstrated that the levels of the tRNAThr(CGU) decoding the ACG codon were similar (and hence, equally rare) in CFBE41o- and 4 CF patient–derived primary HBE cells (Fig 4D and S8D Fig). To verify this result, we performed northern blotting using specific probes that spanned the anticodon loop tRNAThr(CGU). No tRNAThr(CGU) signal was detected in CFBE41o- cells, CF patient–derived primary HBE cells, or pulmonary tissue from a non-CF individual, implying very low abundance, below the detection limit of the northern blot (Fig 4E). Of note, the higher ribosomal occupancy of approximately 50% at the ACG codons compared to the ACT codons (Fig 4B) mirrors the concentration difference between their cognate tRNAs (Fig 4C and 4D; see also S1 Text). Thus, the slower translation of ACG compared to ACT detected by ribosome profiling (Fig 4B) correlates with the rare abundance of its cognate tRNAThr(CGU).
Unexpectedly, we measured far higher concentrations of tRNAThr(CGU) in other human tissues, including the heart, brain, and kidney (Fig 4E). In 2 laboratory cell lines of kidney (HEK 293T) and neuronal (SH-SY5Y) origin, T2562G-CFTR expression did not differ from that of wild-type CFTR (Fig 4F), suggesting that when the cellular level of tRNAThr(CGU) is high (Fig 4A), the effect of T2562G sSNP remains dormant. Of note, the ACG codon is rarely used in CFTR to encode Thr (Fig 4A). Strikingly, among tissues that naturally express CFTR, including the pancreas, colon, and salivary glands, the amounts of all tRNAsThr are equal (Fig 4G), with tRNAThr(CGU) being the rarest among the 3 tRNAThr isoacceptors (for further information, see S1 Text). We conclude that the rare usage of the ACG codon in CFTR (Fig 4A) correlates with the very low abundance of the cognate tRNAThr(CGU) in tissues naturally expressing CFTR. The mutated ACG codon is read by a rare tRNA in the cell lines used in this study (CFBE41o-, CHO, and HeLa cells; Fig 4D, 4E and 4G). However, the levels of other tRNA isoacceptors also differ between CFBE41o- and HeLa cells (S8D Fig). As a result, the effects of other sSNP involving those differing tRNA isoacceptors might vary between CFBE41o- and HeLa.
Finally, we compared the tRNA concentration in both HeLa and CFBE41o- cells with genomic codon usage. The tRNA concentration in both cells correlated poorly with genomic codon usage (S8B and S8C Fig), suggesting that codon usage is a poor predictor of the speed of translation of each single codon. Taken together, variation in the amount of tRNAThr(CGU) measured in different human tissues argues that despite common codon usage, tRNAs modulate the effects of sSNP in a tissue-specific manner.
If the aberrant conformation and function of T2562G-CFTR is indeed a consequence of delayed translation at this sSNP by the low concentration of the cognate tRNAThr(CGU), then the phenotype should be rescued by increasing the speed of translation at this codon through elevation of cellular tRNAThr(CGU) levels. To test this idea, we transiently cotransfected T2562G-CFTR and in vitro synthesized, uncharged tRNAThr(CGU). Elevated levels of the rare tRNAThr(CGU) caused a tRNA concentration–dependent increase in T2562G-CFTR expression (Fig 5A and 5B and S9A Fig) and restored the aggregation temperature of T2562G-CFTR to that of wild-type CFTR (Fig 5C). Transfection of tRNA was without effect on translation (S9A–S9C Fig), arguing against global alteration of translation. Moreover, the transfected tRNAThr(CGU) was translationally active, as it was present also in the polysomal fraction (S9D Fig). Conversely, down-regulation of tRNAThr(CGU) using shRNA treatment noticeably decreased the T2562G-CFTR steady-state protein level (S9E and S9F Fig) with only marginal effects on global translation (S9G Fig).
Finally, we investigated whether accelerating ribosomal passage at the sSNP with tRNAThr(CGU) also ameliorated its functional defects. Elevating the tRNAThr(CGU) concentration rescued the conductance defect of the sc population of T2562G-CFTR channels (Fig 5D–5H), albeit without altering its minor gating defect. Only large-amplitude openings of T2562G-CFTR were observed, with conductance similar to that of wild-type CFTR (Fig 5D–5H and S2 Table). We conclude that the effects of the T2562G sSNP are clearly tRNA-dependent, and increasing the cellular tRNAThr(CGU) concentration rescued both the expression and single-channel conduction defects of T2562G-CFTR.
Our results demonstrate that the T2562G sSNP induces local changes in translation velocity, giving rise to more stable channels with a greatly reduced single-channel conductance (Fig 6A and 6B). The effect is clearly tRNA-dependent, as increasing the cellular concentration of tRNA cognate to the mutant codon rescues the stability and conductance defects of T2562G-CFTR (Fig 6C).
Translation rate is maximized at codons read by highly abundant tRNAs and minimized at codons with rare tRNAs. Thereby, the selection of codons translated at different speeds is not random [3] and shapes the kinetics of mRNA translation to regulate cotranslational protein folding [13, 41–43]. Consistent with this idea, substitution with slow-translating codons in fast-translating regions (and vice versa) might be incompatible with cotranslational protein folding and deleterious for protein integrity. Our results suggest that the T2562G sSNP–induced CFTR dysfunction is through inversion (i.e., slowing down) of local ribosomal speed at the mutant ACG codon. Of note, these data provide the first direct evidence for the central role of tRNA in mediating the effects of an sSNP. The T2562G sSNP exchanges the ACT codon for the ACG triplet. Despite the similar usage of ACT and ACG, which differ only by 2–3-fold (Fig 4A), the ACG triplet has a measurable effect on translation kinetics. tRNAThr(CGU) cognate to the ACG triplet is present at very low abundance in human bronchial epithelial cells, rendering the ACG codon very slowly translated in these cells. tRNAs of the same kind are not uniformly distributed within the cell, which creates inherent variation in the speed of translation of a codon. We reason that such local differences in the tRNA concentration and in general the stochastic nature of tRNA delivery to the ribosome result in production of 2 populations of T2562G-CFTR channels: one identical to wild-type (wtl) and a second with reduced (sc) single-channel conductance (Fig 6B). tRNAThr(CGU) is far more abundant in other tissues. Hence, the ACG codon will be translated with different velocities in diverse tissues, suggesting a tissue-specific effect of the T2562G sSNP, which for CFTR might be restricted only to epithelial cells.
Strikingly, codon usage within the CFTR sequence deviates greatly from human genomic codon usage, with the ACG codon being rarely used in CFTR to encode Thr (Fig 4A). The very low concentration of the cognate tRNAThr(CGU) in bronchial epithelial cells and tissues, which are a key site of CFTR expression, suggests that the ACG codon might have been under selection pressure. Consistent with this idea, wild-type CFTR has in total 3 ACG codons, all of which are located in the context of fast-translating codons pairing to highly abundant tRNAs. Thus, most likely the effects of these 3 naturally occurring ACG codons in wild-type CFTR remain dormant. By contrast, the T2562G sSNP introduces a slowly translating ACG codon in a region enriched with other slowly translating codons, emphasizing the importance of sequence context for the effects of an sSNP. It should be noted that sSNPs have diverse effects that are not always tRNA-dependent. For example, the sSNP at Ile507 of CFTR, which accompanies the ΔF508 mutation, stabilizes secondary mRNA structure in the vicinity of the mutation and most likely contributes to ΔF508-CFTR misfolding by decreasing the mRNA translation rate [44].
Folding of CFTR is cotranslational [45], guided by intensive interactions with cytosolic and ER-resident chaperones [30] and most likely orchestrated by translation kinetics [46]. Importantly, the structural coupling of independently folded CFTR domains and extensive domain–domain interface contacts maintain channel stability and function [47–50]. The ΔF508 mutation, which causes a dramatic reduction of functional plasma membrane–resident CFTR, has minimal impact on the protein backbone and folding of NBD1, but greatly compromises interdomain interfaces [47, 50]. Since the T2562G mutation causes no discernible effects on the core glycosylated form (band B), this suggests that T2562G-CFTR folds close to its native conformation. The larger effect on stability and function of the complex glycosylated form (band C) suggests that the T2562G sSNP likely affects the conformational dynamics of CFTR. The atomic structure of zebrafish (Danio rerio) CFTR identifies an N-terminal ‘lasso’ motif, which wraps around the central axis of the protein and likely regulates channel gating through interactions with the R domain [51, 52]. Although it is not possible to identify Thr854-interacting partners in the atomic structure of zebrafish CFTR [52], it is conceivable that Thr854 might interact with the lasso motif and, hence, that the T2562G sSNP might disrupt this interaction. Thr854 is located at the edge of the R domain, close to membrane-spanning domain 2 (MSD2), in a structured α-helical segment (S1A Fig); thus, it is positioned at the interface between the lasso motif and MSD2 [52]. T2562G-induced changes in Thr854 codon speed might therefore alter this α-helix, destroying interface interactions, leading to changes in CFTR conductance. Alternatively, the T2562G sSNP might impact interdomain interactions of the R domain [53] with other domains of CFTR to influence channel conductance. A further possibility is that the T2562G mutation alters R domain phosphorylation, leading to reduced CFTR expression [54]. Clearly, despite the local subtle effects of the T2562G sSNP, which are only in the vicinity of the mutation, the T2562G sSNP appears to stabilize the final 3D topology of CFTR, most likely by influencing R domain–mediated interdomain interactions, which reduce conformational dynamics and perturb channel conductance.
CFTR is one of the most polyvariant human genes and it is now recognized that CFTR mutant alleles frequently carry combinations of mutations [23, 55]. T2562G is one of the most common SNPs in the CFTR gene [56, 57] with a prevalence of 34% in the general population. Although this sSNP, by itself, does not cause CF, it is prevalent in patients with CFTR-related disorders [26, 58–60], which argues that sSNPs have the potential to epistatically modulate the effects of disease-causing mutations, thereby modifying disease severity. Building on this idea, the Ile507 sSNP along with the ΔF508-mutation alters the response of ΔF508-CFTR to small molecule CFTR correctors [61], implying the potential contribution of epistatic mutations to personalized medicine.
Translation kinetics is an integral feature of CFTR folding and tunes synthesis at critical nodes of the CFTR cotranslational folding landscape [46]. Consistent with this idea, 2 studies show that modulation of translation velocity is a robust strategy to correct folding errors in CF mutants [62, 63]: First, a global decrease of translation kinetics by varying codon choice [62] and second, knockdown of a ribosomal protein greatly increases the folding efficiency of ΔF508-CFTR [63]. In conclusion, our results demonstrate that the complex effects of the T2562G sSNP on CFTR function are most likely a result of altered local kinetics of mRNA translation. We speculate that the inversion of codon speed induced by the T2562G sSNP is applicable to other proteins in a tissue-specific fashion. SNPs introduce variability into an individual’s genome composition, which might influence disease risk, the spectrum of disease symptoms, and ultimately, therapeutic response.
Pulmonary tissues from lung biopsies were provided by Dr. Raymond Frizzell and the Health Sciences Tissue Bank (HSTB). HSTB is covered by the University of Pittsburgh IRB approval #0506140.
HeLa cells (American Type Culture Collection [ATCC] no. CRM-CCL-2), HEK 293 cells (ATCC no. CRL-1573), SH-SY5Y (German cell collection (DSMZ no. ACC209) and immortalized CFBE41o- cells (kind gifts of Karl Kunzelmann, University of Regensburg, Germany; Eric Sorscher, University of Alabama, United States; and Dieter Gruenert, University of California San Francisco, US) were maintained in Dulbecco's modified Eagle's medium (DMEM; PAN Biotech) or Earle's minimal essential medium (MEM; Biochrom), supplemented with 10% fetal calf serum (FCS; PAN Biotech) and 2 mM L-glutamine (Gibco). CHO cells (ATCC no. CCL-61) were grown in Ham’s F-12 nutrient medium supplemented with 10% fetal calf serum (both from Life Technologies). For patch-clamp experiments, HeLa or CHO cells were seeded onto glass coverslips 24 h before transfection. All cells were incubated at 37°C in a humidified atmosphere with 5% CO2. CFBE41o- cells were cultivated on collagen- and fibronectin-coated cell culture dishes. All cells were frequently checked for mycoplasma contamination (VenorGeM mycoplasma detection kit; Biochrom).
The cDNAs of wild-type CFTR, ΔF508-CFTR and CFTR sSNP variants were subcloned into the pcDNA3 vector (Life Technologies) and transfected by using either Lipofectamine LTX, Lipofectamine Plus (Life Technologies), or by using polyethylenimine (PEI, linear [25,000], Polysciences). Constructs were verified by sequencing. In patch-clamp experiments, CFTR cDNAs were cotransfected with those of GFP and then 36–60 h later, GFP-expressing cells were selected for study [64].
Human tRNA sequences were extracted from the Genomic tRNA Database (http://gtrnadb.ucsc.edu). Single-stranded DNA oligonucleotides resembling full-length tRNAs were annealed to obtain double stranded, full-length tDNA templates flanked at the 5′-end with T7 promoter sequence and at the 3′-terminus with CCA [65]. tDNA was transcribed in vitro with T7-RNA polymerase (Fermentas) in the presence of 5-mM GMP (Sigma-Aldrich) and purified by using denaturing polyacrylamide gel electrophoresis (PAGE). Thereafter, tRNAs were denatured at 95°C for 2 min and refolded by cooling to 22°C for 3 min and incubated for 5 min at 37°C prior to transfection. tRNAs were stored at –80°C until further use [65]. The full-length tRNAThr(CGU) sequence was as follows (5′ to 3′): GGCGCGGTGGCCAAGTGGTAAGGCGTCGGTCTCGTAAACCGAAGATCRCGGGTTCGAACCCCGTCCGTGCCTCCA.
tRNAs were transfected with Lipofectamine 2000 (Life Technologies) [65]. 38-ng or 150-ng tRNAs together with 600-ng plasmid DNA were incubated with 5 μl Lipofectamine 2000 in 100 μl Opti-MEM (Gibco) at 22°C for 30 min and added to subconfluent cells in 3.5-cm dishes. A detailed protocol is available at protocols.io (https://doi.org/10.17504/protocols.io.hetb3en).
CF patient–derived primary HBE cells (patient 1, ΔF508/ΔF508; patient 2, ΔF508/G551D; patients 3 and 4, ΔF508/3849 + 10kbC>T) and pulmonary tissue from a non-CF individual were kindly provided by Raymond Frizzell and Matthew Glover (University of Pittsburgh, Pittsburgh, US). Cells were isolated after informed patient consent (HSTB, University of Pittsburgh, IRB approval #0506140) and lung transplantation at the Human Airway Cell Core of the University of Pittsburgh, cultivated on transwell filters at 37°C and 5% CO2, trypsinized, pelleted, and stored in RNAlater (Ambion) at –80°C until further use. Total RNA from various human tissues was purchased from commercially available sources: brain (no. 540157), heart (no. 540165), and kidney (no. 540169) from Agilent and colon (no. 636553), pancreas (no. 636577), and salivary gland (no. 636552) from TAKARA/ClonTech.
The following antibodies were used in this study: mouse anti-CFTR NBD1 (660; dilution 1:1,000), mouse anti-CFTR NBD2 (596; dilutions 1:100 and 1:2,500), and mouse anti-CFTR NBD2 (769; dilution 1:100) (all kindly provided by John R. Riordan and Tim Jensen, University of North Carolina, Chapel Hill, US, and Cystic Fibrosis Foundation Therapeutics, Bethesda, US), anti-CFTR NBD1 (Mr. Pink), rabbit anti-neomycin phosphotransferase II (anti-NPT; dilution 1:2,000; Merck Millipore, no. H06-747), mouse anti–β-actin (anti-ACTB; dilution 1:4,000; Sigma-Aldrich, no. A228), anti–β-catenin (dilution 1:100; Zymed), goat anti–mouse-HRP (dilution 1:10,000; BioRad, no. 170–5047), mouse anti-HA (dilution 1:1,500; Covance, no. MMS-101P), goat anti–rabbit-HRP (dilution 1:3,000; BioRad, no. 170–5046), and goat anti-mouse and goat anti-rabbit Alexa fluor-conjugated secondary antibodies (1:1,000; Molecular Probes, Inc.).
CFTR Cl- channels were recorded in excised inside-out membrane patches by using an Axopatch 200B patch-clamp amplifier and pCLAMP software (both from Molecular Devices) as described previously [64, 66]. The pipette (extracellular) solution contained: 140 mM N-methyl-D-glucamine (NMDG), 140 mM aspartic acid, 5 mM CaCl2, 2 mM MgSO4, and 10 mM N-tris[hydroxymethyl]methyl-2-aminoethanesulfonic acid (TES), adjusted to pH 7.3 with Tris ([Cl-], 10 mM). The bath (intracellular) solution contained: 140 mM NMDG, 3 mM MgCl2, 1 mM CsEGTA, and 10 mM TES, adjusted to pH 7.3 with HCl ([Cl-], 147 mM; free [Ca2+], < 10−8 M) and was maintained at 37°C. A large Cl- concentration gradient was imposed across the membrane patch (internal [Cl-] = 147 mM; external [Cl-] = 10 mM) and voltage was clamped at –50 mV to enhance the amplitude of CFTR channel openings.
CFTR Cl- channels were activated promptly following membrane patch excision by the addition of the catalytic subunit of protein kinase A (PKA [purified from bovine heart], 75 nM; Calbiochem) and ATP (1 mM; Sigma-Aldrich) to the intracellular solution. To prevent channel rundown, PKA and ATP were added to all intracellular solutions. In this study, membrane patches contained ≤5 active channels, determined by using the maximum number of simultaneous channel openings as described previously [66].
We recorded, filtered, and digitized data as described previously [64], but additionally digitally filtered small-conductance T2562G-CFTR openings at 50 Hz prior to analysis. To measure single-channel current amplitude (i), Gaussian distributions were fit to current amplitude histograms. For open probability (Po) and burst analyses, lists of open and closed times were created by using a half-amplitude crossing criterion for event detection. For wild-type CFTR, T2562C, and wtl openings of T2562G-CFTR, transitions <1 ms in duration were excluded from the analysis, whereas for sc openings of T2562G-CFTR, transitions <4 ms in duration were excluded. Dwell time histograms were plotted with logarithmic x axes with 10 bins decade-1, and the maximum likelihood method was used to fit 1- or 2-component exponential functions to the data. Burst analysis was performed as described by Cai et al. [66] by using a tc (the time that separates interburst closures from intraburst closures), which was determined from analyses of closed time histograms. The mean interburst interval (TIBI) was calculated by using the following equation [66]:
Po=Tb(TMBD+TIBI)
(1)
where Tb = (mean burst duration) x (open probability within a burst). Mean burst duration (TMBD) and open probability within a burst (Po(burst)) were determined directly from experimental data by using pCLAMP software. Burst analysis was not performed on small-conductance T2562G-CFTR openings because of their small size. Only membrane patches that contained a single active CFTR Cl- channel were used for burst analyses.
To investigate whether wtl and sc T2562G-CFTR channels are distinct channels that do not interact with each other, we performed a binomial analysis of single-channel data. Assuming that wtl and sc channels behave independently, in a membrane patch with N channels (each with an open probability of p), the probability of k channels being simultaneously open (P(k)) follows the binomial distribution:
P(k)=N!k!(N−k)!pk(1−p)N−k
(2)
Because of the small single-channel conductance of sc channels, we selected for analysis excised inside-out membrane patches with 1 active wtl channel and 1 active sc channel.
To perform binomial analysis on wtl channels and sc channels that exhibit different Po, we adopted the approach of Manivannan et al. [67] for cooperativity in 2-channel current amplitude histograms:
P(0)=(1−p1)(1−p2)
(3)
P(1)=p1(1−p2)+p2(1−p1)
(4)
P(2)=p1p2
(5)
where P(0), P(1) and P(2) are the probability of channels residing in the closed state (L0), open level 1 (L1), and open level 2 (L2), respectively (S6 Fig). p1 is the open probability of sc channels and p2 is the open probability of wtl channels. In practice, values of p1 and p2 were acquired from excised inside-out membrane patches containing only a single active channel by measuring open and closed times as described above. Using these values of p1 and p2, predicted values of P(0), P(1), and P(2) were determined by using Eqs 3–5. These predicted values were then compared with experimental values of P(0), P(1), and P(2), determined from the fit of Gaussian functions to single-channel current amplitude histograms (S6D Fig). If the predicted and experimental values of P(0), P(1), and P(2) diverge from each other, the sc and wtl channels exhibit dependency (i.e., their gating behaviour demonstrates cooperativity) or these channels are nonidentical (i.e., they have different Po values).
To distinguish between dependent and nonidentical channels, we calculated the cooperativity ratio (CR) [38]:
CR=(P(1)P(1))(P(0)P(2))/(2N)(N−1)
(6)
where N is the maximum number of active channels in the excised membrane patch (i.e., N = 2 for membrane patches with 1 wtl channel and 1 sc channel). When N = 2, if CR = 1, the channels are independent and identical; if CR > 1, the channels are nonidentical (i.e., the channels have unequal Po) and if CR < 1, the channels exhibit cooperativity [38].
Well-differentiated primary cultures of CF bronchial epithelia (genotype: ΔF508/ΔF508) cultured at an air–liquid interface were infected with recombinant adenovirus serotype 5 at a multiplicity of infection of 100 for 1 h. The vectors encoded the cDNAs of CFTR constructs driven by a cytomegalovirus promotor. Four days after gene transfer, epithelia were examined by immunocytochemistry. Epithelia were fixed with 4% paraformaldehyde (Electron Microscopy Sciences), permeabilized with 0.3% Triton X-100 (Thermo Fisher Scientific) and blocked with 10% normal goat serum (Jackson Immunologicals) in SuperBlock (Thermo Fisher Scientific). The epithelia were incubated with the anti-CFTR antibodies 769 and 596 (1:100) and anti–β-catenin primary antibody (1:100), followed by Alexa Fluor–conjugated secondary antibodies (Molecular Probes).
Cells were lysed for 15 min on ice in MNT buffer (20 mM MES, 100 mM NaCl, 30 mM Tris-HCl, pH 7.5, supplemented with 1% Triton X-100 and 1x complete protease inhibitor [Roche]) and centrifuged for 5 min at 14,000 x g (4°C) to remove cell debris. Equal amounts of lysates were mixed with SDS-loading buffer, incubated for 10 min at 37°C, separated by SDS-PAGE, and blotted onto a PVDF membrane (Millipore). Western blots were probed with the corresponding primary antibodies at 4°C overnight, followed by detection with HRP-labeled secondary antibodies and visualized by using ECL and the FujiFilm Las-4000 system (GE Healthcare). CFTR protein band intensities were normalized to the expression level of (i) NPT, which is also expressed as a selection marker on the same pcDNA3 plasmid, and (ii) ACTB, to account for differences in transfection efficiency and sample loading on the gel, respectively. The linear range for CFTR and NPT western blot assays ranged from 6.25 μg to 100 μg total protein lysate with R2 of 0.9418 and 0.9475, respectively. Data analysis was performed by using Prism 5 (GraphPad Software Inc.) software.
The plasma membrane stability of T2562G-CFTR was compared to that of wild-type CFTR by using a procedure conceptually similar to the protocol described in [30]. CFBE41o- cells transfected with either wild-type or T2562G-CFTR were seeded into 6-well dishes (600,000 cells per well for each time point). At 70% confluency, 100 μg/ml cycloheximide (CHX) was added to all wells to inhibit de novo CFTR synthesis and cells were further incubated at 37°C to evaluate the plasma membrane stability of CFTR. At each time point, a well of cells was labeled at 4°C for 1 h with noncleavable biotin reagent (EZ-Link-Sulpho-NHS-S-S-biotin, Thermo Fisher Scientific), which does not penetrate the cell and attaches to amino groups located at the membrane surface. Cells were lysed in MNT buffer, and the biotin-labeled entities were immunoprecipitated with Streptavidin-sepharose bead conjugate (Cell Signaling) and washed twice with 10 mM Tris-HCl pH 8.5, containing 300 mM NaCl, 0.05% Triton X-100, and 0.1% SDS. The beads were boiled in SDS-loading buffer, and spotted onto PDFV membrane (Millipore) through a slot–blot manifold. CFTR-positive biotinylated conjugates were detected by using mouse anti-CFTR NBD1 antibodies. As a control, an aliquot of the beads was immunostained with β-actin antibodies; it was empty, showing the complete lysis of the cells and no retention of whole cells. The change in plasma membrane stability was determined by the time-dependent decrease of membrane localized CFTR-positive signal in the time course of incubation at 37°C and was compared to the zero time point at which de novo CFTR synthesis was inhibited.
Thermoaggregation assays were performed as described previously [32]. Briefly, cells transiently transfected with different CFTR variants were lysed in MNT buffer (supplemented with 1% Triton X-100 and 1x complete protease inhibitor) 24 h after transfection. Equal amounts of cleared lysates were mixed with SDS-loading buffer and incubated for 10 min at different temperatures ranging from 37°C to 100°C. Aggregates were removed by centrifugation at 17,000 x g (4°C) for 5 min, and the remaining CFTR protein was analyzed by immunoblotting with anti-CFTR NBD2 antibody (596). The intensities of the B- and C-bands were quantified from immunoblots and fitted with the Boltzmann Sigmoidal equation by using Prism 5 software. Thermal aggregation temperature (Ta) was defined as the temperature at which 50% of the protein remained soluble [32].
Total RNA was extracted by using TRI Reagent (Sigma-Aldrich) according to the manufacturer's protocol. RNA concentration was measured by using the NanoDrop photometer (PEQLAB Biotechnology), and RNA integrity was checked with the 2100 Bioanalyzer and RNA6000Nano Chips (both Agilent) or by denaturing agarose gel electrophoresis. RNA was stored at –80°C for further use.
Steady-state mRNA levels were determined by qRT-PCR. Total RNA was pretreated with DNase I (Fermentas) and reverse transcribed by using oligo-(dT)18 primers and Revert Aid H Minus M-MuLV Reverse Transcriptase (Fermentas). Amplification was performed in clear 96-well plates (Sarstedt) sealed with adhesive tape (Sarstedt) in a Mx3005P qPCR cycler (Agilent) by using QuantiFast SYBR Green PCR master mix (Qiagen) containing 6-carboxyl-X-rhodamine (ROX) as reference dye. CFTR and NPT were amplified with the following primer pairs (5′ to 3′): CFTR forward, CCTATGTCAACCCTCAACACG and CFTR reverse, ACTATCACTGGCACTGTTGC; NPT forward, TGCTCCTGCCGAGAAAGTAT and NPT reverse, GCTCTTCGTCCAGATCATCC. Primers were used at a final concentration of 300 nM. Relative expression levels were calculated by using the ΔΔCT-method and normalized to NPT signals. Each reaction was performed in duplicate with the following controls included in each run: a no-template control (NTC) and a not-reverse-transcribed sample. qRT-PCR assays for CFTR and NPT displayed a linear range over 6 (R2, 0.9952) and 5 (R2, 0.9944) orders of magnitude, respectively. qPCR efficiencies were 99% (slope, 3.345) and 108% (slope, 3.145) for CFTR and NPT, respectively. Data analysis was performed by using MxPro QPCR (Agilent) and Prism 5 software.
tRNAThr(CGU) levels were analyzed by using the QuantiFast SYBR Green RT-PCR kit (Qiagen) containing ROX reference dye. tRNAThr(CGU) and 5S rRNA were amplified by using the following primer pair (5′ to 3′) at a final concentration of 1 μM: tRNAThr(CGU) forward, GGCCAAGTGGTAAGGC and tRNAThr(CGU) reverse, AGGCACGGACGGG; 5S rRNA forward, CCATACCACCCTGAACGC and 5S rRNA reverse, GTATTCCCAGGCGGTCTC. tRNA signals were normalized to 5S rRNA values.
For northern blotting, equal amounts of total RNA were separated by denaturing PAGE, in some cases stained with SYBR gold (Sigma-Aldrich) for visualization and subsequently transferred onto HyBond-N+ membrane (GE Healthcare). tRNAThr(CGU) was detected by overnight hybridization at 60°C with Cy3-labeled probes complementary to the tRNA anticodon loop (5′-Cy3-CTTCGGTTTACGAGACCGACGCCTTA-3′). Blots were subsequently washed at 35°C 3 times with 6x SSC (supplemented with 0.1% SDS), followed by 1 wash with 6x SSC, 1 wash with 2x SSC, and a final wash with 0.2x SSC and imaged on the FujiFilm Las-4000 system (GE Healthcare). Intensities were normalized to 5S rRNA probed also with 5′-Cy3-labeled oligonucleotide (5′-AAGTACTAACCGCGCCCGAC-3′).
tRNA microarrays [68] were performed with tRNA probes covering the full-length sequence of 41 cytoplasmic tRNA species complementary to 49 nuclear-encoding tRNA families with sequences described previously [20]. Each microarray consisted of 12 identical blocks, each containing 2 probes for each tRNA (i.e., in total 24 measurements for each tRNA). Total RNA was extracted from the cells by using TRI Reagent and deacetylated for 45 min at 37°C with 100 mM Tris-HCl (pH 9.0). Fluorescent stem-loop RNA/DNA oligonucleotide [20] bearing a Cy3 or Atto647 fluorescent dye (Microsynth) was ligated overnight at 16°C with T4 DNA ligase (NEB) to all deacetylated tRNAs [68]. Ligation efficiency was analyzed by resolving the samples with denaturing 10% PAGE and detected by fluorescence (Fujifilm LAS-4000) and SYBR gold (Invitrogen) staining. Fluorescently labeled tRNAs were hybridized on the microarrays for 16 h at 60°C in the Hyb4 microarray hybridization system (Digilab). Subsequently, the microarrays were washed once in 2x SSC/0.1% SDS (50°C), once in 1x SSC/0.1% SDS (42°C) and then 3 times in 0.1x SSC (42°C) before scanning by using a GenPIX 4200A (Molecular Devices) scanner [68]. Analysis of the scanned arrays was performed by using the GenPix Pro 7 (Molecular Devices) software. For normalization, identical amounts of in vitro–synthesized tRNA standards (i.e., Escherichia coli tRNALys(UUU), E. coli tRNATyr(AUA), and Saccharomyces cerevisiae tRNAPhe(GAA)) were added to each total RNA sample prior to deacetylation. Quantification was performed by normalizing the median of the Cy3-tRNA signal of each tRNA species to the corresponding Atto647-labeled HEK tRNA signal. Note that the human genome lacks a gene for tRNAThr(GGU) reading the Thr codon ACC (http://gtrnadb.ucsc.edu), which is likely to be decoded by tRNAThr(IGU) via deamination of adenine (A) to inosine (I) in the tRNA wobble position [69].
Absolute tRNA levels of HeLa cells were determined by using the tRNA microarrays. Gel-purified tRNA from the total RNA was ligated to increasing concentrations of the Cy3-labeled fluorescent stem-loop RNA/DNA oligonucleotide. Only values in the linear range (0.57 μM, 1.13 μM, and 2.25 μM) were considered from several microarrays. A detailed protocol is available at protocols.io (https://doi.org/10.17504/protocols.io.hfcb3iw).
tRNA silencing was performed by using shRNAs targeting the anticodon loop of tRNAThr(CGU) [70]. In brief, shRNAs were cloned into the pSUPER vector (Oligoengine) by using BglII and HindIII restriction sites according to the manufacturer's instructions. shRNA bearing pSUPER plasmids (1 μg) were cotransfected together with CFTR constructs (1 μg) into subconfluent HeLa cells (3.5-cm dish) by using PEI. Mock-transfected cells were cotransfected with CFTR constructs and an empty pSUPER vector. Two different shRNA sequences targeting tRNAThr(CGU) were used (5′ to 3′):
shThr1, GGCGTCGGTCTCGTAAACCGAAGTTCAAGAGACTTCGGTTTACGAGACCGACGCC and shThr2, GTGGTAAGGCGTCGGTCTCGTAATTCAAGAGATTACGAGACCGACGCCTTACCAC (underlined nucleotides denote the sense tRNAThr(CGU) target sequence). Cells were analyzed for CFTR protein level or tRNAThr(CGU) level 48 h after transfection. A detailed protocol is available at protocols.io (https://doi.org/10.17504/protocols.io.hgfb3tn).
Approximately 5 million CFBE41o- cells, in 3 independent biological replicates, were used to isolate mRNA-bound ribosome complexes, followed by extraction of RNase I digestion-derived, ribosome-protected fragments (RPFs) as described in [14]. Cells were collected by flash-freezing without preincubation with antibiotics. CHX (100 μg/ml) and harringtonine (2 μg/ml) were present in the lysis buffer and the sucrose gradient buffer to prevent ribosome dissociation in the postprocessing steps. The cDNA libraries from RPFs were prepared by using a modified protocol for miRNA with direct ligation of the adapters and were sequenced with TruSeq SBS kits (Illumina) on a HiSeq2000 (Illumina) machine. Sequenced reads were trimmed by using fastx-toolkit (0.0.13.2; quality threshold: 20), and sequencing adapters were cut by using cutadapt (1.2.1; minimal overlap: 1 nt). Processed reads were uniquely mapped to the human genome (GRCh37) by using Bowtie (0.12.9), allowing a maximum of 2 mismatches (parameter settings: -l 16 -n 1 -e 50 -m 1—strata—best y). RPFs were binned in groups of equal read length, and each group was aligned to the start or stop codons as described in [21, 71]. Taking into account that the P site of the ribosome covers the start codon, for each read length we calculated the distance between the middle nucleotide in the A site and 5′ of the read by using this distance to determine the center of each A site codon along each mRNA. We used 5′ calibration of the reads, as the RNase I cleavage was more variable on the 3′ side of the ribosome-protected fragment, which is consistent with prior studies [21, 71]. The majority of our sequence reads were 27–29 nts as expected. The ribosome dwelling occupancy per codon was calculated as described [40]. The reads over each position i in a gene were normalized to the average number of footprints across this gene. These ratios were then averaged across all genes to give the ribosome dwelling occupancy of a given codon in the transcriptome. The Met codon was excluded from these calculations because N-terminal Met is influenced by the presence of harringtonine used in the processing buffers. The first 51 nt were excluded from the calculations to avoid depletion of ribosomes at the beginning of genes by runoff elongation during cell harvesting [40].
If not stated otherwise, results are expressed as means ± SEM of n observations. Sample sizes of cell analyses were selected to demonstrate reproducibility among independent biological replicates and with adequate power to resolve significant differences among conditions. Patient samples were blindly allocated during experiments without prior knowledge of genotype; each separate patient-derived primary HBE cell sample was considered a single biological replicate. Differences between groups were evaluated by using 2-tailed Student t test implemented in Prism 5 or SigmaPlot 12 (Systat Software Inc.) software. Differences were considered statistically significant when P < 0.05.
tRNA microarray data have been deposited with the Gene Expression Omnibus (GEO) under the accession number GSE53991. The sequencing data were also submitted to GEO under the accession number GSE74365.
|
10.1371/journal.pbio.1001809 | A Functional Screen Reveals an Extensive Layer of Transcriptional and Splicing Control Underlying RAS/MAPK Signaling in Drosophila | The small GTPase RAS is among the most prevalent oncogenes. The evolutionarily conserved RAF-MEK-MAPK module that lies downstream of RAS is one of the main conduits through which RAS transmits proliferative signals in normal and cancer cells. Genetic and biochemical studies conducted over the last two decades uncovered a small set of factors regulating RAS/MAPK signaling. Interestingly, most of these were found to control RAF activation, thus suggesting a central regulatory role for this event. Whether additional factors are required at this level or further downstream remains an open question. To obtain a comprehensive view of the elements functionally linked to the RAS/MAPK cascade, we used a quantitative assay in Drosophila S2 cells to conduct a genome-wide RNAi screen for factors impacting RAS-mediated MAPK activation. The screen led to the identification of 101 validated hits, including most of the previously known factors associated to this pathway. Epistasis experiments were then carried out on individual candidates to determine their position relative to core pathway components. While this revealed several new factors acting at different steps along the pathway—including a new protein complex modulating RAF activation—we found that most hits unexpectedly work downstream of MEK and specifically influence MAPK expression. These hits mainly consist of constitutive splicing factors and thereby suggest that splicing plays a specific role in establishing MAPK levels. We further characterized two representative members of this group and surprisingly found that they act by regulating mapk alternative splicing. This study provides an unprecedented assessment of the factors modulating RAS/MAPK signaling in Drosophila. In addition, it suggests that pathway output does not solely rely on classical signaling events, such as those controlling RAF activation, but also on the regulation of MAPK levels. Finally, it indicates that core splicing components can also specifically impact alternative splicing.
| The RAS/MAPK pathway is a cornerstone of the cell proliferation signaling apparatus. It has a notable involvement in cancer as mutations in the components of the pathway are associated with aberrant proliferation. Previous work has focused predominantly on post-translational regulation of RAS/MAPK signaling such that a large and intricate network of factors is now known to act on core pathway components. However, regulation at the pre-translational level has not been examined nearly as extensively and is comparatively poorly understood. In this study, we used an unbiased and global screening approach to survey the Drosophila genome—using Drosophila cultured cells—for novel regulators of this pathway. Surprisingly, a majority of our hits were associated to either transcription or mRNA splicing. We used a series of secondary screening assays to determine which part of the RAS/MAPK pathway these candidates target. We found that these factors were not equally distributed along the pathway, but rather converged predominantly on mapk mRNA expression and processing. Our findings raise the intriguing possibility that regulation of mapk transcript production is a key step for a diverse set of regulatory inputs, and may play an important part in RAS/MAPK signaling dynamics.
| The RAS/MAPK pathway consists of a core module of three kinases (RAF, MEK, and ERK/MAPK) that transmit signals downstream of the small GTPase RAS. Upstream factors such as receptor tyrosine kinases (RTKs), which respond to extracellular signals, lead to RAS activation by a guanine nucleotide exchange factor (GEF). GTP-loaded RAS then triggers the sequential activation of RAF, MEK, and MAPK; active RAF phosphorylates and activates MEK, which in turn phosphorylates and activates MAPK [1]. Unlike RAF and MEK, MAPK has a variety of cytoplasmic and nuclear substrates that include transcription factors such as c-Jun, c-Fos, p53, ELK1, c-Myc, c-Myb, STAT1/3, SRF, and SMAD1/2/3/4 [2]–[4]. Phosphorylation of these targets, and others, by MAPK induces a wide range of cellular responses that include proliferation, differentiation, and survival [5]. Also, RAS/MAPK signaling's important role in oncogenesis and various developmental disorders has been recognized early on and abundantly studied [6],[7].
Over the last two decades, genetic screens in metazoan models such as Drosophila and Caenorhabditis elegans have been instrumental in identifying a growing list of key regulators of the RAS/MAPK pathway such as sos [8], csw [9], ksr [10]–[12], Cbl [13], dos [14], mts/PP2A [15], sur-8/soc-2 [16],[17], cnk [18], spry [19], sur-6 [20], PTP-ER [21], let-7 [22], alph/PP2C [23], and hyp/ave [24],[25]. Thus, these studies and research conducted in other systems have revealed a large network of factors whose regulatory activity converges on the core MAPK module [1],[5],[6],[26],[27]. This regulatory network includes complex features such as feedback loops [28]–[30], compartmentalization [1],[31], crosstalk with other signaling pathways [32], allosteric modulation via dimerization [27],[33], and the formation of larger order complexes called nanoclusters [34]. While the function of the core module is well characterized, many aspects of the network that surround it are still poorly understood, including its protein composition. Also, many of the identified regulators influence RAS-mediated RAF activation, which is in agreement with the fact that this particular step is subjected to a tight and complex regulation [27]. In comparison, fewer positively acting components have been found to act downstream of RAF, suggesting that MEK and MAPK activation depends on more simple regulatory mechanisms. Alternatively, such modulators might have eluded detection. Finally, most of the regulatory input that has been described so far acts at the post-translational level. Comparatively little is known on how RAS/MAPK component expression is controlled.
The success of the aforementioned genetic screens typically relied on the qualitative modification of a visible phenotype. This consideration, together with the technical limitations associated with genetic screening procedures, usually limit results to a handful of confirmed hits. RNA interference (RNAi) used as a functional genomics tool provides the possibility of a more comprehensive type of analysis providing a systematic means to functionally annotate the genome [35],[36]. Moreover, the possibility of using quantitative assays, in particular, allows for the identification of a much wider range of regulators [37]. However, the considerable number of candidates often identified by this methodology has made the perspective of rapid functional annotation a daunting task.
Here, we present the results of a genome-wide RNAi screen in Drosophila S2 cells that specifically focused on signal regulation between RAS and MAPK. Validated hits were submitted to a series of secondary assays aimed at positioning their regulatory input with respect to the three core kinases. In addition to identifying and correctly positioning most of the components previously known to mediate RAS-induced MAPK activation, the screen led to the discovery of several new factors that act at different steps along the pathway. Notably, we identified five novel components that act upstream of RAF. The homologs of these five proteins are part of a complex named striatin-interacting phosphatase and kinase complex (STRIPAK) [38] that also includes PP2A, which is known to regulate RAF activation [39],[40]. Unexpectedly, the majority of our candidates did not map to the interval between RAS and RAF, but were instead positioned further downstream. These included some transcription factors that we found regulate the transcript abundance of mek, mapk, or of the MAPK phosphatase PTP-ER. However, most of the novel factors were associated with mRNA processing and were found to act downstream of MEK and to regulate mapk splicing. Among these were components of the exon junction complex (EJC), which we and others have previously reported to be involved in regulating the splicing of the mapk pre-mRNA [41],[42]. In particular, depletion of the EJC was found to alter the splicing of mapk's long introns and cause a reduction in the amount of functional protein product. In this study, we focus on the function of a larger group of canonical splicing factors that also regulate mapk splicing. We show that the impact of these factors on alternative splicing (AS) of mapk differs from what we previously described for the EJC, indicating that two different types of regulatory input act on this step in mapk expression.
Thus, in addition to providing a comprehensive view of regulatory factors influencing signal transmission between RAS and MAPK, this work suggests that pathway output does not solely rely on post translational regulatory events, such as those controlling RAF activation, but is also tightly governed by the regulation of the expression of core components. In particular, the expression of MAPK emerges as a focal point for multiple different regulatory inputs.
To systematically search for and categorize new factors that specifically modulate signaling between RAS and MAPK, we employed a screening strategy that involved three distinct steps: (1) a primary genome-wide RNAi screen, (2) a validation screening step aimed at eliminating false positives, and (3) validated candidates were submitted to a series of a secondary screens to establish the position of their regulatory input relative to known pathway components (epistasis) and assess their specificity to RAS/MAPK signaling (Figure S1A).
We employed an automated immunofluorescence-based microscopy assay that quantitatively detected variations of dually phosphorylated MAPK (pMAPK) in Drosophila S2 cells. This assay was used to screen a genome-wide long double-stranded RNA (dsRNA) library for modulation of pMAPK levels induced by RASV12 expression (Figure S1B and S1C). The results from this primary screen and all subsequent screens are made available online at the IRIC RNAi database (http://www.bioinfo.iric.ca/iricrnai). 309 hit genes, which reproducibly altered pMAPK signal, were identified in the primary screen (Table S1). Importantly, core RAS/MAPK pathway components (e.g., raf/phl, mek/Dsor1, mapk/rl, cnk, and ksr) were amongst the strongest hits that decreased the pMAPK signal (Figure 1A; Table S1). Other known positively acting genes were also identified, such as 14-3-3ζ and the RAF chaperone Cdc37 [43]. Another expected hit was βggt-I, which encodes a factor involved in RAS prenylation [10],[44],[45]. Also expected, the PP2C phosphatase alph was identified as a negative regulator [23].
We next conducted two successive validation steps to address readily identifiable sources of false positives, namely effects on the pMet-RASV12 expression system and dsRNA off-target effects. 101 genes of the initial 309 primary hits passed both validation criteria (Figure S2; Table S1; Text S1). Validated genes were then assigned to broad functional categories on the basis of their associated gene ontology (GO) terms and on the functions of predicted homologs. Interestingly, transcription and mRNA processing factors composed, together, roughly half of our candidates (Figure 1B), and mRNA processing was the most highly enriched GO term of our hit set (Table S2). Despite the fact that mRNA splicing factors are often enriched in RNAi screen hit lists [46], we chose not to apply a selection bias against any group of genes at this stage. Therefore, all of the candidate genes passing both primary and validation screen criteria were evaluated in secondary screens without distinction.
To further characterize the 101 candidate genes, we conducted a series of secondary screens that can be subdivided into three groups: (1) MAPK activation induced by stimuli upstream of RAS, (2) epistasis screens involving MAPK activation at the level of or downstream of RAS, and (3) JNK activation screens aimed at addressing specificity to the MAPK pathway context (Figure S3; detailed results in Tables S3, S4, S5).
Four distinct MAPK activation assays using RTKs or GAP RNAi (Figure S3) were conducted to assess the degree to which pathway activity could be perturbed by depletion of the candidate genes in different activation contexts occurring upstream of RAS. Although a few exceptions were found, in most cases, we observed signal modulation that was generally consistent with our RASV12 results.
Next, epistasis experiments were carried out using S2 cell lines expressing either constitutively activated forms of RAF or MEK (Figures 2 and S3). The aim of these experiments was to position the identified genes in relation to the core kinases of the pathway by comparing the values from the RAS, RAF, and MEK activation assays. To do this, we calculated the correlation of our screening data with theoretical profiles of hypothetical components acting within three possible epistasis intervals (RAS-RAF, RAF-MEK, and MEK-MAPK) using an uncentered Pearson's correlation metric (Figure 2C; Table S3; Text S1). All known pathway components were positioned correctly by this approach. For example, Ras85D, ksr, cnk, hyp/ave, 14-3-3ζ, 14-3-3ε, and βggt-I are part of a group of genes that suppressed RASV12, but not activated RAF or activated MEK; these components were thus correctly positioned in the RAS-RAF interval (Figure 2B and 2C). Nine additional genes also fell into this category (Table S3) and therefore represent potentially novel pathway regulators acting at this level. Strikingly, while only eight hits (including the RAF chaperone Cdc37) mapped between RAF and MEK, most of the candidates (69) were assigned to the MEK-MAPK interval (Table S3), with the majority of these being factors not previously linked to RAS/MAPK signaling. This unexpected finding suggests that additional regulatory events that escaped prior detection are lying downstream of MEK.
RAC1V12 and peptidoglycan (PGN) were then used as stimuli in two JNK activation assays as a proxy to evaluate specificity to the RAS/MAPK signaling context (Figure S3). Very few of the candidates modulated pJNK to a similar extent as they did pMAPK (Table S4). One of these was the ALPH PP2C phosphatase, whose depletion increased both pMAPK and pJNK signals. This is consistent with our recent findings demonstrating that ALPH negatively regulates both MAPK and JNK signaling [23],[47]. Remarkably, the vast majority of the RNA processing factors identified in the primary screen did not modulate pJNK levels and thereby argued for their specific role in RAS/MAPK signaling (Table S4).
We next submitted our secondary screen data to unsupervised hierarchical clustering to group candidates with similar profiles together (Figure 3A). Bona fide pathway components with similar functions are clearly grouped together by this analysis. For example, Ras85D, ksr, cnk, and hyp all act at the level of RAF activation and all show very similar profiles. Both 14-3-3 isoforms, which also act at this level, are grouped together and are also close to the first group of genes involved in RAF activation, as is βggt-I, a component involved in RAS prenylation. On the basis of these findings, we can expect that candidates who have a similar profile to bona fide RAS/MAPK pathway components might in fact share the same function as these components.
Following this, we sought to identify putative protein complexes as well as related factors in our set of candidates by constructing a protein interaction network (PIN) based on publicly available protein and genetic interaction data (Figure 3B). The canonical RAS/MAPK components are clearly grouped together in this network and at least two other complexes can be clearly distinguished. The first consists of components of the STRIPAK complex and the second is composed primarily of mRNA processing factors. Remarkably, the components of both complexes also group together in similar functional profiles in our clustering analysis (groups “1” and “2” in both panels of Figure 3).
Given that several of our candidate genes are linked to RNA processing and transcription, we hypothesized that these factors might be acting on the expression of one or multiple RAS/MAPK pathway components. We first investigated the impact of our candidates on the expression of core RAS/MAPK pathway component transcripts (Ras85D, raf, mek, mapk, ksr, cnk, and PTP-ER) by quantitative PCR (qPCR) (Figure S4A; Tables S5 and S8). A specific effect on the mapk transcript was observed upon depletion of mago and eIF4AIII (Figure 4A; Tables S5 and S6) as we have previously reported [41]. At least three other factors (Cdk12, Fip1, and CG1603) also seemed to modulate the transcript levels of mapk. Moreover, two factors, gfzf and CG4936, were found to modulate the levels of mek and PTP-ER, respectively (Figure 4A; Table S6). CG1603, gfzf, and CG4936 were subsequently tested in larval eye disc tissue where similar results were obtained (Figure 4B).
Surprisingly, aside from the candidates mentioned above, most hits did not appear to cause a significant change in the transcript levels of pathway components. We had previously observed that qPCR assays targeting mapk are not always strongly affected by the splicing changes induced by EJC depletion. On the other hand, a reverse transcription PCR (RT-PCR) assay spanning the whole mapk transcript is a more sensitive tool allowing for detection of small splicing changes [41]. Based on this premise, and on the fact that almost all the splicing factors we identified mapped downstream of MEK, we decided to systematically examine the impact of these factors on mapk splicing. Consequently, we used the RT-PCR assay that had been used with the EJC to examine mapk splicing. Interestingly, not only did this experiment reveal that nearly all the splicing factors in our set caused shifts in the mapk RT-PCR profile, but these RT-PCR profiles were also clearly different from those produced by EJC depletion (Figure S5). Thus, while the impact of most of our candidates on mapk expression may not be apparent when measuring total transcript abundance, a clear impact on the different mapk isoforms can be observed by RT-PCR, indicating that these factors regulate AS of mapk.
In the case of the EJC, the splicing changes were accompanied by a corresponding decrease in MAPK protein levels. Because of this decrease, we decided to also measure the impact of our candidates on MAPK protein levels using quantitative immunofluorescence. This analysis confirmed that most of the factors positioned downstream of MEK (including most of the RNA processing factors) also caused a reduction of MAPK protein levels. Conversely, AKT protein levels, which were used as a control, were not generally sensitive to depletion of these same factors (Figure 5A). We also verified the impact on MAPK and two other pathway components (RAS and CNK) by Western blot. Most candidates that caused a reduction in MAPK levels did not impact the levels of RAS, CNK, or AKT (Figure S6), which mirrored the results from the immunofluorescence experiment. Thus, both evaluations of MAPK protein levels agreed with the RT-PCR experiments suggesting that the changes in splicing results in a reduction of MAPK protein abundance.
Most of the other hits, including those positioned at the RAS-RAF and RAF-MEK intervals, did not appear to cause a change in the expression of RAS/MAPK pathway components. However, we observed that a minority of candidates seemed to cause fluctuations in multiple proteins or transcripts. Likewise, these candidates also tended to impact some or all of the non-RAS/MAPK related assays we tested them in (Table S5). Moreover, these hits were also more frequently present in hit lists of other published RNAi screens (Table S4). Consequently, in order to discriminate such non-specific hits from higher quality candidates, we derived a scoring system that factored in results from the pJNK assays, pMet-green fluorescent protein (GFP) expression, hit occurrence in previously published RNAi screens, and the impact on measured protein and transcript levels (see Table S4; Text S1). As expected, a few of the splicing factors in our list had low specificity scores. On the other hand, most of the factors that selectively affected MAPK levels, including the majority of the splicing factors, were not present in previous screen hit lists more frequently than bona fide RAS/MAPK factors and generally displayed a good specificity to the RAS/MAPK context (Tables S4 and S5). This finding suggests that discriminating against an entire category of genes on the basis of the enrichment of that category in previous screen hit sets is a strategy that can lead to elimination of meaningful candidates.
Most of the positive regulators positioned in the RAS-RAF interval were factors that had previously been linked either to RAS prenylation or regulation of RAF activation. As previously mentioned, one of the RAS prenylation factors we identified was βggt-I, which was originally identified in Drosophila [10],[45]. In addition to βggt-I, two other factors that are known to function in RAS prenylation in other organisms were also identified: Hmgcr and Fnta (CG2976). FNTA is the alpha farnesyltransferase subunit for mammalian RAS proteins [48]. The hydroxymethylglutaryl-CoA reductase HMGCR functions in the cholesterol biosynthesis pathway and is required for farnesylation of RAS and other membrane-associated proteins [49],[50]. In our Western blot experiments, all three of these factors were observed to cause a mobility shift in RAS suggesting that RAS geranylgeranylation is impaired [51], and thus that these factors act on RAS in Drosophila (Figure S6A, S6C, and S6J).
RAF activation is arguably the most tightly regulated step of the MAPK module [1],[27],[52]. Multiple components are involved in a series of events that link up RAF to RAS, anchor RAF at the plasma membrane, allow RAF to adopt and maintain an active conformation and finally enable efficient substrate targeting [52]. Phosphorylation and dephosphorylation events control progression throughout these steps [52]. Of these, the removal of the phosphate moiety on the S346 residue of Drosophila RAF (equivalent to S259 of human RAF1) is one of the pivotal regulatory events as it is thought to trigger the release of 14-3-3, which otherwise sequesters RAF in the cytoplasm [52]. Of the factors involved in RAF activation, all of the expected factors (ksr, cnk, hyp, 14-3-3ε, and 14-3-3ζ) were correctly positioned at the RAS-RAF interval (Figure 2B). Two other candidates, Pp1-87B and Sur-8, also clustered together with the set of known factors acting in the RAS-RAF interval. These two factors had not previously been shown to act on MAPK signaling in Drosophila, but evidence from other organisms indicates that they might act at this level [16],[17],[40], which is consistent with our results. Of particular interest, one study has found a mammalian complex composed of PP1, SUR-8, and MRAS and linked it to dephosphorylation of the S259 residue on C-RAF [53].
The only other positive regulators in the RAS-RAF epitasis group that were not previously linked to RAS/MAPK signaling were five components that form the smallest of two complexes in our network (CKA, STRIP, SLMAP, FGOP2, and MOB4) (Figure 3B). Three of these components are distantly related to budding yeast alpha factor arrest (FAR) complex components, which are involved in signaling G1 arrest upon alpha factor stimulation [54]. More recently, a protein complex comprising Striatin, the catalytic subunit of PP2A, the STE20 family kinase STK24, and four additional core proteins was identified in human cells and named the STRIPAK complex [38]. The core of this complex was suggested to serve as a protein platform that specifies PP2A and/or STK24 action. Remarkably, the five RAS-RAF proteins identified are homologs of the non-catalytic members that make up the core STRIPAK complex. CKA, which is the fly Striatin homolog, has previously been demonstrated genetically to act as a positive regulator of JNK signaling [55]. MOB4 has also been previously studied genetically in Drosophila, where it appears to participate in mitotic spindle assembly [56]. The three other members, STRIP, FGOP2, and SLMAP, have not been extensively studied in flies and are named on the basis of their mammalian counterparts.
Consistent with their ability to work together as a complex, the five STRIPAK homologs had similar effects in all the secondary screens and epitope-tagged variants co-immunoprecipitated in binary co-expression experiments (Figures 3A and S7). Notably, their depletion also suppressed JNK activation induced by RAC1V12, suggesting that CKA/Striatin modulates signaling through this pathway and that the other STRIPAK members act in conjunction with Striatin in this context. This is also consistent with the findings that TRAF3-interacting JNK-activating modulator (T3JAM), one of the SLMAP homologs, is linked to JNK signaling [57] and with a recent report that identified Cka as a suppressor of JNK signaling [58]. Furthermore, depletion of STRIPAK components reduced pMAPK signal induced by insulin, activated Sevenless RTK (SEVS11) and GAP RNAi, but only marginally affected EGFR signaling (Figure S5). This suggests that the role of STRIPAK differs depending on the MAPK and JNK activation contexts.
To confirm the involvement of STRIPAK complex components in RAS/MAPK signaling in vivo, we conducted genetic interaction experiments using Cka/Striatin mutant alleles [55]. RAS/MAPK activity is required for neuronal photoreceptor and cone cell differentiation during Drosophila eye development [59],[60]. Expression of RasV12 under the control of the eye specific sev promoter/enhancer regulatory sequences produces extra photoreceptor cells, which causes a characteristic rough-eye phenotype (Figure 6B) [61]. This rough eye phenotype was dominantly suppressed in a Cka heterozygous mutant background (Figures 6E and S8A). Extra wing vein material produced by a constitutively active Egfr allele, EgfrElp, was also dominantly suppressed by Cka mutant alleles to a degree comparable to a weak loss-of-function allele of rl/mapk (Figure S8E and S8J). In agreement with these results, wing vein deletions were significantly enhanced in a shp-2/csw hemyzygous mutant background when a Cka mutation was introduced in this context (Figure S8D, S8I, and S8K). Moreover, consistent with the role of STRIPAK in RAS/MAPK signaling, loss-of-function of Cka activity impaired R7 photoreceptor cell differentiation, which is a classical RAS/MAPK-dependent developmental event (Figure S9).
In addition to PP2A/mts, we noted that some of the other STRIPAK components described in Goudreault and colleagues [38] were not identified in our primary screen. However, one of these, GckIII, was one of the validated regulator in the InR-driven MAPK screen reported by Adam Friedman and Norbert Perrimon [62]. This raised the possibility that GckIII might have an impact on pathway activity in alternate activation contexts. To address this, we examined the effects of depleting GckIII in RASV12, insulin, and GAP RNAi assays. We found that while this had little impact on RASV12-induced MAPK activation, GckIII depletion had an effect comparable to Fgop2 depletion in the insulin and GAP RNAi contexts (unpublished data). This indicates that GckIII may function upstream or in parallel to RAS and raises the intriguing possibility that the STRIPAK complex regulates multiple aspects of the larger RTK/RAS/MAPK pathway.
Since Striatins are defined as PP2A regulatory (B) subunits [63] and STRIPAK was initially described as a PP2A associated complex, it is possible that STRIPAK assumes this role in the context of RAF activation. We observed a modulation of RASV12 signaling upon depletion of the catalytic subunit of PP2A (mts), but not of the regulatory B subunit, tws, which support the notion that STRIPAK components are functioning as PP2A regulatory (B) subunits in this context (unpublished data). A similar function for STRIPAK has recently been described in the context of Hippo signaling, where it was found to associate with PP2A and HPO [64]. Finally, in agreement with our findings, another recent report has linked the CKA subunit to RAS/MAPK signaling [65] and another study conducted in Neurospora has indicated that MAPK regulates STRIPAK function, suggesting the possibility of regulatory feedback phosphorylation [66].
Only 13 of our candidates were found to act downstream of RAF and upstream of MEK. Of these, the RAF chaperone Cdc37 was the only bona fide pathway regulator. Among the others, CG8878 was of interest as it is the homolog of the mammalian Vaccinia-Related Kinase (VRK) genes of which two (VRK1 and VRK3) have been recently identified in a recent screen for KRAS synthetic-lethal factors [67]. Also, another study has shown that VRK2 interacts with MEK and KSR1, potentially acting on MEK activation [68]. However, in this context, VRK2 acts as a negative regulator whereas CG8878 appears to act positively on RAS signaling in our experiments.
The most interesting candidate that fell within the RAF-MEK interval was GST-containing FLYWCH zinc-finger protein (gfzf). GFZF was initially found because of the property of its GST domain to bind to glutathione sepharose beads [69]. It does not, however, have any clearly identified function assigned to it, though there are indications that it may act as a co-factor for the E2F transcription factor [70]. Also, recent RNAi screens have suggested that gfzf may be acting downstream of PDGF to control cell size [71] and as a factor functioning in the G2/M DNA damage checkpoint [72]. In our qPCR experiments, gfzf clearly stood out from other candidates as it was the only factor that caused a strong reduction in mek transcript levels and had the closest profile to the mek RNAi itself (Figures 4A, 4B, and S4A). This observation also extended to the protein product as we observed a clear reduction in MEK levels upon gfzf knockdown (Figure 4C). In flies, two different gfzf loss of function alleles suppressed the RasV12 rough eye phenotype (Figures 6F and S8A). gfzfcz811also increased the severity of wing vein deletions in hemizygous cswlf males (Figure S8D and S8K). Finally, knocking down gfzf reduced RASV12-induced hemocyte proliferation (Figure 7A). Together, these observations suggest gfzf is regulating mek, possibly by acting as a positive transcription factor. Since FLYWCH domain proteins have been found to negatively control miRNA expression in C. elegans [73], one interesting alternative is that gfzf represses the production of a miRNA that targets the mek transcript.
Another candidate, CG9797, was clustered close to gfzf through the functional screen data and the qPCR results (Figures 3A and S4A), and its protein product is predicted to interact with GFZF (Figure 3B). In the qPCR screen, CG9797 knockdown caused a weak reduction in mek transcript levels; it is the fourth strongest hit in terms of reducing mek levels, after the dsRNAs targeting gfzf, mek, and RpL24 (Table S5). From this, and since CG9797 encodes another zinc finger protein, it is possible that this factor works in conjunction with GFZF as a transcriptional regulator. Recently, the chromatin remodeling factors Geminin and Brahma have been found to modulate MEK protein expression in drosophila wing discs [74], raising the possibility that gfzf might be acting in conjunction with these factors.
mRNA processing and transcription factors formed the largest group of hits in our study and also the largest complex in our network analysis (Figures 3B and S1B), and almost all these factors mapped to the MEK-MAPK epistasis interval (Table S5). While the majority of the candidates in this category cause changes in mapk expression, one clear exception was CG4936. This gene encodes a zinc finger protein of unknown function that is distantly related to human ZBTB20, a BCL-6 like transcription factor that is expressed in hematopoietic tissues and lymphoid neoplasms [75]. CG4936 also mapped downstream of MEK, but was not found to influence mapk expression. A first clue as to the function of this factor was provided by the observation that CG4936 behaved very similarly to PTP-ER in our functional screens; of all the candidates, it has the closest profile to PTP-ER (Figure 3A). One of the two MEKEE-based MAPK activation assays involved the use of PTP-ER RNAi to increase the pMAPK signal. Surprisingly, while both PTP-ER and CG4936 RNAi significantly increase MEKEE-induced pMAPK levels, combining CG4936 and PTP-ER RNAi did not have an additive effect on pMAPK levels (Figure 2B; Table S5). This finding suggests that these factors work together in the same regulatory pathway. CG4936 also clustered close to PTP-ER in our qPCR screen (Figure S4A). Consistent with this, both qPCR and protein analysis revealed that CG4936 knockdown respectively caused a specific reduction in PTP-ER transcript and protein levels (Figure 4A, 4B, and 4D). Furthermore, in our genetic interaction experiments, flies trans-heterozygous for two P-element insertion alleles of CG4936 (CG4936EY10172/CG4936DG10305) showed an enhancement of the RasV12 rough eye phenotype (Figure 6G) and suppressed lethality and phenotypes caused by a homozygous mapk/rl1 hypomorphic allele (Figure S8C, S8G, and S8H). CG4936EY10172 alone suppressed the lethality and wing vein deletions of cswlf hemizygous males (Figure S8I and S8K), as did the double CG4936 mutant (unpublished data). Moreover, cswlf homozygous females were observed in a background heterozygous for CG4936EY10172, which is indicative of suppression as cswlf is recessive lethal. Altogether, these data suggest that CG4936 acts on PTP-ER transcription and that this action has bearing on RAS/MAPK dependent developmental processes.
Excepting CG4936, most of the other candidates positioned downstream of MEK (39) caused a decrease in MAPK protein levels (<−0.25 log2-fold and p<1×10−5) (Table S5). A few of these factors caused a significant and reproducible reduction in mapk transcript levels as well, without significantly impacting the levels of other RAS/MAPK pathway components (Figure S4A; Table S5). These factors include two EJC components (eIF4AIII and mago) as well as three other candidates: Cdk12, Fip1, and CG1603. Depletion of these latter candidates led to a significant drop in mapk transcript levels as measured by qPCR (Figure 4A; Table S6). However, these three factors could be distinguished from the EJC components by the fact that they did not cause a shift in the mapk RT-PCR profile (Figure S5), but only a decrease in the overall transcript abundance, suggesting that they do act at a different step of mapk expression. Finally, lethal alleles of CG1603 and Fip1 dominantly suppressed a RasV12 induced rough eye phenotype (Figures 6H and S8A). The CG1603 allele also suppressed the extra wing vein phenotype caused by EgfrElp, as did Cdk12KG05512 (Figure S8E and S8J). However, none of the alleles had a readily observable impact on cswlf phenotypes (unpublished data); although Fip1 did cause a slight enhancement of cswlf lethality (Figure S8I). Finally, consistent with our cell culture data, knockdown of CG1603 in larval imaginal disc clones was found to cause a pronounced decrease in MAPK protein levels (Figure 7C).
Cdk12 is the main RNA polymerase II C-terminal domain kinase. Phosphorylation of the C-terminal domain of RNAP II is required for transcription elongation, RNA processing, and splicing [76],[77]. Thus it is possible that Cdk12 influences either of these steps in the case of mapk expression although it does not seem to induce AS. CDK12 has previously been shown to function in conjunction with Cyclin L in regulating AS [78]. Consistent with these data, the Drosophila CycL ortholog, CG16903, was also a hit in our screen that mapped to the MEK-MAPK interval (Table S3) and caused a shift in the mapk RT-PCR profile (Figure S5). However, CG16903 did not significantly change mapk transcript levels measured by qPCR (Table S5).
Fip1 is also involved in transcript processing; its best studied ortholog in yeast is part of the pre-mRNA cleavage and polyadenylation complex [79]. mRNA polyadenylation plays an important part in stabilizing spliceosome assembly on the 3′ most exon of the transcript, is an important pre-requisite for effective mRNA export, and also influences mRNA stability [80]–[82]. Thus, like Cdk12, Fip1 may be regulating mapk expression by controlling transcript abundance or by influencing splicing efficiency.
CG1603 encodes a protein of unknown function that contains MADF type zinc finger domains that are related to Myb DNA binding domains [83]. CG1603 is poorly conserved in humans, bearing distant homology to ZNF664 and ZNF322. Interestingly, ZNF322 was found to act as a transcriptional co-activator of SRF and AP-1 in humans, which would position it downstream of MAPK signaling [84], thus representing a potential feedback mechanism.
The majority of the factors that were observed to lower MAPK levels were either spliceosome components or factors associated to the splicing machinery (Figure 5A; Table S5). Intriguingly, multiple lines of evidence suggested that these splicing components can also play a specific role in modulating RAS/MAPK signaling: (1) most splicing factors did not have any detectable impact on CNK, AKT, and RAS levels, even though these proteins are all derived from intron-containing genes (Figure S6), (2) depletion of most splicing factors by RNAi did not significantly modulate PGN (peptidoglycan) and RAC1V12-induced JNK activation (Figure 3A; Table S4), and (3) these splicing factors scored as hits in previous screens less often than bona fide RAS/MAPK pathway components (Table S4). Accordingly, the majority of these splicing factors were categorized in the high specificity group. Importantly, many of these had impacts on MAPK levels that were comparable to or greater in strength than those splicing factors of the lower specificity group. This finding suggests that the higher specificity score is not simply attributable to lower knockdown efficiency or a weaker impact on constitutive splicing.
Interestingly, one indication that the canonical splicing factors might be acting differently on MAPK signaling than the EJC came from our qPCR expression data. Namely, some high specificity splicing factors such as the CG10754 (the counterpart of human SF3A2, a U2 small nuclear ribonucleic particle [snRNP] associated factor involved in branch point binding [85]), CG3198 (the ortholog of LUC7L3, a predicted splicing factor [86]), Prp19 (the central component of the PRP19 spliceosomal complex involved in C complex assembly [87]), as well as CG4849 and CG6686 (two predicted tri-snRNP components [88],[89]) all caused a reduction in MAPK protein levels comparable to the EJC factors mago and eIF4AIII. However, no reduction of mapk mRNA was observed by qPCR (Figure 4A; Table S5). Furthermore, another important difference between the EJC and the canonical splicing factors was that the depletion of many candidates of the latter group caused an increase in nuclear pre-mRNA retention as measured by fluorescence in situ hybridization using a poly-A probe (Figure S4B and S4C) [90]. This indicates—as might be expected—that most of these splicing factors also play a more general role in splicing.
The key difference between the EJC and canonical splicing factors suggested by the qPCR and pre-mRNA export data was readily observable using the whole-transcript mapk RT-PCR assay; both groups caused an alteration in the mapk RT-PCR product, but the canonical splicing factors produced a clearly different pattern (Figure S5). Since most of the candidates in the canonical splicing factors group produced similar shifts in the mapk RT-PCR profile, we selected two representative factors for more detailed follow-up experiments. The first, Prp19, was selected because it is a core spliceosome component that caused a strong reduction in MAPK levels and had a clear impact on the mapk RT-PCR profile (Figures 5A, 5B, and 8B). The second, Caper, was selected because it is a serine rich (SR) AS factor [91] that did not perturb global pre-mRNA export, and it caused a weaker reduction in MAPK protein levels and a less severe change in the mapk RT-PCR profile (Figures 5A, 5B, 8B, and S5). Consistent with the protein expression data, while both factors clearly altered the mapk RT-PCR products, similar RT-PCR assays targeting Ras85D, raf, mek, cnk, or ksr did not display any obvious change (unpublished data). Also, Western blot experiments showed a clear drop in MAPK protein levels with no effect on RAS, RAF, MEK, or AKT (Figure 5B); this suggests that splicing of mapk, and not of other pathway components, is affected.
Also lending strength to the idea that splicing factors may have a specific role in the RAS/MAPK context was the fact that three Prp19 alleles and one Prp8 allele, were isolated by our group in an independent genetic screen for modifiers of a dominant-negative form of CNK (Figure S8B and CB, ML, MS, and MT, unpublished data). In addition to this finding, Prp8 has been previously found to enhance the small wing phenotype induced by expression of the Egfr inhibitor aos [92]. Consistent with this result, the Prp19 and Prp8 alleles dominantly suppressed the RASV12 rough eye phenotype, as did an allele of Caper (Figures 6I, 6J, and S8A). The Prp19 alleles also dramatically enhanced cswlf lethality and suppressed the EgfrElp wing phenotype (Figure S8E, S8I, and S8J). Moreover, in mapk/rl1 homozygous flies, Prp19 and Caper alleles enhanced the severity of wing vein deletions and rough eye phenotypes (Figures 6K–6R, S8G, and S8H). Importantly, although flies carrying one copy of both Caperf07714 and Prp19CE162 were perfectly viable, this allelic combination was entirely lethal in a mapk/rl1 homozygous background (Figure S8G). This result constitutes another indication that these splicing factors are acting in concert on mapk expression. Finally, splicing of mapk was found to be altered in Prp19CE162 flies also homozygous for rl1 (Figure S10A and S10B). The rl1 mutant alone reduces mapk transcript levels without altering the RT-PCR splicing profile [93; and unpublished data].
Complementing the genetic interaction experiments, expression of Prp19 RNAi reduced the RasV12-induced proliferation of larval hemocytes (Figure 7A). In addition, clonal tissue expressing Prp19 RNAi in wing imaginal discs consistently caused a reduction in MAPK levels (Figure 7D), although the reduction was not as pronounced as that of mapk RNAi. Clonal regions sometimes showed signs of apoptosis (in one of the two RNAi constructs tested and in wing discs in particular) suggesting that these tissues may be more sensitive to knockdown of Prp19 than S2 cells. Finally, splicing of mapk was found to be altered in wing disc segments where Caper had been knocked down (Figure S10C).
Altogether, these experiments suggest that these transcription and splicing factors are important in regulating MAPK levels, and thus are important for MAPK signaling. Furthermore, our genetic interaction data suggests that they can act in a number of different in vivo contexts. In particular, the Prp19 alleles had an impact in all our RAS/MAPK genetic interaction experiments and displayed some of the strongest phenotype modifications. Thus, of the different groups of candidates, it is quite possible that splicing factors are relevant to the broadest range of RAS/MAPK regulatory and developmental contexts.
Prp19 and Caper RNAi display similar mapk RT-PCR profiles, but Caper produces a more subtle shift in product size with less lower size bands observable (Figure 8B). To verify that these changes in the RT-PCR profile of mapk were due to altered splicing, we cloned and sequenced the RT-PCR products. We found that the lower size mapk transcripts produced by both Caper and Prp19 RNAi were generated by a series of exon skipping events and, to a lesser extent, by retention of the 5′ most intron of the RB/RE isoforms (Figure 8A–8D; Table S7). Prp19 and Caper RNAi produced many single exon skipping events with exons IV and VII being the most frequently skipped. These AS events differed from those we had previously observed in EJC depleted samples, where skipping of multiple consecutive exons was observed immediately to the 3′ end of exon I [41]. Interestingly, exon skipping events associated with the EJC mostly resulted in frameshifting due to the loss of the start site in exon II and/or skipping of exon III. On the other hand, the skipping of exons IV and VII associated with Prp19 and Caper produce an in-frame deletion potentially giving rise to a truncated protein product. The fact that we did not observe any smaller size products may be due to the epitope being removed or to the smaller products being unstable and degraded (MAPK is mostly composed of a kinase domain and it is likely that these deletions would disrupt proper folding).
In order to confirm that depletion of Prp19 and Caper produced AS changes in mapk that were different from those produced by EJC depletion, we designed RT-PCR assays aimed at detecting specific AS events. Using primer pairs in which the 5′ primer overlapped the exon junction between exons I and III or exons I and IV, we were able to detect an enrichment in exon II–III and II–IV skipping in eIF4AIII depleted samples, consistent with our previous results. However, these exon skipping events were not enriched in Caper and Prp19 RNAi samples. Conversely, using the same strategy, increased skipping of exons IV and VII was observed for Caper and Prp19, but not for eIF4AIII (Figure 8E). Also, using primers targeting the exons on either side of the skipped exons, we were able to detect both the canonical and the alternative transcripts produced by either Prp19 or Caper depletion in both S2 cells and wing disc tissue (Figures 8E and S10C). In sum, multiple lines of evidence indicate that Prp19 and Caper cause specific changes in mapk splicing that differ from those associated to the EJC. This finding suggests that at least two different types of regulation act on mapk AS. By extension, it is likely that the other splicing factors identified in our screen can be grouped with Caper and Prp19—and not the EJC—since they produced mapk RT-PCR profiles similar to these factors.
In this report, we presented the results of an RNAi screen for factors influencing signaling between RAS and MAPK in Drosophila. Most previously known pathway components have been identified in our screen, including a few that had not yet been found in flies (e.g., Sur-8, Pp1-87B, Hmgcr, and Fnta). On the basis of analyses of our secondary screen results and publicly available data, we assessed the specificity of our candidates, grouped them into protein complexes, and positioned their effect relative to the core RAS/MAPK pathway components (Figure 2). We also evaluated the impact of our candidates on the expression of core components of the RAS/MAPK pathway. From these data, we identified distinct groups of factors with different roles in modulating RAS/MAPK signaling. Furthermore, we show that the largest group of candidates—which is composed of splicing factors—acts specifically on mapk expression. This discovery is surprising since most of the previously described regulators of MAPK signaling act at the level of RAF activation. Thus, our results uncover an unappreciated point of control governing Drosophila RAS/MAPK signaling that takes place through the control of mapk expression (Figure 9). Moreover, the identification of two other factors (gfzf and CG4936) that act to control mek and PTP-ER expression adds another layer to the gene expression control network surrounding the RAS/MAPK pathway.
When comparing our results to those of three previous RNAi screens examining insulin receptor (InR) and EGFR induced MAPK signaling in Drosophila cells in culture [37],[62], we found that our validated hit set had a relatively limited overlap with those studies; 44 of our 78 positive regulators were present in their list of 986 positive regulators and 18 of our 28 negative regulators were also present in their group of 1266 negative regulators (Figure S11A and S11B). 331 hits reported in the InR screen were tested for their ability to modulate RASV12 in an assay similar to the one used in our screen. However, very few of the candidates found to modulate RASV12 signaling in their secondary screen were also present in our validated hit set; 14 of our 106 validated genes were found to alter RASV12 induced pMAPK beyond 5% of controls in their study (Figure S11C and S11D). Furthermore, nine of these 14 genes were bona fide pathway components. This limited overlap between our studies can be explained by two things: First, their S2R+ InR assay involved the use of an exogenous source of YFP-tagged MAPK where the YFP signal was used to normalize pMAPK signal. This strategy makes detection of factors that modify endogenous MAPK expression impractical (a large proportion of our candidates are involved in exactly this type of regulation). Second, when selecting which hits to follow up in secondary screening, the authors elected to exclude certain genes linked to large molecular complexes, most likely excluding many of the splicing factors that were retained in our study.
With the exception of the well described regulation of RAS by let-7 family miRNAs [22], surprisingly little is known on the regulation of MAPK module component expression. It was therefore surprising to find that our largest group of hits specifically decreased MAPK levels. Yet, control of MAPK expression is not unprecedented. For example, in yeast, both the FUS3 and HOG1 MAPKs are transcriptional targets of their respective pathways [94]. Pumilio mRNA binding proteins are also known to reduce MAPK activity by lowering the expression of the C. elegans MAPK, Mpk-1, as well as ERK2/MAPK1 and p38α/MAPK14 in human embryonic stem (ES) cells, through the binding of specific sites on the 3′ UTR of their respective transcripts [95]. Finally, LARP-1 RNA-binding proteins have been found to control the abundance of the transcripts of Mpk-1 and other pathway members in the C. elegans germ line [96].
In this study, we found that multiple splicing factors act downstream of MEK to specifically control MAPK levels. While some of these factors were associated with the regulation of AS, most were components of the spliceosome or factors that co-purify with the spliceosome. Interestingly, many lines of evidence suggest that AS can be modified by spliceosomal factors [97]–[106]. Global analyses of splicing events in yeast using a series of temperature sensitive alleles and deletions of splicing factors, found that spliceosome components differed in terms of which splicing events they altered [98],[99],[105]. Among these, the yeast homologs of Prp19, Prp8, U4-U6-60K, l(3)72Ab, Prp6, SmD2, SF3a120, CG4849, and CG10333—factors included in our set of hits—were found to differentially modulate the splicing of specific sets of transcripts [98]. Moreover, studies conducted in Drosophila using an RNAi-based strategy have demonstrated that knocking down many so-called “core” spliceosome components caused specific changes in specific AS contexts [100],[101]. In particular, some of the “core” components identified in our screen have been shown to have selective effects. For example, knocking down Prp6, l(3)72Ab, crn, Caper, SF3a120, and CG10418, was found to differentially influence AS of Dscam, para, TAF1, and/or dAdar [100],[101]. In mammals, SmB was also recently shown to be involved in regulating inclusion of alternative exons [103]. Finally, our observation that a Prp19 knockdown had specific effects on MAPK protein levels and the identification of Prp19 (as well as Prp8) in a separate CNK-based genetic screen supports the notion that these splicing factors are important for mapk expression in vivo. Thus, the fact that we identified specific spliceosome components in our screen and not others may reflect the importance of these particular components in regulating mapk splicing.
The specific elements of the mapk gene structure that dictate requirement of particular splicing factors have yet to be determined. Likewise, many—if not most—other genes may have structures that preclude AS regulation by components of the spliceosome. Still, AS processes in at least four other Drosophila genes seem to involve some of the same components of the splicing machinery that we have linked to mapk, indicating that a common characteristic may dictate the involvement of these spliceosome components in AS. On the other hand, the fact that some of the spliceosome factors are involved in specific AS contexts and not others, suggests that key differences exist. This observation is important as it implies that specific AS events can be modulated by controlling the activity of these spliceosomal factors. In support of this idea, one of the previously mentioned studies indicates that the control of TAF1 AS by a series of splicing factors—which include Caper and other spliceosome components—is downstream of a Camptothecin-induced ATR pathway [101]. Another example is the regulation of CD44 AS by RAS/MAPK-induced phosphorylation of Sam68, which has been found to function by regulating the activity of the U2AF65 spliceosome component [107]. In addition to Sam68, other splicing factors and spliceosome components—including some that were identified in our screen—were recently found to be the targets of ERK phosphorylation [108]–[110]. Finally, a general splicing repressor, SRp38, has also been shown to function as a sequence specific splicing activator upon phosphorylation in response to cellular stress [111].
One explanation for our results, as well as some of the previous observations, would be that some of the spliceosomal factors in our set may be interchangeable or function as non-essential co-regulators. Thus, removing these components would not greatly disrupt general spliceosome function, but rather lead to an altered spliceosome activity that selectively impacts sensitive AS contexts. An example of this is the interplay between PUF60 and U2AF65, which both function in 3′ splice site recognition; the two factors can either work cooperatively or independently, producing different splicing outcomes on the basis of the presence or absence of either protein [112]. A further example is the stress-induced relocalization of certain spliceosome and splicing factors that leads to changes in spliceosome configuration and AS [113]. Cellular stress has also been associated with the production of “non-productive” AS variants or silent messengers (transcripts that are either degraded by the quality control machinery or that do not encode functional proteins) [114],[115]. In fact, the AS we observed in mapk is reminiscent of the AS observed for the E3 ligase MDM2 following camptothecin-induced genotoxic stress [116]. In this study, the authors found that stress induced a number of non-productive MDM2 transcript isoforms that resulted in lower MDM2 protein levels and a stabilization of the MDM2 target, p53. It will be important to determine whether stress—or another signal—acts to control MAPK protein levels by inducing the AS we observe in our experiments.
The discovery that MAPK expression is specifically modulated by mRNA processing factors raises multiple questions. Not only will it be vital to define the upstream signals that dictate this activity but it will also be important to assess which other genes are similarly regulated and to identify the common characteristic that renders them sensitive to this type of regulation. Also, the time frame within which these changes occur will have implications as to the role that this type of regulation can play in MAPK signaling. Typically, RAS/MAPK signal modulation has been observed to occur through either rapid post-translational mechanisms or through slightly slower mechanisms involving control of protein stability and transcriptional control. However, the reduction in protein levels we observed as a consequence of AS are only apparent over a period of days, probably because the MAPK protein is relatively stable. This finding indicates that this regulation will not be relevant over the shorter timeframes of previously characterized regulatory events and also implies that a prolonged stimulus will be necessary to produce the effects we observe. Therefore, control of mapk splicing may be more important in the context of certain tissues and organs, in development or in disease. Interestingly, the abundance of core spliceosome components has been shown to be regulated and vary both temporally and across different tissues [117]–[121]. What is more, disruption in core spliceosome components has also been found to cause changes in AS in diseases such as spinal muscular atrophy and retinis pigmentosa [122]. Mutations in splicing factors have also been found to occur in a large proportion of myelodysplasia patients [104],[123] as well as in melanoma [106]. Another study has shown that, in glioblastomas and astrocytomas, splicing factors controlled by c-Myc play a role in controlling the expression of pyruvate kinase, a factor that is important for aerobic glycolysis [124]. It will be interesting to explore whether, in contexts such as these, the altered activity of splicing factors may regulate MAPK levels with important functional consequences for either normal or diseased cellular function.
pMet-RasV12 S2 cells diluted in Schneider medium (to a concentration of 1×106 cells/ml) were distributed in 96 well clear plates (Corning) containing 5 µl dsRNA aliquots at a concentration of ∼200 ng/µl. Plates were placed in plastic containers to reduce evaporation and incubated at 27°C for four days. RasV12 expression was induced by adding 0.7 mM CuSO4 to medium 24 h prior to fixation. Cells were resuspended and transferred to concanavalin A coated plates and allowed to settle and adhere for 1 h. Cells were then fixed in 4% paraformaldehyde/PBS, washed, and blocked in 0.2% Triton X-100/0.2% BSA/PBS (PBT/BSA) and incubated overnight with an anti-pMAPK antibody (1/2,000; Sigma number M8159), washed in PBT/BSA, and revealed using an anti-mouse Alexa Fluor 555-conjugated secondary antibody (1/1,000; Invitrogen number A-21424). DAPI (0.04 µg/ml) was used to stain nuclei. Mowiol (9.6% PVA, Fluka) was added to wells prior to imaging. An automated fluorescence microscopy system (Zeiss Axiovert) was employed for plate imaging. Autofocus, image acquisition, and analysis were conducted using MetaMorph (Molecular Devices) software. The cell-scoring application in MetaMorph was used for cell segmentation and quantification of fluorescent signal.
Candidates were assigned to one of three possible epistasis intervals (RAS-RAF, RAF-MEK, or MEK-MAPK) on the basis of the data from the following secondary screens: (1) RASV12, (2) RAFED, (3) RAFCT, (4) RAFEDCT, (5) MEKEE, (6) MEKEE+PTP-ER dsRNA. The correlations between normalized secondary screen log transformed values and three predetermined epistasis profiles were calculated using a modified uncentered Pearson's correlation:Where r is the correlation value [−1,1], x is the secondary screen value (for screens number 1, 2, 3, 4, 5, and 6), y is the predetermined profile value, and w is the weight applied to a given screening experiment (where w = [3 1 1 1 1 2]; see Text S1). The following predetermined epistasis profiles y were used:
RAS−RAF = [1 0 0 0 0 0]
RAF−MEK = [1 1 1 1 0 0]
MEK−MAPK = [1 1 1 1 1 1]
A negative r indicates reverse correlation and is observed for positive pathway regulators.
Fly husbandry was conducted according to standard procedures. All crosses were performed at 25°C. The sev-RasV12 line has been described previously [45]. EgfrElp was described in [125]. The Cka alleles [55] were kindly provided by S. Hou. The cswlf [126] flies were originally obtained from L. Perkins. The mapkE1171 allele was identified in a genetic screen as a dominant suppressor of a dominant negative form of KSR [127]. The Prp19CE162, Prp19CE40, Prp19TE1036, and Prp8CE309 alleles were recovered in a genetic screen for modifiers of a dominant negative form of CNK (CB, ML, MS, and MT, unpublished data).
RNAi fly lines were obtained from the VDRC [128]. All other fly lines described herein were obtained from the Bloomington stock center.
Adult fly eyes were imaged using a stereomicroscope (Leica MZ FL III) and CombineZP, a freely available software package (http://www.hadleyweb.pwp.blueyonder.co.uk/CZP/News.htm), was used for focus stacking. Adobe Lightroom and GNU Image Manipulation Program (GIMP) were used for image processing. Wings were mounted in Permount (Fisher) on glass slides and imaged using a Nanozoomer (Hamamatsu).
RNAi clones were generated using a line carrying a heat shock inducible flip-out actin promoter driving the expression of GAL4 and GFP in clonal tissues (hs-flp;; Act5C>CD2>Gal4, UAS-GFP). L1 larvae were heat shocked for 15 minutes at 37°C and later collected for dissection upon reaching late L3 (wandering) stage. Third instar eye-antennal and wing discs were dissected in Schneider medium, fixed, and stained with DAPI and an anti-MAPK antibody (1/1,000, Cell Signaling number 4695) following the same procedure described above for S2 cells.
The following Drosophila genes are named on the basis of their human counterparts: Fnta (farnesyl transferase alpha; CG2976), Fgop2 (fibroblast growth factor receptor 1 oncogene partner 2; CG10158), Slmap (sarcolemma associated protein; CG17494), Strip (Striatin interacting protein; CG11526). Ras85D (refers to the Drosophila gene encoding RAS), ras (gene encoding an IMP dehydrogenase) is referred to by its full name, “raspberry,” to avoid confusion. Following the nomenclature recommended by Flybase, gene symbols are in lower-cased italics and protein symbols are in upper-case without italics. hyp (hyphen; designates the ave/hyp gene).
|
10.1371/journal.pntd.0005791 | Proteogenomic analysis of the total and surface-exposed proteomes of Plasmodium vivax salivary gland sporozoites | Plasmodium falciparum and Plasmodium vivax cause the majority of human malaria cases. Research efforts predominantly focus on P. falciparum because of the clinical severity of infection and associated mortality rates. However, P. vivax malaria affects more people in a wider global range. Furthermore, unlike P. falciparum, P. vivax can persist in the liver as dormant hypnozoites that can be activated weeks to years after primary infection, causing relapse of symptomatic blood stages. This feature makes P. vivax unique and difficult to eliminate with the standard tools of vector control and treatment of symptomatic blood stage infection with antimalarial drugs. Infection by Plasmodium is initiated by the mosquito-transmitted sporozoite stage, a highly motile invasive cell that targets hepatocytes in the liver. The most advanced malaria vaccine for P. falciparum (RTS,S, a subunit vaccine containing of a portion of the major sporozoite surface protein) conferred limited protection in Phase III trials, falling short of WHO-established vaccine efficacy goals. However, blocking the sporozoite stage of infection in P. vivax, before the establishment of the chronic liver infection, might be an effective malaria vaccine strategy to reduce the occurrence of relapsing blood stages. It is also thought that a multivalent vaccine comprising multiple sporozoite surface antigens will provide better protection, but a comprehensive analysis of proteins in P. vivax sporozoites is not available. To inform sporozoite-based vaccine development, we employed mass spectrometry-based proteomics to identify nearly 2,000 proteins present in P. vivax salivary gland sporozoites. Analysis of protein post-translational modifications revealed extensive phosphorylation of glideosome proteins as well as regulators of transcription and translation. Additionally, the sporozoite surface proteins CSP and TRAP, which were recently discovered to be glycosylated in P. falciparum salivary gland sporozoites, were also observed to be similarly modified in P. vivax sporozoites. Quantitative comparison of the P. vivax and P. falciparum salivary gland sporozoite proteomes revealed a high degree of similarity in protein expression levels, including among invasion-related proteins. Nevertheless, orthologs with significantly different expression levels between the two species could be identified, as well as highly abundant, species-specific proteins with no known orthologs. Finally, we employed chemical labeling of live sporozoites to isolate and identify 36 proteins that are putatively surface-exposed on P. vivax salivary gland sporozoites. In addition to identifying conserved sporozoite surface proteins identified by similar analyses of other Plasmodium species, our analysis identified several as-yet uncharacterized proteins, including a putative 6-Cys protein with no known ortholog in P. falciparum.
| Malaria is one of the most important infectious diseases in the world with hundreds of millions of new cases every year. Malaria is caused by parasites of the genus Plasmodium which have a complex life cycle, alternating between mosquito and mammalian hosts. Human infections are initiated with a sporozoite inoculum deposited into the skin by parasite-infected mosquitoes as they probe for blood. Sporozoites must locate blood vessels and enter the circulation to reach the liver where they invade and grow in hepatocytes. In the case of Plasmodium vivax, one of the two Plasmodium species responsible for the majority of the disease burden in the world, the parasite has the ability to persist for months in the liver after the initial infection and its activation causes the recurring appearance of the parasite in the blood. Though all clinical symptoms are attributable to the blood stages, it is only by attacking the transmission stages before the formation of hypnozoites (the persisting parasites in the liver) that an impact on the burden of vivax malaria can be achieved. We used state-of-the-art mass spectrometry-based proteomics tools to identify the total protein make-up of P. vivax sporozoites. By analyzing which proteins are exposed to the parasite surface and determining the degree of protein’s post-translational modifications, our investigation will aid the understanding of the novel biology of sporozoites and importantly, advise the development of potential vaccine candidates targeting this parasite stage.
| Malaria is a major global infectious disease, responsible for nearly 429,000 deaths and 212 million new cases annually (World Malaria Report 2016, WHO). This disease, found in much of the tropical and subtropical regions of the world, is caused by parasites of the genus Plasmodium, transmitted to humans by the bite of infected anopheline mosquitoes. Parasites (sporozoites) that have infected the mosquito salivary gland are transmitted to the human host as the mosquito injects saliva while taking a blood meal. These sporozoites find their way to the liver where they invade hepatocytes and reproduce asexually. The mature liver stages rupture and the release of exoerythrocytic merozoites that are ready to invade erythrocytes causes the clinical symptoms of malaria.
The majority of human malaria cases are caused by P. falciparum and P. vivax. A large proportion of malaria research efforts focus on P. falciparum infections, motivated by the severity of clinical symptoms and the high mortality rate that is especially evident among children in sub-Saharan Africa. In contrast, P. vivax malaria affects more people in a wider global range [1, 2], but infections with P. vivax often do not cause disease that matches the severity observed for P. falciparum infections. P. vivax-infected individuals of all age groups may still endure repeated, debilitating febrile attacks, severe anemia, and respiratory distress that are more frequently fatal than previously appreciated [3]. Additionally, with only one exposure to infectious mosquito bite, P. vivax can initiate not only one symptomatic infection but a series of reoccurring onsets of malaria episodes that, if not treated, can last for months. These recurring infections are due to a distinctive property of P. vivax liver infection: formation of hypnozoites, a portion of P. vivax liver-stage parasites that becomes dormant and can reactivate weeks to months or even years later [4].
The malaria elimination strategies of vector control and treating symptomatic blood-stage infection with anti-malarial drugs are not as effective against P. vivax as against P. falciparum because P. vivax can persist in the liver as dormant hypnozoites, and because P. vivax gametocytes develop earlier and can be transmitted before onset of clinical symptoms [5–7]. Currently, the only approved treatment for P. vivax is primaquine. Primaquine, however, comes with major complications: its short half-life translates to long dosage regimens, its toxicity for patients with glucose-6-phosphate-dehydrogenase deficiency requires pre-screening of recipients [8], and limited effectiveness in patients with certain cytochrome P450 2D6 polymorphisms will require consideration [9].
An alternative route to reducing the burden of vivax malaria would be the development of an effective vaccine against P. vivax. Targeting P. vivax pre-erythrocytic stages (the sporozoite stage and the liver stage) for vaccine development not only has the advantage that these initial stages of infection involve only a small number of parasites and are completely asymptomatic, but also that such a vaccine could prevent relapsing infections. In fact, one of the most effective experimental vaccination strategies against P. falciparum infection is the use of live attenuated sporozoites (damaged by irradiation) that are effective in inducing complete immune protection by their ability to mount humoral and cellular immune responses against the sporozoite and the liver stage of the parasite [10]. This method of vaccination was recently tested in P. vivax showing encouraging protective efficacy [11]. Nevertheless, the major obstacle for a successful pre-erythrocytic vaccine lies in the required threshold for vaccine efficacy: to protect against infection, the pre-erythrocytic vaccine must be completely effective. Full development of a single liver-stage parasite and exoerythrocytic merozoite release results in full-blown blood stage infection and all the clinical consequences of the disease. This requirement–inducing sterile immunity by targeting the liver stages–may be especially difficult to achieve for vaccines targeting the liver stages of P. vivax due to the ability of the parasite to form hypnozoites. It is presently unknown if vaccination regimens that target developing liver stages would also be able to target hepatocytes harboring hypnozoites.
Thus, an effective subunit vaccination strategy that targets the parasite before it enters the hepatocyte could be the most plausible solution for preventing a P. vivax hepatocyte infection, development of liver-stage parasites, and hypnozoite formation. It has been shown that antibody responses against the circumsporozoite protein (CSP), a major surface protein on the Plasmodium sporozoite, can lead to sterile protection against infection, but in most cases these responses offer only partial protection in P. falciparum [12]. A recent clinical study in which a P. vivax CSP-based subunit vaccine was used showed no sterile protection, but a significant delay in the onset of parasitemia was observed [13]. As opposed to P. falciparum infection where partial protection offers only limited benefits, partial protection that could be observed after immunizations against P. vivax has the potential to considerably affect the hypnozoite burden in the liver by limiting the number of sporozoites reaching the hepatocyte and developing into hypnozoites, thereby directly decreasing the chances of relapse malaria [14]. A vaccine targeting P. vivax sporozoites is therefore highly desirable.
The identification of non-CSP antigens that can be included into a multi-antigen subunit vaccine has recently gained momentum for P. falciparum, but such an effort has not yet been initiated for P. vivax. After mosquito transmission, sporozoites embark on a complex route of infection in the human host and three biological activities of the sporozoite are essential for their success, namely, gliding motility, cell traversal, and cell invasion. All of these activities require engagement of sporozoite surface and secreted proteins with the host environment and thus might be blocked by antibodies. Therefore, the discovery of new P. vivax sporozoite surface antigens, together with CSP-based antigens, may allow the development of a better antibody-based, anti-infection vaccine [15].
Mass spectrometry (MS)-based proteomics has previously been employed to catalogue the protein complement of P. falciparum, P. yoelii and P. berghei salivary gland sporozoites with the goal of identifying new targets for therapeutics and new antigens for subunit-based vaccines [16–20]. The most comprehensive proteomic analyses to-date of P. falciparum salivary gland sporozoites detected over 2000 of the approximately 5000 gene products predicted from the Plasmodium falciparum genome [19] and identified 42 putatively surface-exposed sporozoite proteins [20] by a chemical labeling strategy. Here, by applying similar proteomics techniques to the analysis of the proteins present in P. vivax salivary gland sporozoites of field isolates, a combined total of 1970 P. vivax proteins were identified, of which 36 have been categorized as putative sporozoite surface proteins. Post-translational modification of sporozoite proteins by glycosylation and phosphorylation have also been evaluated to further aid the development of subunit vaccines.
The human blood collection protocol was approved by the Ethical Committee of the Faculty of Tropical Medicine, Mahidol University. All adult subjects participating in this study provided written informed consent. No child participants were included in this study.
Anopheles dirus mosquitoes (from the Mahidol University colony maintained at the Faculty of Tropical Medicine laboratories) were infected with blood collected from patients who were confirmed positive for only P. vivax malaria via microscopy at local health centers in close proximity to the Kanchanaburi Campus, Mahidol University. In brief, 150 μL of red blood cell pellet from blood samples was suspended in pooled normal AB serum to a packed cell volume of 50%. The suspension was fed for 30 min to 100 female mosquitoes (5-7 days old) via an artificial membrane attached to a water-jacketed glass feeder maintained at 37°C. Unfed mosquitoes were removed and fed mosquitoes were maintained on a 10% w/v sucrose solution and incubated at 26°C and 80% humidity for at least 14 days. Salivary gland dissections were performed at days 14-19. CSP haplotype (VK210 or VK247) was determined by PCR.
Salivary glands from P. vivax-infected mosquitoes were harvested by microdissection and homogenized by grinding. Sporozoite preparations were purified from mosquito debris on an Accudenz discontinuous gradient as previously described [21]. Total sporozoite numbers were counted on a hemocytometer. For the total proteome analyses, 3.5 × 106 VK210 and 4.5 × 106 VK247 sporozoites were pelleted for 3 min at 16,000 × g, re-suspended in 1 × PBS pH 7.4, pelleted, and stored at -80°C. Prior to protein separation by SDS-PAGE, the pellet was re-suspended in an equal volume of 2 × sample buffer and heated for 5 min at 70°C. For the surface proteome samples, 2 × 106 VK210 and 1.8 × 107 VK247 sporozoites were pelleted for 3 min at 4,000 × g at 4°C and re-suspended with ice-cold 1 × PBS pH 8.0. The VK247 sample was evenly split into two tubes and all three samples were pelleted again. One of the VK247 samples was set aside as an unlabeled control. The remaining two samples were re-suspended in 40 μL ice-cold 1 × PBS pH 8.0 per 106 sporozoites. A 10 mM stock solution of EZ-Link Sulfo-NHS-SS-Biotin (Thermo Fisher Scientific, part number 21331) was added to a final concentration of 2 mM and the samples were incubated for 1 h at 4°C. The sporozoites were pelleted and re-suspended in 500 μL ice-cold 1 × Tris-buffered saline (TBS) pH 8.0 and incubated for 5 min on ice to quench excess biotin label. The sporozoites were then pelleted for 2 min at 16,000 × g and washed a second time in 1 × TBS, each time removing as much supernatant as possible without disturbing the pellet. The samples were stored at -80°C until lysis. The sporozoites were lysed by re-suspending the pellet in 100 μL lysis buffer (1% w/v SDS, 4 M urea, 50 mM Tris-HCl pH 8.0, 150 mM NaCl, 1 × protease inhibitor (Roche cOmplete)) and incubating for 30 min at 4°C with end-over-end rotation. The samples were diluted to 1 mL in 1 × PBS pH 7.4, added to 1 mg of magnetic streptavidin beads (Dynabeads MyONe Streptavidin T1) which had been washed three times in 1 × PBS, and incubated for 1 h at 4°C with end-over-end rotation. The beads were washed sequentially with the following: 1) 2% w/v SDS; 2) 0.1% w/v SDS, 6 M urea, 1 M NaCl, 50 mM Tris-HCl pH 8.0; 3) 0.1% w/v SDS, 4 M urea, 200 mM NaCl, 1 mM EDTA, 50 mM Tris-HCl pH 8.0; 4) 0.1% w/v SDS, 50 mM NaCl, 50 mM Tris-HCl pH 8.0. Bound proteins were eluted by adding 40 μL 2 × sample buffer (4% w/v SDS, 125 mM Tris-HCl pH 6.8, 20% v/v glycerol, 0.02% w/v bromophenol blue) to which dithiothreitol (DTT) was added to a final concentration of 50 mM and heating the tube for 7 min at 70°C. The eluted sample was transferred to a new tube and stored at -80°C until separation by SDS-PAGE.
SDS-PAGE pre-fractionation and in-gel tryptic digestion were performed essentially as described in [19]. Extended methods are provided in S1 File. Briefly, samples were electrophoresed through a 4-20% w/v SDS-polyacrylamide gel (Pierce Precise Tris-HEPES). Gels were stained with Imperial Stain (Thermo Fisher Scientific), de-stained in Milli-Q Water (Millipore), and cut into fractions (S1 Table). Gel pieces were then de-stained with 50 mM ammonium bicarbonate (ABC) in 50% acetonitrile (ACN) and dehydrated with ACN. Disulfide bonds were reduced with 10 mM DTT and cysteines were alkylated with 50 mM iodoacetamide in 100 mM ABC. Gel pieces were washed with ABC in 50% ACN, dehydrated with ACN, and rehydrated with 6.25 ng/μL sequencing grade trypsin (Promega). The supernatant was recovered and peptides were extracted by incubating the gel pieces with 2% v/v ACN/1% v/v formic acid, then ACN. The extractions were combined with the digest supernatant, evaporated to dryness in a rotary vacuum, and reconstituted in liquid chromatography (LC) loading buffer consisting of 2% v/v ACN/0.2% v/v trifluoroacetic acid (TFA).
LC and MS parameters were essentially as described previously [19, 20]. Extended method details are provided in S1 Table. Briefly, LC was performed using an Agilent 1100 nano pump with electronically controlled split flow or a Proxeon Easy nLC. Peptides were separated on a column with an integrated fritted tip (360 μm outer diameter (O.D.), 75 μm inner diameter (I.D.), 15 μm I.D. tip; New Objective) packed in-house with a 20 cm bed of C18 (Dr. Maisch ReproSil-Pur C18-AQ, 120 Å, 3 μm). Prior to each run, sample was loaded onto a trap column consisting of a fritted capillary (360 μm O.D., 150 μm I.D.) packed with a 1 cm bed of the same stationary phase and washed with loading buffer or buffer A (0.1% v/v formic in water). The trap was then placed in-line with the separation column for the separation gradient. The LC mobile phases consisted of buffer A and buffer B (0.1% v/v formic acid in ACN). The separation gradient was 5% B to 35% B over 60 min for the surface-labeled samples and 5% B to 25% B over 120 or 180 min for the whole proteome samples. Tandem MS (MS/MS) was performed with an LTQ Velos Pro-Orbitrap Elite (Thermo Fisher Scientific). Data-dependent acquisition was employed to select the top precursors for collision-induced dissociation (CID) and analysis in the ion trap. Dynamic exclusion and precursor charge state selection were employed. Two nanoLC-MS technical replicates were performed for each fraction, with roughly half the available sample injected for each replicate.
The MS data generated for this manuscript, along with the search parameters, analysis parameters and protein databases can be downloaded from PeptideAtlas (www.peptideatlas.org) using the identifiers PASS00976 (whole proteome) and PASS00977 (surface-labeled). Mass spectrometer output files were converted to mzML format using msConvert version 2.2.0 (whole proteome data) or 3.0.5533 (surface-labeled data) [22] and searched with Comet version 2015.02 rev.0 [23]. The protein sequence database is described in the following section. The precursor mass tolerance was ±10 ppm, and fragment ions bins were set to a tolerance of 1.0005 m/z and a monoisotopic mass offset of 0.4 m/z. Semi-tryptic peptides and up to 2 missed cleavages were allowed. The search parameters included a static modification of +57.021464 Da at Cys for formation of S-carboxamidomethyl-Cys by iodoacetamide and potential modifications of +15.994915 Da at Met for oxidation, -17.026549 Da at peptide N-terminal Gln for deamidation from formation of pyroGlu, -18.010565 Da at peptide N-terminal Glu for loss of water from formation of pyroGlu, -17.026549 Da at peptide N-terminal Cys for deamidation from formation of cyclized N-terminal S-carboxamidomethyl-Cys, and +42.010565 for acetylation at the N-terminus of the protein, either at N-terminal Met or the N-terminal residue after cleavage of N-terminal Met. Additionally, the search parameters for sporozoite surface samples included a potential modification of +145.019749 Da at Lys for addition of the biotin label, the disulfide bond of which had been cleaved and alkylated. The MS/MS data were analyzed using the Trans-Proteomic Pipeline (TPP) [24] version 5.0.0 Typhoon. Peptide spectrum matches (PSMs) were assigned scores in PeptideProphet [25], peptide-level scores were assigned in iProphet [26], and Protein identifications were inferred with ProteinProphet [27]. Additional TPP parameters are available in S1 File. In the case that multiple proteins were inferred at equal confidence by a set of peptides, the inference was counted as a single identification and all relevant protein IDs were listed. Only proteins with ProteinProphet probabilities corresponding to a false discovery rate (FDR) less than 1.0% (as determined from the ProteinProphet mixture models) were reported. For comparison with P. falciparum salivary gland sporozoites, a publically available data set [19] (available from PeptideAtlas using the identifier PASS00095) was re-analyzed with the same software and parameters described above. The spectra were searched against a database comprising P. falciparum 3D7 [28] (PlasmoDB v.30, www.plasmodb.org [29]), Anopheles stephensi Indian AsteI2.3 [30] (VectorBase, www.vectorbase.org [31]), and a modified version of the common Repository of Adventitious Proteins (v.2012.01.01, The Global Proteome Machine, www.thegpm.org/cRAP) with the Sigma Universal Standard Proteins removed and the LC calibration standard peptide [Glu-1] fibrinopeptide B appended. Decoy proteins with the residues between tryptic residues randomly shuffled were generated using a tool included in the TPP and interleaved among the real entries. P. falciparum protein annotations were updated from PlasmoDB v.32.
A protein database containing sequence polymorphisms of P. vivax proteins occurring in Thailand was created by aligning DNA-seq and RNA-seq reads from field isolates to the P. vivax Sal-1 genome [32] (PlasmoDB v.26). DNA-seq reads from 19 Thai field isolates were obtained from www.plasmodb.org and aligned using Burrows Wheeler Aligner (v.0.7.12) and SNVs were called using the Genome Analysis Toolkit (v.3.6). RNA-seq reads for 13 isolates [33] (obtained from https://www.ncbi.nlm.nih.gov/bioproject/, accession number PRJNA376620) were aligned using STAR (v.2.5) and SNVs were called using the Genome Analysis Toolkit (v.3.6). All proteins with sequences different from the Sal-1 reference proteome were compiled (S2 File) and added to a database comprising P. vivax Sal-1 [32] (PlasmoDB v.31), P. vivax P01 [34] (PlasmoDB v.31), Anopheles stephensi Indian AsteI2.3 [30] (VectorBase), and the modified version of the cRAP proteins described above. Decoy proteins were generated as above. Mass spectra from the two whole-proteome samples and the two surface-labeled samples were searched against the database with Comet as described above and the resulting PSMs were analyzed with PeptideProphet and iProphet as described above except that the NSP model was enabled in iProphet. All P. vivax peptides identified with iProphet probabilities corresponding to a model-estimated FDR less than 1.0% were aligned against the P. vivax P01 reference proteome. A new P. vivax P01 reference proteome was assembled incorporating polymorphism-bearing peptides identified by the above-described search. If a detected peptide was associated with a given P. vivax protein in at least one of the field isolates but did not align with the P. vivax P01 reference sequence due to sequence polymorphisms, the variant peptide sequence was appended to the end of the reference protein sequence entry in the fasta database. This modified P. vivax P01 reference proteome was added to the An. stephensi and cRAP databases described above. Additionally, the entry for CSP (PVP01_0835600), which contains the tandem repeat region specific to the VK210 haplotype, was appended with the sequence of the VK247 tandem repeat region [35]. Decoys were generated as above. This database was used for all subsequent analysis of the MS data. P. vivax protein annotations were updated from PlasmoDB v.32.
Relative protein abundance within and between samples was estimated using a label-free proteomics method based on spectral counting. Extended method details are provided in S1 File. The spectral counts for a protein were taken as the total number of high-quality PSMs (identified at a PeptideProphet probability corresponding to an FDR less than 1.0%) that identified the protein. PSMs from degenerate peptides (peptides whose sequences were found in multiple proteins in the database) were split among proteins containing that peptide in a weighted fashion [36, 37]. Relative protein abundance within samples was ranked using the normalized spectral abundance factor. The spectral abundance factor (SAF) for a given protein was calculated as the quotient of the total PSMs identifying that protein and the protein's length. The SAF for each Plasmodium protein was normalized to the sum of all Plasmodium SAF values obtained from the same sample, and this normalized SAF (NSAF) was natural log-transformed to ln(NSAF) [38, 39]. The population of ln(NSAF) values for each sample assumed a normal distribution, as did the population of log-transformed protein abundance fold-change ratios between samples, calculated as ln(NSAF)A-ln(NSAF)B where A and B are two different samples in which the same protein was observed. Each of these distributions was fit with a Gaussian curve in Microsoft Excel using minimum residual sum of squares and goodness-of-fit was evaluated with the R2 coefficient of determination (S1 and S2 Figs). To assess the relative abundance of proteins between the two samples, PSM counts for all proteins were first increased by 1 in order to assign non-zero values to proteins detected in one sample but not the other [40]. These adjusted spectral counts were then normalized so that the sum of all PSMs was the same in both samples. The abundance ratio for a given protein between a two samples was then calculated as
RA:B=cAcB
Where RA:B is the protein abundance ratio of a protein between sample A and sample B and cA and cB are the adjusted and normalized spectral counts for the protein in sample A and sample B, respectively. In order to assess the error in spuriously large protein ratios obtained from proteins with low spectral counts, the G-test of significance was applied to the adjusted and normalized spectral counts for each protein pair as
G=2[cAln(cA(cA+cB2))+cBln(cB(cA+cB2))]
and a p-value was assigned by calculating the probability that a χ2 distribution with one degree of freedom was more extreme than the G statistic [40]. The distribution of the log2(RA:B) values of all proteins detected in both samples was fit with a Gaussian curve as above. Protein abundance ratios were corrected for systematic bias by subtracting the mean of this distribution (which was near 0 in all cases) from each log-transformed protein ratio. In order to assess the probability that a protein ratio was more extreme than the normal distribution of protein ratios, a p-value was calculated for each ratio using the complementary error function as
p=ERFC|log2(RA:B)−μσ2|
where μ is the mean and σ is the standard deviation of the fit Gaussian. The FDR arising from multiple hypothesis testing was assessed by the Benjamini-Hochberg method for both tests independently, and protein ratios with an FDR less than 5.0% by both the G-test and ERFC were considered significant.
Phosphorylated peptides were identified by searching the MS data with the same parameters listed above with the additional potential modification mass of +79.966331 Da at Ser, Thr, and Tyr. The PSMs generated from these searches were analyzed separately by PeptideProphet as above, except that the DECOYPROBS option was used so that decoy peptides were assigned probabilities and included in the output. Decoy peptides were used to calculate an FDR among the subset of PSMs exhibiting phosphorylation. Due to the infrequent occurrence of phosphopeptides in these un-enriched samples, the decoy-estimated FDR for phosphopeptide PSMs was as high as 24% in the VK210 sample and 19% in the VK247 sample at the probabilities corresponding to a 1.0% decoy-estimated FDR for the entire population of PSMs. The more stringent cut-off to achieve 1.0% FDR among phosphopeptide PSMs was used to identify high-confidence phosphopeptides. Phosphopeptide PSMs within each sample were only counted if the phosphopeptide was identified by at least one PSM at the high-stringency cut-off and by at least two PSMs in the population-level cut-off. The number of PSMs identifying a phosphopeptide and the number of PSMs identifying the same peptide in un-phosphorylated form were used to estimate the percentage of that peptide that was phosphorylated in the sample. Localization of phosphate groups within phosphopeptides was confirmed and/or corrected using a development version of PTMProphet (source code available at https://sourceforge.net/p/sashimi, SVN revision number 7584. See S1 File for complete parameters).
Experimental and theoretical evidence was used to identify high-confidence putatively surface-exposed proteins from among those P. vivax proteins identified by surface labeling live sporozoites with the biotin tag. Proteins were taken for further consideration if they were identified by at least two peptides and three PSMs in at least one of the two labeled samples. Proteins were considered high-quality candidates if they possessed predicted characteristics of a surface protein, i.e., transmembrane (TM) domain(s), a signal peptide, or a glycophosphatidylinositol (GPI) anchor, or if they exhibited spectral evidence for incorporation of the biotin label. Theoretical evidence for presence of surface protein characteristics was determined from protein primary sequences using established tools: the number of predicted TM domains was obtained from THMM2 [41] via PlasmoDB.org (P. vivax P01 v.31), presence of a signal peptide was predicted by SignalP version 4.1 [42] (http://www.cbs.dtu.dk/services/SignalP/) and presence of a glycosylphosphatidylinositol (GPI) anchor was predicted using PredGPI [43] (http://gpcr2.biocomp.unibo.it/gpipe/index.htm). A protein was considered to have spectral evidence for labeling if a non-degenerate component peptide displaying the addition of the biotin tag was identified from at least one high-quality PSM (PeptideProphet probability corresponding to an FDR less than 1.0%). Non-specific binding to the streptavidin beads was assessed by comparing the VK247 labeled and unlabeled samples, which were split from the same sample and processed in parallel with or without labeling. In order to identify those proteins with the highest value as potentially surface-exposed targets based on the theoretical and experimental evidence, proteins were assigned priority tiers (1 being highest) as follows: 1) possessing predicted TM domain(s), signal peptide or GPI anchor and exhibiting spectral evidence of incorporation of the biotin tag; 2) exhibiting spectral evidence of incorporation of the biotin tag but lacking predicted TM domain(s), signal peptide or GPI anchor; 3) possessing predicted TM domain(s), signal peptide or GPI anchor but lacking spectral evidence of incorporation of the biotin tag; 4) lacking predicted TM domain(s), signal peptide or GPI anchor as well as lacking spectral evidence of incorporation of the biotin tag. Tiers one, two and three were considered high-quality candidate surface proteins. Proteins identified from fewer than two peptides and three PSMs in at least one sample were not assigned priority tiers.
MS-based proteomics was used to survey the proteins present in P. vivax salivary gland sporozoites. Two independent sporozoite samples were obtained from mosquitoes fed on blood obtained from volunteers who presented with clinical malaria at treatment centers in Thailand. Peptide spectrum matches were analyzed using the Trans-Proteomic Pipeline [24]. Proteins identified with scores corresponding to an FDR less than 1.0% were reported. A total of 1711 P. vivax proteins were identified from 3.5 × 106 sporozoites bearing the VK210 haplotype of circumsporozoite protein (CSP), of which 1492 (87.2%) were identified by at least two non-degenerate peptides. A total of 1747 P. vivax proteins were identified from 4.5 × 106 sporozoites bearing the VK247 CSP haplotype, of which 1572 (90.0%) were identified by at least two non-degenerate peptides. A combined total of 1970 P. vivax proteins were identified from the two samples, of which 1733 (88.0%) were identified from at least two non-degenerate peptides in at least one of the samples. Of the combined 1970 P. vivax proteins identified, 1488 (75.5%) were identified in both samples (S2 Table). Label-free protein quantification based on spectral counts was used to compare protein abundance between the two samples. NSAF, a technique that normalizes spectral counts for protein length and sample complexity, was used to compare relative protein abundance within and between the two samples, while protein abundance ratios between the two samples were tested for significance using the G-test as well as information extracted from the normal distribution of protein ratios. Comparing the protein abundances showed largely similar protein composition and protein abundance (Fig 1). The proteins identified in both samples included all of the proteins in the top quartile of abundance in each sample and 968 of 983 proteins (98.5%) in the upper half of abundance in each sample. Furthermore, 218 of 223 (97.8%) of the proteins unique to the VK210 sample were in the lower half of abundance, with 155 (69.5%) in the bottom quartile. Likewise, of the proteins unique to the VK247 sample, 249 of 259 (96.1%) were in the lower half of abundance, with 192 (74.1%) in the bottom quartile. These results suggest that differences in proteins detected between the two samples arose primarily from technical issues affecting limit-of-detection rather than unique protein expression in one sample or the other. Likewise, observed differences in relative protein abundance observed between the two samples were likely predominantly technical in origin rather than biological. The populations of ln(NSAF) values from the two field isolate samples could be fitted with Gaussian curves with similar means and variance, and the population of log-transformed abundance ratios for proteins detected in both samples assumed a normal distribution with a mean near zero, i.e., a protein ratio of essentially 1:1 (S1 Fig). Fitting the population of ratios to a Gaussian allowed measurement of the deviation from the mean of 1:1, which was low (less than one standard deviation) for high-abundance proteins and generally increased at lower protein abundances (Fig 1A), a known phenomenon of spectral counting [39, 44]. To identify proteins with significantly different abundances between the two samples, a likelihood ratio test (G-test) was applied to the protein ratios obtained from comparing spectral counts of each protein as observed in the two samples [40]. All spectral counts were increased by 1 in order to obtain ratios for proteins observed in only one sample [40]. Additionally, a Gaussian curve was fit to the distribution of log-transformed abundance ratios for proteins observed in both samples (Fig 1B) and the complementary error function was used to obtain a p-value indicating the probability that the normal distribution was more extreme than any give protein ratio. Combining these two tests identified protein ratios that deviated significantly from the mean while accounting for the increased quantification error at low spectral counts (Fig 1C). After correcting for multiple hypothesis testing by the Benjamini-Hochberg procedure, 119 proteins were identified with p-values corresponding to an FDR less than 5.0% by both methods. Of these, 35 were identified in both samples (2.4% of all proteins identified in both samples) and 84 were identified only in one sample or the other (17% of all proteins identified in only one of the two samples; S2 Table).
In order to compare salivary gland sporozoite proteomes of P. vivax and P. falciparum, a previously published proteomic analysis of P. falciparum salivary gland sporozoites [19] was re-analyzed with the same informatics pipeline used here, identifying 2010 proteins, of which 1798 (89.5%) were identified by at least two peptides (S3 Table). The same quantitative approach described above was used to compare the relative abundance of protein orthologs between the two species. The spectral counting methods were expected to be less accurate when comparing orthologs between species than when comparing the same proteins detected in different samples of the same species because, all else being equal, two protein orthologs with sufficiently different sequences could produce different numbers of PSMs due to differences in the number of tryptic peptides produced and the detectability of these peptides by LC-MS determined by sequence-specific chemical properties. Nonetheless, there was a large overlap in both protein detection and relative protein abundance between protein orthologs detected in the P. vivax and P. falciparum samples (Fig 2). Of the all the proteins detected in either of the P. vivax samples or the P. falciparum sample, 2314 had annotated orthologs in both P. falciparum and P. vivax, and 1609 of these (69.5%) were detected in the sporozoite samples of both species analyzed here. As with the comparison between the two P. vivax samples, the population of log-transformed ratios of proteins identified in both P. vivax and P. falciparum had a mean near 1:1 with little deviation from the mean among the high-abundance proteins and increasing deviation at low spectral counts. Most of the protein orthologs identified in one species and not the other were low-abundance proteins. Of the 332 orthologs not detected in the P. falciparum sample, 300 (90.4%) were in the lower half of abundance and 189 (56.9%) were in the bottom quartile of abundance. Of the 373 orthologs not detected in the P. vivax samples, 325 (87.1%) were in the lower half of abundance and 224 (60.1%) were in the bottom quartile of abundance (S3 Table, Fig 2A). The most highly abundant proteins detected in the P. vivax sporozoites were also highly abundant in P. falciparum sporozoites, including several with critical roles in invasion, e.g., CSP, thrombospondin-related anonymous protein (TRAP; PVP01_1218700, PF3D7_1335900), gamete egress and sporozoite traversal protein (GEST; PVP01_1258000, PF3D7_1449000), cell traversal protein for ookinetes and sporozoites (CelTOS; PVP01_1435400, PF3D7_1216600), apical membrane antigen 1 (AMA1; PVP01_0934200, PF3D7_1133400), sporozoite invasion-associated protein 1 (SIAP1; PVP01_0307900, PF3D7_0408600), and sporozoite protein essential for cell traversal (SPECT1 PVP01_1212300, PF3D7_1342500) (Table 1). Even so, a number of high-abundance proteins were identified that were of significantly higher abundance in one species than the other (S3 Table). For example, PVP01_0314600 (conserved Plasmodium protein, unknown function) was in the top decile of abundance in both P. vivax sporozoite samples, while its syntenic ortholog PF3D7_0718900 (conserved Plasmodium protein, unknown function) was not detected in the P. falciparum sporozoite sample, or for that matter, in any of the P. falciparum proteomics datasets on PlasmoDB spanning the entire P. falciparum lifecycle. In the P. falciparum sporozoite sample, two conserved Plasmodium proteins of unknown function, PF3D7_0215200 and PF3D7_0410500, were in the top decile of protein abundance but not detected at all in either P. vivax sample. Both proteins are up-regulated in P. falciparum salivary gland sporozoites based on transcriptomic and proteomic data compiled at PlasmoDB.org. In addition to differentially expressed orthologs, a number of identified proteins had no orthologs in the other species compared. Of the combined 1970 P. vivax proteins identified, 29 (1.47%) had no P. falciparum ortholog. These included three proteins annotated as PIR proteins (Plasmodium interspersed repeats, species-specific immunovariant proteins [34, 45]) and 17 unannotated proteins (i.e., “conserved Plasmodium protein, unknown function”). The most abundant P. vivax protein with no P. falciparum ortholog identified in the samples was a putative 6-Cys domain protein (PVP01_0303900). This protein was in the top decile of abundance in both P. vivax samples, and is putatively surface-exposed on salivary gland sporozoites (discussed below).
Mass spectra were searched against the P. vivax P01 reference proteome [34] (PlasmoDB v.31[29]). Current high-throughput MS approaches require a precise knowledge of the genome of the organism under study; a protein can only be identified if its exact sequence is contained in the database against which the mass spectra are searched. Because the samples were obtained from field isolates and not laboratory strains, they were expected to contain protein sequence polymorphisms that would not be represented in the reference proteome. In order to increase the likelihood of identifying isolate-specific polymorphisms, the P. vivax protein database against which the mass spectra were searched was augmented with potential polymorphisms obtained from genomic and transcriptomic analyses of Thai P. vivax field isolates. A total of 13 RNA-seq and 19 DNA-seq data sets were aligned against the P. vivax Sal1 reference genome and a reference proteome was generated containing any protein with an amino acid sequence differing from the reference. Only 22% of the proteins in the reference proteome had completely conserved sequences across all 33 datasets (the 32 field isolates plus the reference genome). Over 50% of the proteins had four or more unique amino acid sequences arising from various combinations of sequence polymorphisms, and 10% had 15 or more unique sequences. One protein, RNA pseudouridylate synthase (PVX_080660) had a unique sequence in all 33 genomes aligned (S4 Table). These P. vivax Sal-1 variants and the P. vivax Sal-1 reference proteome were appended to the P. vivax P01 reference proteome and used to identify polymorphisms in the analyzed samples. A total of 301 identified P. vivax proteins contained polymorphisms that were not present in the P. vivax P01 reference proteome (S5 Table). The four identified P. vivax salivary gland sporozoite proteins exhibiting the most polymorphisms not present in the P. vivax P01 reference proteome were surface proteins: AMA1 (PVP01_0934200), TRAP-like protein (TLP; PVP01_1132600), TRAP (PVP01_1218700) and GPI-anchored micronemal protein (GAMA; PVP01_0505600). Each of these proteins also exhibited a high degree of polymorphism in the compared field isolate genomes (95th, 89th, 98th, and 97th percentiles, respectively, of the number of unique protein sequences arising from polymorphisms among the compared genomes). Except for seven proteins identified from a single peptide, all of these polymorphism-bearing proteins could be detected without the additional knowledge of polymorphisms obtained from the field isolates. However, failing to detect peptides would have led to increased errors in protein quantification by spectral counts. Furthermore, knowledge of non-synonymous substitutions in the P. vivax genome was critical to accurately characterizing proteins detected in the samples. For example, the P. vivax haplotype designations VK210 and VK247 are based on differences in the sequence of the repeat region of CSP. The VK210 haplotype bears tandem repeats of the sequence GDRA(D/A)GQPA, while the VK247 haplotype bears tandem repeats of the sequence ANGA(G/D)(N/D)QPG. In the VK247 whole proteome analyzed here, the repeat region of CSP was poorly detected due to a lack of Lys or Arg residues that would result in tryptic peptides. However, hundreds of PSMs identified a tryptic peptide at the C-terminal end of the tandem repeat region which is distinct in VK247 [35], and no independent evidence was observed for VK210-specific peptides (S3 Fig, S5 Table). Conversely, in the VK210 sample, peptides covering the entire CSP tandem repeat region were identified from hundreds of PSMs, owing to the presence of regularly interspersed Arg tryptic cleavage sites. Interestingly, the VK210 sample appeared to contain a mixed infection of at least two distinct field isolates. The same discriminating peptide at the C-terminal end of the tandem repeat region was identified by hundreds of PSMs for semi-tryptic fragments of various lengths containing the VK210-specific sequence found in the P. vivax P01 reference proteome. However, a semi-tryptic variant of the peptide found in the P. vivax Sal-1 version of CSP was also identified, as were semi-tryptic peptides matching portions of the VK247 version of the peptide. There was not enough independent evidence to determine if the VK247 haplotype was present in the sample (S3 Fig, S5 Table). Evidence for a mixed infection was also found in TRAP. Seven sequence polymorphisms not present in the P. vivax P01 reference proteome were identified in TRAP in the samples analyzed, four of which were present in the P. vivax Sal-1 reference proteome and three of which were only found in field isolates. As was observed for CSP, the VK247 sample appeared to contain a single haplotype of TRAP, while there were at least two haplotypes of TRAP in the VK210 sample (S4 Fig, S5 Table). In addition to accurate quantification and correct characterization of proteins, knowledge of sample-specific polymorphisms was critical to identifying post-translational modifications (discussed below).
It was recently shown that CSP and TRAP are glycosylated in P. falciparum salivary gland sporozoites [20]. Here we report that these proteins are similarly modified in P. vivax sporozoites. The motif CX2-3(S/T)CXXG in thrombospondin repeat (TSR) domains can be modified with an O-linked fucose at the Ser/Thr [46], and this fucose can be further modified with glucose to produce a β1,3-linked disaccharide [47, 48]. Additionally, the WXXW and WXXC motifs of TSR domains can be modified with a C-linked mannose at Trp [49, 50]. These potential glycosylation motifs are present in the TSR domains of both CSP and TRAP in all Plasmodium species. The TSR domain of P. vivax CSP contains the tryptic peptide ATVGTEWTPCSVTCGVGVR with potential O-fucosylation and C-mannosylation sites. Modification of the peptide with O-linked glycans could not be directly detected by the spectral search engines due to the fact that O-linked glycans are highly labile in the gas phase [51, 52] and are lost during collision-induced dissociation (CID) used to generate the identifying fragment spectra. However, as was previously demonstrated with P. falciparum salivary gland sporozoites [20], it was possible to infer the presence of the O-linked glycan through manual interpretation of the mass spectra (S5 and S6 Figs). The analysis showed that this peptide was modified with a mass matching that of an O-linked deoxyhexose. No evidence for C-mannose was observed (Fig 3). While neither the identity of the deoxyhexose nor its attachment site in the peptide could be determined from the data, we presume it to be a fucose attached to the C-terminal Thr based on knowledge of the sugars and enzymes present in Plasmodium [53], TSR domains in other species, and the fact that this residue has been shown to be O-fucosylated in crystal structures of PfCSP expressed in mammalian cells [54]. Evidence for O-fucosylation of CSP was observed in both samples. Based on the signal intensity of the LC peaks, it appears that the majority of CSP (~90%) was glycosylated while a portion was unmodified (S5 and S6 Figs). In P. falciparum sporozoites, the majority of CSP was also observed to be modified with a single deoxyhexose while a small portion was unmodified, though some CSP was also observed to be further modified with an additional hexose, consistent with O-linked fucose-β1,3-glucose [20]. No evidence for modification of CSP with a disaccharide was observed in these P. vivax sporozoite samples.
The TSR domain of P. vivax TRAP contains the tryptic peptide VANCGPWDPWTACSVTCGR which includes potential O-fucosylation and C-mannosylation motifs. Critically, the TRAP in the VK247 sample and some of the TRAP in the mixed-infection VK210 sample exhibited an Arg→Lys substitution at this peptide. Knowledge of this polymorphism was only obtained from the field isolate genomes, so lacking that data would have prevented detecting glycosylation in the samples bearing the substitution. As with CSP, TRAP was observed to be modified with a gas-phase labile modification (S7 and S8 Figs) which was presumed to be O-fucose attached at the C-terminal Thr, again based on the TSR motif as well as crystal structures of PvTRAP and PfTRAP expressed in mammalian cells [55]. C-mannosylation of the WDPWTAC sequence was not observed (Fig 1), even though in P. falciparum sporozoites the C-terminal Trp of WDEWSPC was modified with a mass matching that of hexose, likely C-mannose [20]. Based on chromatographic peak areas, virtually all TRAP in both samples was completely glycosylated (S7 and S8 Figs).
The MS data were further analyzed for evidence of protein phosphorylation, a reversible PTM that is often involved in signaling and control of cellular function. Proteomic analysis of phosphoproteins has been performed for asexual stages of P. falciparum [56–60] but not sporozoites. Typical phosphoproteomic analyses employ affinity techniques to enrich for phosphorylated peptides prior to LC-MS. While that approach was not feasible for this study due to the limited sample material available, it was still possible to detect the presence of this modification in proteins that were highly abundant and/or heavily modified in the samples. Evidence for phosphorylation was found for a total of 139 proteins in either or both of the samples (S6 and S7 Tables). Among the detected phosphoproteins with GO terms, the most prevalent functional class was proteins with DNA or RNA-binding activity (21.6% of the phosphoproteins) and the second most prevalent class was proteins with ATP activity, e.g., ATP binders, kinases and phosphatase (20.9% of the phosphoproteins). Also well-represented were components of the gliding machinery, including Myosin A (MyoA; PVP01_1212200), the glideosome-associated proteins GAP40 (PVP01_1018200), GAP45 (PVP01_1440900) and GAPM2 (PVP01_0532000), several inner membrane complex (IMC) proteins, and the calcium-dependent protein kinase CDPK1 (PVP01_0407500). The P. falciparum salivary gland sporozoite data were searched in the same fashion, identifying 91 phosphorylated proteins (S6 and S8 Tables). All but four of these had syntenic orthologs in P. vivax, and 48 of these (55%) were among the phosphoproteins identified from the P. vivax samples. The list of sporozoite phosphoproteins was compared against thirteen proteomic analyses of P. falciparum blood-stage parasites [16, 56–66] (including five analyses of phosphopeptides enriched from blood-stage parasites [56–60]) and three proteomic analyses of P. vivax blood-stage parasites [61, 67, 68] available on PlasmoDB.org. The majority (74%) of phosphoproteins identified from either P. vivax or P. falciparum sporozoites were also identified in phosphorylated form in P. falciparum blood stages (S6 Table). Table 2 lists 16 P. vivax sporozoite phosphoproteins whose orthologs were identified in P. falciparum blood stages but for which no evidence of phosphorylation was observed, potentially representing sporozoite-specific phosphorylation. Table 3 lists 18 P. vivax sporozoite phosphoproteins that were not detected at all (either phosphorylated or unphosphorylated) in proteomic analyses of P. falciparum and P. vivax blood stages, representing known and potentially novel proteins specific to the sporozoite stage.
In order to identify surface-exposed proteins on P. vivax salivary gland sporozoites, a chemical labeling approach was employed based on the recent analyses of the P. falciparum salivary gland sporozoite surface proteome [19, 20]. Live sporozoites were labeled with a membrane-impermeable, amine-reactive tag that covalently labeled solvent-exposed lysines with a biotin tag. The parasites were then lysed and labeled proteins were recovered with streptavidin beads. Two parasite samples were analyzed, one containing 2 × 106 sporozoites bearing the VK210 CSP haplotype and one containing 1.8 × 107 sporozoites bearing the VK247 CSP haplotype. The VK247 sample was split in two and half was left unlabeled in order to assess non-specific binding. A total of 90 Plasmodium proteins were identified from the VK210 sample, of which 61 (68%) were identified from two or more peptides, and 221 Plasmodium proteins were identified from the labeled VK247 sample, of which 147 (67%) were identified from two or more peptides. The combined samples identified 239 Plasmodium proteins, of which 72 (30%) were identified in both samples. The 129 proteins identified from two or more peptides and three or more PSM in at least one sample were taken for further analysis. Some proteins could be seen to exhibit direct evidence for incorporation of the biotin label in the identifying mass spectra. Absence of spectral evidence for labeling does not mean that the protein was not labeled [69], but observing labeling in highly abundant sporozoite surface proteins such as CSP and TRAP provides evidence that the labeling and enrichment protocol successfully identified surface-exposed proteins. The non-specific binding was very low—only eight Plasmodium proteins were identified from the unlabeled sporozoites (five identified from two or more peptides) compared to the 221 Plasmodium proteins identified from an equal number of labeled sporozoites from the same sample. The eight Plasmodium proteins in the control were identified by 49 PSMs, more than 30-fold fewer than the 1604 PSMs obtained from the labeled sample (S9 Table). The labeled and unlabeled VK247 sporozoites were split from the same batch of purified sporozoites and, except for the labeling steps, were processed identically in parallel along with the labeled VK210 sample, including lysis, capture on magnetic biotin beads, washes, elution, SDS-PAGE and in-gel digestion, and all three samples were analyzed by LC-MS one after the other on the same column. As such, the raw number of spectral counts gives the best estimate of relative abundance when comparing the relative abundance of a protein identified in both the labeled and unlabeled VK247 samples. Seven of the eight proteins identified from the unlabeled control were also identified in the labeled sample. Although there was insufficient data to assess statistically significant enrichment of labeled versus unlabeled proteins, all seven proteins were at least two-fold more abundant in the labeled sample based on the number of PSM. Proteins identified in the unlabeled control included the known sporozoite surface proteins TRAP and sporozoite surface protein essential for liver stage development (SPELD; PVP01_0938800) [70], as well as actin (PVP01_1463200), which has been detected on the surface of ookinetes [71]. These proteins exhibited direct evidence from the identifying mass spectra that they had been labeled with the biotin tag the labeled samples. They were also among the most abundant proteins in the sporozoite proteome (top 2%), so their presence among non-specifically binding proteins is not surprising. Given the above, the contribution of non-specific binding to the proteins identified in both samples was assumed to be minimal. In order to rule out low-confidence identifications, only proteins identified from two or more peptides and three or more PSM were taken for further analysis.
Although the biotin tag for surface labeling is putatively membrane-impermeable [72], based on previous work, some portion of sporozoites were assumed to have compromised plasma membranes, resulting in labeling of intracellular proteins [20, 73]. Therefore, combined theoretical and experimental evidence were used to identify the strongest candidates for surface-exposed proteins from among all those identified by the surface protein enrichment strategy. Proteins that were identified with high confidence as described above were assigned a priority tier (1 being highest) as follows: Tier 1) possessing predicted transmembrane (TM) domain(s), signal peptide or glycophosphatidylinositol (GPI) anchor and exhibiting spectral evidence of incorporation of the biotin tag; 2) exhibiting spectral evidence of incorporation of the biotin tag but lacking predicted TM domain(s), signal peptide or GPI anchor; 3) possessing predicted TM domain(s), signal peptide or GPI anchor but lacking spectral evidence of incorporation of the biotin tag; 4) lacking predicted TM domain(s), signal peptide or GPI anchor as well as lacking spectral evidence of incorporation of the biotin tag. These criteria produced a list of 36 high-quality candidate surface proteins (Table 4). Of these, 31 orthologs were also detected by similar analyses of putatively surface-exposed proteins on P. falciparum [20] or P. yoelii [74] salivary gland sporozoites. Several of these are known to be secreted and/or surface-exposed on sporozoites, including CSP, TRAP, SPELD [70], GEST, sporozoite surface protein 3 (SSP3; PVP01_1427900), hexose transporter (HT; PVP01_0420400) and CelTOS.
The sole function of a Plasmodium sporozoite injected into the skin of the host during a mosquito bite is to find its way to the liver and initiate liver stage development. To achieve this aim, the parasite must be mobile, traverse various tissue barriers, and finally recognize and infect a hepatocyte in the liver. These complex processes rely on interaction of various parasite proteins with the host tissues and represent a bottleneck of Plasmodium infection, as only small fraction of sporozoites produced in a mosquito makes it to the host liver. Impediment of the parasite-host interaction presents an opportunity to interfere with the parasite life cycle.
Presented here is an effort to identify and characterize the proteins in P. vivax salivary gland sporozoites. While the total number of proteins identified with high confidence is comparable to the most comprehensive analyses of its kind performed on P. falciparum and P. yoelii sporozoites [19], the list of identified proteins presented here is not assumed to be complete. The shotgun proteomics methods for high-throughput proteomic profiling employed here are inherently biased toward highly abundant proteins and are affected by sample complexity and the dynamic range of protein concentrations. These limitations are especially pronounced when analyzing mosquito-stage Plasmodium parasites. Obtaining sufficient sample material for analysis is difficult, as it requires dissecting hundreds of mosquitoes and extracting the sporozoites from the salivary glands. There is unavoidable loss of sporozoites during the purification process, but this step is absolutely critical, otherwise the signal from contaminating mosquito proteins masks parasite peptides in the mass spectrometer. Assuming that there are more proteins present in sporozoites than detected here, the identification of these proteins will likely require further improvements in techniques for purifying large numbers of sporozoites along with continued improvements in mass spectrometer detection limit and duty cycle.
Previous efforts to catalogue the protein complement of Plasmodium sporozoites have used laboratory strains, whereas the sporozoites analyzed in this work were obtained from clinical samples isolated from natural infections. Because of the scarcity of the samples, each of the four samples analyzed here (two whole-proteome and two surface-enriched) were different field isolates of P. vivax. To account for expected polymorphism among field isolates, the mass spectra were searched against a reference proteome supplemented with protein sequences bearing polymorphisms observed in 32 different Thai field isolates. This analysis showed that 301 proteins in the samples exhibited sequence polymorphisms not found in the P. vivax P01 reference proteome. While nearly all of these of proteins could have been identified from conserved peptides present in the reference proteome, knowledge of polymorphisms gained from genomic and transcriptomic analyses of field isolates was critical for accurate analysis of the samples. At a qualitative level, the VK210 and VK247 haplotypes of CSP could be confirmed, and the VK210 whole proteome sample appeared to have contained a mixed infection of at least two different VK210 field isolates. Additionally, it was not possible to detect O-fucosylation of TRAP in some of the samples without the knowledge that the TRAP peptide containing the O-fucosylated Thr can contain an Arg→Lys substitution, a polymorphism that was present in a third of the analyzed field isolate genomes but in neither the P. vivax Sal-1 nor the P. vivax P01 reference proteomes. Protein quantification by spectral counting revealed important information about the relative abundance of proteins within and between salivary gland sporozoite samples of the same species and across species. The high-throughput proteomics methods employed here require exact knowledge of the protein sequence in order to detect component peptides, thus accurate quantification of proteins bearing amino acid substitutions required knowledge of protein sequence polymorphisms that were not reflected in the reference proteome.
Label-free protein quantification based on spectral counting was used to compare relative protein abundance within and among samples. In addition to comparing the two P. vivax salivary gland sporozoite proteomes to each other, a union list of identified P. vivax proteins was compared with a P. falciparum salivary gland sporozoite dataset obtained from re-analysis of published proteomic data [19]. The quantitative data was useful for identifying general trends, e.g., highly abundant proteins that were identified in all datasets or proteins that were abundant in one species but whose orthologs were conspicuously absent in another. Identifying a protein in one sample and not the other when comparing two similar samples is common in MS-based proteomics; a protein may not be detected because it is truly absent in the sample, because it is below the detection limit of the instrument, or due to some technical issue such as interference from some other species in the sample or the stochastic sampling of the ion stream by the mass analyzer [75]. In the comparison of the two P. vivax salivary gland sporozoite samples, the proteins identified in one sample but not the other were primarily low-abundance, suggesting that detection limit was the primary source of differences in proteome coverage between the two samples. Likewise, when the relative abundance of any given protein was compared between the two samples, high-abundance samples showed little deviation from the population mean of a 1:1 ratio, while the deviation tended to increase at lower concentrations. This increasing deviation with decreasing spectral counts is a known phenomenon in spectral counting methods [39, 44]. Statistical tests based on independent assessment of protein ratios as well as the population of protein abundances in the samples were used to identify proteins with significantly different protein abundance between samples. When biological replicates are feasible, it is common to employ a paired t-test of protein abundance ratios obtained from spectral counts [38]. When biological replicates are not available, a conservative likelihood ratio test, the G-test, can be applied to the pooled spectral counts from LC-MS technical replicates [40, 44]. The advantage of this test for spectral counting is that at low spectral counts where quantification error is high, only the largest protein ratios achieve significance. However, at high spectral counts, even protein ratios near 1:1 can be assigned significance. Based on the observation that the population of log-transformed abundance ratios of proteins detected in both samples was Gaussian, it could be assumed that only proteins at the extreme ends of the distribution were truly significant, and the complementary error function provided a p-value as a metric for this deviation. In conjunction with the G-test to eliminate spuriously large protein ratios obtained from low spectral counts, it was possible to identify a small number of proteins that may have had truly different protein expression between the two P. vivax samples (though technical sources of variance cannot be ruled out). Conversely, the analysis showed that the majority of identified proteins exhibited similar protein expression levels in the field isolates examined. Such observations are of value when considering targets for novel vaccines or therapeutics. The same quantitative approach used to compare the two P. vivax samples also enabled comparison of the P. vivax and P. falciparum samples and demonstrated that for most proteins identified in one species, the ortholog was identified at a similar relative abundance in the other species, especially among high-abundance proteins. In addition to proteins identified in each species for which there is no annotated ortholog in the other species, the significance thresholds provided by the statistical tests identified a number of proteins with putatively different expression in the two species that warrant further exploration, including proteins identified in one species whose ortholog was not detected at all in the other. Further investigation will be required to determine if these sorts of proteins are truly expressed at greater levels in P. vivax compared to P. falciparum and whether they may play a specialized role in P. vivax biology.
Examining the curated functional annotation of the Plasmodium proteome revealed that many of the proteins that are known or predicted to be involved in invasion in blood stages or in sporozoites of different Plasmodium species are also expressed in P. vivax sporozoites. Multiple annotation sources were combined in order to compile a list of dense granule, microneme, rhoptry, rhoptry neck, and glideosome proteins, and protein detection was compared between P. vivax and P. falciparum salivary gland sporozoites (Fig 4, S10 Table). There was a very high overlap in the invasion-related proteins detected in the two species, and the handful of proteins detected in only one species or the other were detected only at low abundance. For example, the most abundant of these invasion-related proteins identified in P. vivax but not P. falciparum was the micronemal protein merozoite TRAP-like protein (MTRAP; PVP01_0613800). This protein was confidently identified in both of the P. vivax samples analyzed here, though only in the second quartile of relative abundance. Its syntenic ortholog (PF3D7_1028700) was not among the 2010 proteins identified from P. falciparum sporozoites, but the transcript of MTRAP has previously been detected in P. falciparum salivary gland sporozoites [76], suggesting that the failure to detect this protein by proteomics may reflect limit of detection rather than true biological difference between the two Plasmodium species.
In addition to detecting the proteins present in P. vivax salivary gland sporozoites, the data were analyzed for evidence of post-translational modifications, specifically phosphorylation and glycosylation. Protein phosphorylation is of interest for development of antimalarial drugs [77, 78] because the reversible modification is involved in regulation of essentially every aspect of the complex Plasmodium life cycle, yet the parasite and the mammalian host are sufficiently phylogenetically divergent that many Plasmodium protein kinases can, in theory, be selectively inhibited [78, 79]. Although the limited amount of starting material available for this work precluded phosphopeptide enrichment, 139 proteins were identified with evidence of phosphorylation, and the orthologs of 48 of these were also phosphorylated in P. falciparum salivary gland sporozoites. The most prevalent functional class of proteins exhibiting phosphorylation was proteins involved in transcriptional and translational regulation, including DNA- and RNA-binding proteins and transcription and translation factors, suggesting active phosphorylation-mediated regulation of gene expression in this stage. Many of these proteins were also among the other prevalent class of proteins, those with ATP binding activity. Components of the gliding motility machinery were also well-represented among the observed phosphoproteins, including MyoA, the glideosome-associated proteins GAP40 and GAP45, several inner membrane complex (IMC) proteins, and the calcium-dependent protein kinase CDPK1. A phosphoproteomic analysis of P. falciparum schizonts [58] found evidence that phosphorylation helps to regulate the glideosome machinery, and in vitro work confirmed that GAP45, MyoA, and CDPK1, which are proteins important for motility of the merozoites that emerge from blood-stage schizonts, are substrates of protein kinase A. That these proteins are also phosphorylated in sporozoites suggests that phosphorylation also plays a role in regulating gliding motility in sporozoites. Recently, a glideosome-associated connector (GAC) protein has been identified which links the adhesin MIC2 (the Toxoplasma gondii ortholog of Plasmodium TRAP) to F-actin and is essential for motility and invasion [80]. P. vivax GAC (PVP01_1110200) was in the top decile of protein abundance in the samples analyzed here. Over half of the GAC in P. vivax sporozoites was phosphorylated at Ser10 or Ser15 on the protein N-terminus, and approximately 30% of the protein was phosphorylated at Ser1495 or Ser1496 on a section bearing homology to a protein-binding armadillo-type fold (ARM). Ser1496 on the same peptide in P. falciparum was approximately 38% phosphorylated in the salivary gland sporozoite data analyzed here. Previous phosphoproteomic analyses identified eight different phosphosites on GAC in P. falciparum blood stages [56–60], several of them on regions of conserved sequence between P. falciparum and P. vivax. The ARM phosphosite seen in P. vivax and P. falciparum sporozoites was also observed in P. falciparum blood stages, but the N-terminal phosphosites were not. AMA1 plays a role in adhesion of merozoites to erythrocytes, though its role is dispensable for rodent Plasmodium sporozoite invasion of hepatocytes [81]. The role of AMA1 in human-infecting Plasmodium sporozoites is currently undetermined. The protein has a single transmembrane domain near its C-terminus that serves as an anchor to the parasite plasma membrane. Previous proteomic analyses have shown that several residues on the cytoplasmic tail of AMA1 are phosphorylated in blood stage parasites [56–60], and mutating these residues to prevent phosphorylation resulted in a defect in invasion of erythrocytes [82]. AMA1 in P. falciparum shares an identical sequence with P. vivax at the C-terminus. In the P. vivax sporozoite samples analyzed here, the cytoplasmic tail of AMA1 was observed with a single phosphorylation at Ser551, Thr553 or Thr554 (corresponding to Ser610, Thr612 and Thr613 in P. falciparum). In the P. falciparum salivary gland sporozoite data re-analyzed here, nearby Ser588 on the cytoplasmic tail was phosphorylated. In the P. falciparum and both P. vivax sporozoite samples, the peptides containing the respective phosphosites were never observed in unmodified form, suggesting AMA1 is constitutively phosphorylated in salivary gland sporozoites. These conserved residues have been observed to be variably phosphorylated in AMA1 in P. falciparum blood stages [56, 59, 60]. These results suggest that phosphorylation of AMA1 plays a role in regulating the protein’s function in sporozoites, perhaps by mediating attachment to the glideosome.
Importantly, many of the phosphoproteins identified in the sporozoite samples have not been observed to be phosphorylated in the handful of blood-stage phosphoproteomes published to-date, including several proteins known to be specific to the sporozoite stage. Among these stage-specific phosphoproteins are proteins known to be located on the sporozoite surface. For example, it has been previously determined that sporozoite surface protein 3 (SSP3; PVP01_1427900, PF3D7_0812300) is found on the surface of P. yoelii and P. falciparum salivary gland sporozoites [19, 20, 83], and here we show that it is likely surface-exposed in P. vivax sporozoites as well. The role of SSP3 is not fully understood, but initial work suggests that it plays a role in gliding motility [83]. Approximately 30% of SSP3 in P. vivax sporozoites was phosphorylated at Ser440 near the C-terminus of the protein. A single predicted transmembrane domain at residues 402–424 is likely the point where the protein is anchored to the membrane, placing the phosphosite on the cytosolic portion of the protein. Approximately 40% of SSP3 in P. falciparum sporozoites was similarly phosphorylated at the C-terminal cytoplasmic tail at either or both of two Ser residues, Ser456 (the P. falciparum counterpart of the P. vivax S440 phosphosite) or nearby Ser459. Further experimentation will be required to elucidate any role phosphorylation might play in the function of this and other phosphorylated sporozoite surface proteins, as well any effect on their antigenicity.
It has been recently shown that the major sporozoite surface proteins CSP and TRAP are glycosylated at their TSR domains in P. falciparum salivary gland sporozoites [20] and the data presented here now show that these proteins are also glycosylated in P. vivax sporozoites. Strikingly, both CSP and TRAP in P. vivax were modified only with a single deoxyhexose (presumably O-fucose), whereas in P. falciparum, CSP was observed with either a deoxyhexose or a deoxyhexose-hexose disaccharide (likely O-fucose-β-1,3-glucose), and TRAP was observed with the O-linked mono- or disaccharide as well as with a C-linked hexose (likely C-mannose). The reason for this difference is not immediately clear. The monosaccharide-modified versions of TRAP and CSP were the dominant forms in both P. falciparum and P. vivax sporozoites, and a putative O-fucosyltransferase POFUT2 (PF3D7_0909200, PVP01_0707700), which could hypothetically add O-fucose to TSR domains, was observed to be expressed in sporozoites of both species. In P. falciparum sporozoites, the dissacharide-modified versions of CSP and TRAP were present at lower abundance than the monosaccharide-modified versions, so it is conceivable that in the P. vivax samples the disaccharide-modified versions were present but at concentrations below the detection limit. It is also possible that the necessary glycosyltransferase was not expressed. While no putative β-1,3-glucosyltransferase for adding glucose to O-fucose has been identified in Plasmodium, PfPIESP1 (PF3D7_0310400) has been identified as having sequence homology with human β-1,3-glucosyltransferase and possesses putative glycosyltransferase domains [20]. PIESP1 was expressed in P. falciparum salivary gland sporozoites [19], but its P. vivax homolog (PVP01_0829800) was barely detected in the P. vivax sporozoites analyzed here (identified by a only two PSMs in one sample and not at all in the other). The absence of C-linked hexose on TRAP in the P. vivax sporozoites analyzed here was unequivocal. Unlike O-linked glycans, C-mannose is not gas-phase labile and withstands collision-induced dissociation, giving rise to peptide fragment ions that precisely identify the residue to which the modification is attached. Furthermore, some portion of the C-mannose undergoes cross-ring fragmentation that gives rise to neutral loss species that further corroborate the identity of the C-mannose [20]. The glycosylated TRAP peptide was identified by dozens of spectra in the P. vivax samples, none of which contained evidence for modified Trp. A putative C-mannosyltransferase (PF3D7_0806200, PVP01_0114300) that could hypothetically add C-mannose to TSR domains was expressed in both P. falciparum [19] and P. vivax sporozoites, though this function has yet to be verified experimentally. Interestingly, this disparity in glycosylation was also observed in TRAP expressed in mammalian cells: PfTRAP was C-mannosylated but PvTRAP was not [55]. It is notable that in PvTRAP, the sequence where it would be expected to find C-mannosylation contains prolines that could affect the secondary structure of the motif and disrupt recognition by the glycosyltransferase. Further studies will be required to determine whether the observed differences in CSP and TRAP glycosylation between P. falciparum and P. vivax are due to differences in enzyme function or some other technical or biological reason. Importantly for design of vaccine antigens, O-fucosylation of CSP and TRAP almost certainly affect antigenicity of the proteins. Structural studies of these proteins have shown that the glycans project above the protein surface and yet have structurally constrained orientations [54, 55, 84], and studies of the conserved TSR domain in other species have shown that fucosylated amino acids may be viewed as surrogate amino acids [85], so the protein and carbohydrate elements create unique combinatorial epitopes.
A chemical labeling approach was used to enrich proteins that are surface-exposed on salivary gland sporozoites. As previously discussed [20, 73], this surface biotinylation approach is known to produce spurious results due to labeling of cytosolic proteins presumably arising from a portion of sporozoites that exhibit compromised plasma membranes, an inevitable byproduct of the excessive sample handling involved in dissecting, purifying, and labeling the parasites. The data presented here were curated with theoretical and experimental information to select those identified proteins that are the most likely to be truly surface-exposed and proteins were assigned priority tiers based on this evidence in order to identify high-quality candidates for future efforts to validate and test these proteins as vaccine antigens. This approach is supported by the fact that the list of high-quality candidates includes several known sporozoite surface proteins, including CSP, TRAP, SSP3, and SPELD. Additionally, cross-referencing the results with similar analyses of P. falciparum [20] and P. yoelii [74] salivary gland sporozoites revealed a large overlap in the proteins that the technique identified across species. A notable exception was a 6-Cys protein (PVP01_0303900) that has no ortholog in P. falciparum or P. yoelii but does have syntenic orthologs in the more closely related malaria parasites P. knowlesi and P. cynomolgi. In other Plasmodium species, other 6-cys proteins are known to be found on the sporozoite surface and to play a role in liver invasion [20, 86]. Future studies will determine what role this putative surface protein may have in sporozoites and if it will be useful as a vivax-specific antigen.
The putatively surface-exposed proteins identified here as well as in P. falciparum and P. yoelii sporozoites included known cytosolic proteins. While there remains the possibility that these results represent experimental artifact as discussed, there is increasing evidence that cytosolic proteins can have “moonlighting” roles and be found on the cell surface of Plasmodium and other organisms. For example, the chaperone HSP70-2 (BiP) has been shown to localize at the surface of certain cell types in other organisms [87], and HSP70-2/BiP (PVP01_0716300, PF3D7_0917900) was identified as putatively surface-exposed in both P. vivax and P. falciparum salivary gland sporozoites. Another protein classified as a heat shock protein, HSP20, has been identified by biotinylation of sporozoite surface proteins in both P. falciparum and P. vivax sporozoites. This protein has been demonstrated to be surface-exposed in P. berghei salivary gland sporozoites by immunoelectron microscopy [88]. Other intracellular proteins repeatedly identified as surface-exposed by the surface labeling technique include components of the gliding motility machinery, including actin, MyoA, glideosome-associated proteins (GAP), and inner membrane complex (IMC) proteins. Immunofluorescence assays of un-permeabilized P. falciparum salivary gland sporozoites showed that the glideosome proteins GAP45 and MTIP were accessible to antibodies during gliding [20], though whether this was due to relocation of the proteins to the sporozoite surface or permeability of the plasma membrane in gliding sporozoites is not known. Similarly, immunofluorescence identified actin, which is part of the gliding machinery, at the ookinete surface [71]. Another glideosome-associated protein, GAP50, has been shown to relocate to the surface of gametes where it binds complement regulator proteins and inactivates human complement in the blood meal that would otherwise induce lysis of the parasite [89]. Taken together, this information suggests that even “known” intracellular proteins identified by this surface labeling method can reflect truly surface-exposed proteins and warrant further investigation.
In conclusion, the MS-based proteomics methods employed here enabled the most comprehensive identification to-date of proteins and their post-translational modifications present in P. vivax sporozoites. Combined with the identification of putatively surface-exposed proteins of P. vivax salivary gland sporozoites, these results suggest that the complement of surface-exposed proteins on salivary gland sporozoites may contain many unexpected as well as post-translationally modified proteins that warrant further experimentation to verify their localization and assess their suitability as vaccine antigens.
|
10.1371/journal.pbio.0050106 | Peptides Encoded by Short ORFs Control Development and Define a New Eukaryotic Gene Family | Despite recent advances in developmental biology, and the sequencing and annotation of genomes, key questions regarding the organisation of cells into embryos remain. One possibility is that uncharacterised genes having nonstandard coding arrangements and functions could provide some of the answers. Here we present the characterisation of tarsal-less (tal), a new type of noncanonical gene that had been previously classified as a putative noncoding RNA. We show that tal controls gene expression and tissue folding in Drosophila, thus acting as a link between patterning and morphogenesis. tal function is mediated by several 33-nucleotide–long open reading frames (ORFs), which are translated into 11-amino-acid–long peptides. These are the shortest functional ORFs described to date, and therefore tal defines two novel paradigms in eukaryotic coding genes: the existence of short, unprocessed peptides with key biological functions, and their arrangement in polycistronic messengers. Our discovery of tal-related short ORFs in other species defines an ancient and noncanonical gene family in metazoans that represents a new class of eukaryotic genes. Our results open a new avenue for the annotation and functional analysis of genes and sequenced genomes, in which thousands of short ORFs are still uncharacterised.
| How cells organize into embryos remains a fundamental question in developmental biology. It is likely that significant insights into embryo development will emerge from the characterisation of novel types of genes. Yet most current genome annotation methods rely heavily on comparisons with already-known gene sequences, so genes with previously uncharacterised structures and functions can be missed. Here we present the characterisation of one of these novel genes, tarsal-less. tarsal-less has two unusual features: it contains more than one coding unit, a structure more similar to some bacterial genes; and it codes for small peptides rather than proteins. In fact, these peptides represent the smallest gene products known to date. Functional analysis of this gene in the fruit fly Drosophila shows that it has important functions throughout development, including tissue morphogenesis and pattern formation. We identify genes similar to tarsal-less in other species, and thus define a tarsal-less–related gene family. We expect that a combination of bioinformatic and functional methods, such as those presented in this study, will identify and characterise more genes of this type. These results suggest that hundreds of novel genes may await discovery.
| The work of the last decades has seen a breakthrough in our understanding of the genetic and molecular mechanisms of development. Classical genetic approaches have been complemented by systematic searches for new genes and their functions, resulting in an exponential increase of information. This new knowledge has filtered to related areas such as cell biology, medical research, and increasingly, evolution and population genetics. However, there still remain significant gaps in our understanding, not only of how different aspects of development such as patterning, morphogenesis, and differentiation are organised and implemented at the cellular level, but also in how these different aspects are coordinated. One exciting possibility is that new types of genes with new coding arrangements await discovery and characterisation. The number of known key regulatory genes and signalling proteins remains small, in the region of the hundreds, but sequenced and annotated genomes, including the human genome, still contain thousands of genes and transcripts without known function or sequence similarity to other genes [1–3] or are deemed RNA or noncoding genes [4].
The development of the Drosophila leg offers a good system in which to pursue this analysis further. Fly legs have a high density of pattern elements and a simple developmental topology, with a single main axis of patterning and growth, the PD axis [5,6]. The legs of Drosophila develop from presumptive organs called imaginal discs, and the morphogenesis of these discs, in particular their acquisition of a stereotyped set of folds that prefigure the morphology of the final appendage, is coordinated with patterning and growth [7,8]. An understanding of the main patterning events in leg development has recently been achieved [9,10], and a preliminary understanding of the coordination of a cell-signalling–mediated patterning event with its morphogenesis, in the development of joints, via Notch signalling, has been obtained [11–15]. More genes with well-defined morphogenetic functions await integration into this scheme [16], but the identification of further links between patterning and morphogenesis remains elusive. Our search for these links led us to the isolation and characterisation of a new Drosophila gene that we call tarsal-less (tal). This gene expresses a 1.5-kilobase (Kb) transcript that had been classified as putatively noncoding [17,18]. It contains several open reading frames (ORFs) smaller than 50 amino acids (aa) and thus is putatively polycistronic. Our analysis shows that surprisingly, the peptides translated from ORFs of just 11 aa mediate the function of the gene. Therefore tal has two novel features for eukaryotic coding genes: the direct translation of short, unprocessed peptides with full biological function, and their tandem arrangement in a polycistronic messenger. We identify tal homologous genes in other species and observe that they define a new, noncanonical gene family of ancient origin. We expect that a combination of new bioinformatics and proteomics methods tailored to the search of peptides and small ORFs (smORFs) [19,20], plus a reassessment of classical data, will identify and characterise more new coding genes with similarly important functions in these and other areas of biology.
We identified the tal gene through a spontaneous mutant (tal1) with defective legs in which the tarsal segments [21] do not develop (Figure 1). Meiotic and deficiency mapping, followed by cytogenetic and molecular methods, revealed tal1 to be a small inversion between regions 86E1,2 and 87F15. The tal1 phenotype maps to the 87F15 breakpoint, to the left of the Mst87F gene (Figure 1A). There is no gene prediction in this region, but there is a noncoding cDNA, LD11162 [22], and two lethal P element inserts, S011041 and KG1680, located 5′ and 3′ respectively to LD11162 (Figure 1A). We found KG1680 to be allelic to tal1 and to produce similar phenotypes in legs over a chromosomal deficiency for the tal region. These are regulatory mutants that affect only the imaginal disc function. Mobilisation of both KG1680 and S011041 insertions produced a number of alleles that all define a single complementation group. Alleles producing a deletion of the coding region for LD11162 (talS68, talS18, and talK40; see Figure 1A) behave as nulls.
In addition to LD11162, there are several cDNAs isolated independently [22]. We sequenced one of these, LP10384, that is identical to LD11162. In addition, a single transcript of 1.5 kb corresponding to this cDNA has been identified by Northern blots [17] and reverse-transcriptase PCR (unpublished data). The expression of this transcript is similar to the lacZ reporter S011041 (Figure 2A and 2B), is coincident with the regions affected in tal mutants (Figure 1B and 1C), and is lost in ta1 mutants (unpublished data). To prove definitely that this transcript encodes the function of the tal gene, we performed a rescue experiment. The KG1680 insert was replaced by a Gal4 insert [23]. The resulting Gal4 line (P{GaWB}talKG, subsequently referred to as tal-Gal4) is a regulatory viable allele similar to tal1 and the KG1680 insertion, and produces a tal phenotype in legs (Figure 1B–1D) while simultaneously driving the expression of upstream activating sequence (UAS) constructs [24] in the tal pattern. We generated a construct with the full-length LP10384 cDNA downstream of a UAS promoter (UAS-tal) and tried to rescue mutant animals of the genotype tal-Gal4/talS68 by introducing this UAS-tal construct. In these tal-Gal4/talS68; UAS-tal/+ animals, the phenotypes were rescued to wild type (Figure 1E). This rescue proves that the tal function is encoded by LP10384, which represents the tal RNA. Moreover, ectopic expression of UAS-tal produces mutant phenotypes that are consistent with tal being a tarsal determinant: transformation of distal tibia and fusion to tarsi, where tal is normally expressed (Figure 1F).
tal expression in the leg has the interesting feature of being transient (Figure 2A–2C). The time of tal expression (from about 80 to 96 h after egg laying [AEL]) coincides with the specification of the tarsal region by the activation of specific genes in ring patterns similar to that of tal [9,10]. One of the genes activated transiently at this time and required for tarsal patterning is the zinc-finger transcription factor rotund (rn) [25]. We observe that the expression of rn is lost in tal mutants and is extended following ectopic expression of UAS-tal (Figure 2D–2F). In contrast, loss or excess of function of rn (induced with a UAS-rn construct) has no effect on tal expression (unpublished data). These results show that the rn gene is a downstream target of tal.
Further functions of tal are apparent. In tal mutants, the whole tarsal region is missing, a stronger phenotype than that produced by rn mutants [25], and anti-Caspase 3 staining reveals that this is not produced by cell death (unpublished data). tal expression precedes and then straddles the tarsal furrow within which the tarsal segments develop (Figures 2A, 2B, and 3) [26]. In tal mutant discs, the tarsal fold does not form further than a superficial constriction, subsequent tarsal folds do not form, and the tarsal region does not grow (Figure 3). Reciprocally, ectopic expression of tal induces the appearance of ectopic folds in legs (unpublished data). These morphogenetic phenotypes are not produced by changes of rn expression on its own [25], and the lack of folding is not rescued by inducing expression of rn in tal mutants.
tal null alleles are embryonic lethal. tal expression in the developing embryo is initially segmental (Figure 4A; see also http://www.fruitfly.org), followed by a later and more complex pattern of expression in many organs (Figure 4B–4D). The embryonic mutant phenotypes include broken trachea, loss of cephalopharyngeal skeleton, abnormal posterior spiracles, and lack of denticle belts (Figure 4E–4H). These are the regions where tal is expressed from stage 13 until the end of development (Figure 4C and 4D). This phenotype is identical to a deletion of the entire 87F13–15 region, and is not enhanced by removing any putative maternal contribution in germ-line clones (unpublished data). Ectopic expression of UAS-tal produces reciprocal mutant phenotypes, such as extra sclerotised elements in the cephalopharyngeal skeleton (Figure 4I).
Despite the early segmental pattern of expression, tal mutants do not show any segmentation or homeotic phenotype (Figures 4 and S2). Therefore, the early segmental expression seems to be only a transient state to establish expression in the precursors of the tracheal system (Figure 4B). Although the mutant epidermis lacks denticle belts, segment-specific epidermal sensory organs are present, and segments are formed. Expression of markers such as wingless (Figure 4J), Distal-less, and Ubx (Figure S2) is normal. The late expression of wingless is not expanded and thus is not responsible for the observed loss of denticles [27]. Furthermore, tal function is independent of shaven-baby (Figure 4K) [28]. Altogether these results suggest that tal acts in parallel to the canonical denticle-patterning cascade [29]. Interestingly, tal mutant cells do not undergo the tubulin accumulation and cell morphology changes leading to the differentiation of denticles [30] (Figure 4L and 4M, and unpublished data).
Our results show that tal is required for several key developmental processes. The tal cDNA has been classified as “putatively noncoding” [17,18] on the basis of having no ORF longer than 100 aa and no known homologies. A number of candidate smORFs are present in the tal transcript. We will refer to these smORFs according to their sequence and position from 5′ to 3′ as 1A, 2A, 3A, AA, and B (Figure 5A). The type-A ORFs (1A, 2A, 3A, and AA) include a conserved LDPTGXY motif of 7 aa, and this motif is very strongly conserved in the cDNA of homologous genes that we have identified in other arthropods (Figure 5 and Figure S1). ORF 1A and 2A encode an identical 11-aa peptide. ORF 3A encodes another 11-aa peptide very similar to 1A. ORF AA encodes a 32-aa peptide whose N- and C-termini each contain a LDPTGXY motif (Figure 5A). ORF-B encodes a 49-aa peptide without known domains other than a poly-Arg stretch and is somehow weakly conserved in other insects (Figures 5 and S1).
The conservation of the aa sequences in other species suggests, but does not prove, the translation of these smORFs. With such short sequences, aa conservation cannot be distinguished easily from simple nucleotide conservation, and therefore we decided to study the functional significance of these smORFs and to obtain experimental evidence for their translation. For this, we have built upon our rescue and ectopic expression experiments that proved that tal is encoded by the mRNA represented by LD11162 and LP10384 (Figure 1B–1F). We have tried to rescue tal mutants with UAS constructs containing different directed mutations affecting specific ORFs, and in separate experiments, we have studied the ectopic effects of such constructs and compared them with those of full-length UAS-tal. The results are summarised in Figure 6A.
A construct containing a full-length cDNA from Bombyx mori (Bm-wds) produces the same effects as a full-length Drosophila one. This result validates the comparative results described above and also indicates that tal functionality lies in the ORFs, since these are the only stretches of DNA sequence conserved between Drosophila and Bombyx (Figure S1). Therefore, we next concentrated on dissecting the role of the ORF sequences in the Drosophila cDNA. A deletion construct (AB) leaving only a type-A ORF plus ORF-B is still fully functional. It can rescue tal mutants, and it produces the same ectopic effects as full-length tal. Construct delA deletes the type-A ORF and is just 32 base pairs (bp) shorter than AB, but has lost all functionality, suggesting that the type-A ORF is key for the tal function, and ORF-B is dispensable. It could be argued that the translation initiation context of ORF-B is too weak and that its expression requires an upstream functional type-A ORF. However, the construct ATG-B, in which we have put ORF-B under the control of the Tal1A initiation context, is still unable to reproduce the tal rescue or ectopic effects. Reciprocally, two constructs in which potential translation of ORF-B has been abolished, by either deleting it (delB) or by mutating its start codon (NoB), are fully functional, rescue tal mutants, and produce the same ectopic effects as full-length UAS-tal, including activation of rn expression (unpublished data). Finally, a construct containing only one type-A ORF (1A) is fully functional, and a one-nucleotide insertion that produces a frameshift (1A-FS) abolishes its functions.
Altogether, these results show that (1) an 11-aa type-A ORF provides tal function, and (2) ORF-B has no developmental function.
These functional results indicate that tal function resides in the type-A ORFs, and the results with constructs Bm-wds, 1A, and 1A-FS seem to exclude a model of tal function as a noncoding RNA. Thus we sought direct proof of tal translation.
The small size of the putative tal peptides makes them difficult to detect directly. In order to facilitate their detection in in vitro and in vivo experiments, we have tagged them by introducing the green fluorescent protein (GFP) coding sequence, minus the start and stop codons, in frame and within each of the type-A ORFs and the ORF-B (Figure 6B). Thus, the resulting fusion constructs still have the tal sequences relevant for translation, including the 5′ and 3′ UTRs, the initiation consensi, and start codons. Construct 1A-GFP contains the GFP sequence within the type-A ORF of the AB construct, which was functional and contains the 1A translation initiation environment. 2A-GFP, 3A-GFP, AA-GFP, and B-GFP contain each GFP fusion within a full-length tal cDNA. Expression of these constructs in a reticulocyte in vitro transcription and translation system with [35S]-methionine shows that the fusion peptides are expressed from the 1A-GFP, 2A-GFP, and AA-GFP constructs, but not from the B-GFP (Figure 6C). Transfection of these constructs into Drosophila S2R+ cells confirmed these results and also showed translation of 3A-GFP (Figure 6D). In all cases, we can discard the interpretation that the results are due to translation from a second methionine in the GFP sequence, not only because of the size of the fusion products obtained, but also because these putative peptides would lack the N-terminal sequences that are essential for GFP fluorescence [31].
Thus, our results show that the tal gene is coding, and polycistronic, because several peptides can be synthesised from a single RNA species. The type-A peptides provide the full tal function, and are translated both in vitro and in vivo.
Our results show that translation of an RNA containing smORFs of just 11 aa is required for several important processes during development. Although the tal cDNA contains several copies of the type-A ORFs related by a common LDPTGXY domain, a construct containing just one of them is fully functional. Small peptides are known to have important biological functions, most clearly in endocrine and neural communication [32], but in all described cases, these peptides are mature, cleaved products of a longer ORF. The originality of the tal gene is thus 2-fold. First, smORFs of just 33 nucleotides are fully functional and capable of translation. Second, the carefully regulated local expression of these peptides in complex patterns (as opposed to a systemic release) has important developmental functions. Our genetic and molecular analysis (Figure 1A and unpublished data) show that the tal genomic region contains specific regulatory sequences spread out over a minimum of 25 Kb.
We notice that tal expression and function are often associated with tissues undergoing changes of shape such as folding and invagination. The development of the fly leg is directed by a regulatory cascade involving cell signals and region-specific transcription factors [9,10,33] (reviewed in [6]). Regulatory interactions between these identity-conferring transcription factors refine and stabilise the final pattern [34,35]. This pattern is then translated into morphogenetic movements and position-specific cell differentiation programs [16,36]. tal seems to be an important part of the leg developmental process and to act as a link between patterning and morphogenesis. On the one hand, the transient ring of tal expression appears in the precise time and place to control tarsal patterning, by promoting rn expression and by being involved in further regulatory interactions with other leg-patterning genes (Figure 2 and unpublished data). On the other hand, tal controls folding of the leg tissue independently of these effects. In the wild-type leg imaginal discs, a complex morphogenetic process involving the appearance of extra folds within the tarsal furrow, in correlation with leg growth, is apparent [26]. In tal mutants, this morphogenetic process is compromised, whereas in excess-of-function experiments, ectopic expression of tal induces the appearance of ectopic folds in legs. In the mutant discs, cells undergo an apico-basal constriction, but the tarsal furrow never widens into a fold; the appearance of further tarsal sub-folds is precluded, and the presumptive tarsal region does not grow. In the embryo, tal expression is found in tissues of ectodermal origin that undergo an invagination without compromising their epithelial organisation, such as the foregut (and later on in its derivatives, the proventriculus and the pharynx), the hindgut, the developing trachea, and the spiracles [37]. In mutant embryos, head involution is slow, the pharynx is short and misplaced, and tracheal fusion is incomplete (Figure 4 and unpublished data). The loss of denticles in the epidermis does not seem based on alterations of the segmental patterning cascade, but on cell morphology defects that do not involve defects in apico-basal cell polarity or epidermal integrity (Figure 4 and unpublished data). Altogether, these results suggest that tal is required for the control of cell movements during tissue morphogenesis. Further research beyond the scope of this initial study should identify the cellular and molecular targets of this function.
Our results provide experimental evidence for function and translation of the type-A ORFs. These include the in vitro and in vivo translation assays, functional rescues, and sequence analysis. Our results therefore imply that tal is polycistronic, because several ORFs can be translated from a single RNA molecule. The question arises of how this can be accomplished in an eukaryotic gene, but the literature provides a possible mechanism. Polycistronic genes are known in eukaryotes including Drosophila [38–40], and so in principle, all tal ORFs could be potentially translated simultaneously. Experimental evidence supports three models for translation of polycistronic messengers in eukaryotes, namely “internal ribosomal entry sites (IRES),” “leaky scanning,” and “reinitiation” [41]. There are clear rules backed by experimental data concerning the DNA sequences and transcript structure involved in each of these models. The tal RNA sequence seems to exclude both the IRES and the leaky scanning possibilities. There is not enough space for IRES between the tal ORFs, and the initiation consensi are stronger in the 5′ ORFs than in the 3′ ones, the opposite of conditions favourable for leaky scanning. However, polycistronic translation of type-A ORFs in the tal transcript is possible under the reinitiation model because their spacing is between 40 and 200 bp, and the short type-A ORFs (1A to 3A) are much shorter than 35 aa. In all cases studied, the presence of 5′ ORFs has a dramatic impact on the rate of translation of the 3′ ones, leading in certain conditions, to total blockage of 3′ translation [41]. Accordingly, our in vitro translation experiment shows a diminishing amount of protein arising from each ORF, with highest levels produced by 1A, and lowest by AA (Figure 5C). We would expect, by virtue of its conserved common domain, that these translated type-A peptides will share the same functions. The presence of repeated or similar ORFs is perhaps a device to ensure enough translation of LDPTGXY-containing peptides. This hypothesis coincides with the results of our structure/function analysis, which shows that a single artificial type-A ORF suffices to provide tal function.
These conclusions are further corroborated by our discovery of tal homologous genes in other insects. These genes contain repeated copies of type-A ORFs in varying number from two (crustaceans and primitive insects) to 11 (Bombyx mori), and an evolutionary trend towards accumulation of more type-A ORFs, including duplications of the entire gene, is apparent. The aa sequence of these type-A ORFs is very strongly conserved in their core domain LDPTGXY. The spacing between ORFs is most compatible with the reinitiation model described above. Not only sequence, but also functionality is conserved, as indicated by the rescue of Drosophila mutants with a Bombyx cDNA. The resilience and long age of the evolutionary history of this gene family suggest, not a recently evolved curiosity of some insects, but a peptide with ancestral and current importance.
All available data suggest that the weakly conserved ORF-B is spurious or nonfunctional. In Drosophila, our functional analysis fails to identify any essential function for ORF-B, and both our in vitro and in vivo studies fail to detect its translation. This is in agreement with the fact that the 5′ presence of several type-A ORFs with strong initiation contexts, allied to the weakness of the context for ORF-B, does not favour the translation of ORF-B (Figure 5A). Furthermore, the size of the ORF AA is 32 aa, near the limit of 35 aa required for continued downstream reinitiation at ORF-B. In agreement with this sequence analysis, ectopic expression of the Bombyx Bm-wds construct containing an ORF-B in Drosophila does not produce any additional phenotypes when compared to those produced by the Drosophila constructs, indicating that the Bombyx ORF-B is not functional either. We would surmise that the weak conservation of ORF-B sequences is either related to some functional requirement (other than translation) for the nucleotide sequence in the region of the transcript, or pure chance.
The conservation of aa sequences has been suggested as evidence for the translation of three type-A ORFs and one ORF-B in a homologous gene called milles-pattes (mlpt) found in the flour beetle Tribolium castaneum [42]. These ORFs are of a similar small size as in Drosophila, but again such aa conservation is not conclusive evidence. In the absence of a biochemical and functional analysis of these different ORFs like the one we present here, it is difficult to guess which ORFs are translated and mediate the function of mlpt. The ORF-B of mlpt has been deemed the main functional element of the gene due to its longer length [42], but in fact, the available data belie this interpretation and favour our own conclusion of ORF-B as nonfunctional. The ORF-B of mlpt has no Kozak consensus at all, and its start codon overlaps with the stop codon of the previous 5′ type-A ORF, a situation that seems most unlikely to lead to ORF-B translation, even by a mechanism of readthrough as postulated [42]. Readthrough and ribosome codon slippage always proceed by skipping bases forward, rather than backwards as would be needed here. Further, ORF-B aa conservation is rather weak. Although Savard et al. [42] identify a “poly-Arg” conserved domain in alignments of selected sequences from species of only three insect orders, this conservation disappears when the comparisons are extended to further orders such as in our sequence analysis (Figures 4 and S1). We note that (1) “orphan” AUG codons are not a rare occurrence (about 500,000 in Drosophila; M. Ladoukakis, personal communication), and (2) that the nucleotide sequence in the ORF-B region is thymidine-poor, which produces a bias in its conceptual translation towards certain amino acids, including Arg. In addition, our analysis shows that tal genes without ORF-B exist, and in fact, an ORF-B is only present in some genes from holometabolous insects.
RNA interference (RNAi) analysis of the function of the whole mlpt transcript identifies several functions [42] that seem homologous to the one we have identified in Drosophila, in particular the tarsal-promoting function, and a requirement in the tracheal system. However, Savard et al. [42] also identify a “gap” and homeotic segmentation phenotypes that our expression and functional data results show to be absent in Drosophila (Figures 3 and S2). This functional difference might be due to the different modes of early embryonic development in Drosophila and Tribolium, which also involve a different complement of gap and maternal genes [43]. To clarify whether this segmentation function is ancestral, but has been lost in Drosophila, or whether it is a recently arisen specialization of Tribolium, will require the functional characterisation of tal in other insects.
All sequenced and annotated genomes contain genes and transcripts without known function, sequence homologies, or even known protein domains. In particular, an increasing number of RNA transcripts are being classified as “noncoding” on the basis of not having ORFs longer than 50–100 aa. Furthermore, genomes contain hundreds of thousands of similarly smORFs that are systematically eliminated from gene annotations for statistical reasons. cDNA libraries and expressed sequence tag (EST) collections also discriminate against small cDNAs, perhaps losing many potential transcripts as well [44]. In the rare cases in which smORFs have been identified in longer, polycistronic messengers, studies have centred on the regulatory effect of the 5′ smORFs and resulting peptides on a standard, longer 3′ ORF. Thus, the possibility of smORFs producing peptides with important, independent functions has been largely overlooked outside of yeast, in which there is firm evidence for their existence [19]. Here we identify tal as a functional gene encoding only smORFs, which are translated. The tal type-A peptides define an ancient gene family with at least a crustacean representative (in Daphnia), and thus is not restricted to insects and is older than 440 million years (the estimated time for the origin of insects). We suspect that this new gene family may in fact be a representative of a new and widespread class of genes and that more genes encoding smORFs, either alone or in polycistronic messengers, await isolation and characterisation. Our analysis shows that a good cross-species sample of sequences is required to predict noncanonical peptide-coding genes, but also that these predictions must be validated by functional data, because in its absence, wrong predictions can be made. We expect that a combination of bioinformatic and functional methods tailored to the search of peptides and smORFs will identify and characterize more new gene products and eukaryotic coding genes. Preliminary results in Drosophila (unpublished data), yeast [19], and Hydra [45] suggest that hundreds of such genes may exist.
A synthetic deficiency for the 87F13–15 region was generated in heterozygous Df(3)urd /Df(3)red31 flies. dpp-Gal4 and Dll-Gal4 were used to drive ectopic transgene expression in flies and embryos, respectively. These stocks plus l(3)S011041 ([46]) and KG1680 ([47]) are available from stock centres (http://flybase.bio.indiana.edu). The svb107 enhancer trap line, which reproduces the shaven-baby pattern of expression [28], and the mutant allele svb2 were a gift from F. Payre. Flies and embryos were mounted in Hoyer's for microscopy.
Replacement of the P{SuPor}KG1680 insertion by a P{GaWB} transposable element was done by mobilisation in omb-Gal4; +/CyO Δ2–3; KG1680/TM3Sb flies [23]. The progeny from possible replacements were screened following UAS-GFP expression. All replacements were precise. Mobilisation of P{lacW}l(3)S011041 and P{GaWB}talKG was carried out with the Δ2–3 transgene. Revertants lacking white and yellow markers as appropriate were isolated. Molecular characterisation of these revertants and replacements was done by PCR, Southern blot, and sequencing as needed. talS68 and talS18 are deletions obtained by mobilisation of P{lacW}l(3)S011041, and talK40 from mobilisation of P{GaWB}talKG.
Developing trachea were revealed with the rhodamine-conjugated Chitin-Binding Protein (CBP at 1:500; New England Biolabs, Beverly, Massachusetts, United States). Other antibodies used were anti-β-galactosidase (1:1,000; Sigma, St. Louis, Missouri, United States; and 1:5,000; Cappel, MP Biomedicals, Solon, Ohio, United States); anti-cleaved-Caspase-3 (Asp 175: Cell Signaling Tech. at 1:250), anti-αtubulin (DM1A at 1:500; Sigma), anti-Wingless (1:50; Developmental Studies Hybridoma Bank [DSHB], Iowa City, Iowa, United States), anti-Ubx FP388 (1:20; R. White), and anti-Dll (1:2,000; I. Duncan). In developing leg discs, the actin cytoskeleton was revealed by phalloidin-rhodamine (1:40; Molecular Probes, Eugene, Oregon, United States) and basal membranes by anti-β-integrin (1:500; DSHB). Secondary antibodies conjugated to biotin, rhodamine, and FITC were used (Jackson ImmunoResearch, West Grove, Pennsylvania, United States, and Vector Laboratories, Burlingame, California, United States). Standard protocols for embryo and imaginal disc staining were followed [27]. Images were acquired and processed using a Zeiss LSM 510 confocal microscope (Carl Zeiss, Oberkochen, Germany) and LSM image software.
Standard procedures were followed. DIG-labelled LP10384 was used as a tal RNA probe, and DIG-labelled 4H-3 rn cDNA fragment was used as a rn probe [25].
The tal constructs are based on the LP10384 cDNA cloned in the pOT2 vector. Primer sequences and detailed strategies are available on request. The AB construct was made by digestion of the LP10384 cDNA with BamHI, which cuts in equivalent positions within the conserved regions of the ORF 1A and the last LDPTGXY motif of the ORF AA. The fragment containing the vector and most of the LP10384 sequence was ligated, resulting in a single type-A ORF that codes for a peptide identical to 1A. The rest of the mutant constructs were made by PCR, with primers containing directed mutations and/or restriction sites for ligation. With this strategy, we avoid any alterations to the rest of the cDNA, including UTRs and regions between the ORFs. For the Bombyx construct, the wdS20994 cDNA has been cloned into pPUASt. For the 1A-GFP construct, the sequence of GFP was amplified by PCR from the pEGFP vector with internal primers so that the fragment did not contain start or stop codons, and with a BamHI adapter site. This fragment was BamHI digested and cloned into BamHI linearised AB construct. For the 2A-GFP and 3A-GFP, a SpeI site was introduced at the end of the LP10384 ORF 2A and ORF 3A by directed mutagenesis, then linearised, and the GFP sequence flanked by SpeI adaptors was introduced in frame. For the AA-GFP, a SpeI site was introduced in the middle of the ORF AA, between the two conserved LDPTGXY motifs, by directed mutagenesis, then linearised, and the GFP sequence flanked by SpeI adaptors was introduced in frame in LP10384. For the B-GFP construct, a similar strategy was employed, by introducing a KpnI site in ORF-B. For the generation of transgenic flies or transfection into S2R+ cells, these constructs were excised by double digestion with EcoRI and XhoI, and directionally cloned into pPUASt.
These were carried out using the TNT Quick Coupled Transcription/Translation reticulocyte system (Promega, Madison, Wisconsin, United States). The pool of proteins was separated by PAGE, and incorporation of [35S]-Met allowed the detection of the translated products by autoradiography.
Drosophila S2R+ cells were grown in Schneider's Drosophila medium (Invitrogen, Carlsbad, California, United States) with 10% heat-inactivated foetal bovine serum, 50-units/ml penicillin, 50-μg/ml streptomycin (Invitrogen) at 24 °C. S2R+ cells were removed from the culture flask with Trypsin-EDTA (Invitrogen). Cells were transiently transfected with 2 μg of DNA using FuGene HD (Roche, Basel, Switzerland). Plasmids transfected were pActin-Gal4, pPUASt-DsRedT4NLS, and the appropriate pPUASt-tal-GFP construct. At 48 h after transfection, cells were washed in PBS, fixed for 20 min in 4% paraformaldehyde, washed twice, stained for 10 min with DAPI (Sigma), washed, and then mounted in Vectashield medium.
Drosophila melanogaster cDNAs were obtained from the Berkeley Drosophila Genome Project (BDGP) collection [22]. tal cDNAs are LD11162 and LP10384. LP10384 sequencing revealed it to be identical to LD11162, with a 5′ UTR just 8 bp longer. For the phylogenetic analysis, homologous sequences were identified with the BLAST engine against several databases and obtained by different strategies. We used the following: for Anopheles gambiae, we obtained from the MR4 Anopheles repository, the cDNA 19600449643540 from the MRA-467–43 library [48]; for Lutzomyia longipalpis, two sequenced cDNAs; Bombyx mori cDNA brP0760 and EST wdS20994, which we obtained from the Silkbase EST collection [49] and sequenced; Apis mellifera genomic contig 15.24; and Tribolium castaneum gene mlpt. For the following species, we assembled contigs from the mentioned sequences: four Bicyclus aniana ESTs; three Homalodisca coagulata ESTs; two Aphis gossypii ESTs; three Acyrthosiphon pisum ESTs; a Locusta migratoria EST; a Daphnia pulex EST; and three genomic traces from the NCBI archive.
The National Center for Biotechnology Information (NCBI) (http://www.ncbi.nlm.nih.gov) accession numbers for the genes and gene products discussed in this paper are as follows: Acyrthosiphon pisum ESTs (CV844847, CV848262, and DY229958); Anopheles gambiae cDNA 19600449643540 (EF427621); Aphis gossypii ESTs (DR391935 and DR396643); Apis mellifera genomic contig 15.24 (NW_001253127); Bicyclus aniana ESTs (DY768921, DY768985, DY769016, and DY770310); Bombyx mori cDNA brP0760 (BP115320); Bombyx mori cDNA wdS20994 (EF427620); Daphnia pulex EST (EE682928); Daphnia pulex genomic traces from the NCBI archive (AZSH294914, AZWZ371589, and AZWZ484121); Drosophila melanogaster cDNA LD11162 (AY070879); Drosophila melanogaster cDNA LP10384 (EF427619); Homalodisca coagulata ESTs (CO641298, DN197711, and DN197836); Locusta migratoria EST (DY229958); Lutzomyia longipalpis cDNAs (AM108347 and AM108346); and Tribolium castaneum mlpt (AM269505).
|
10.1371/journal.pbio.0050211 | Nitroimidazole Action in Entamoeba histolytica: A Central Role for Thioredoxin Reductase | Metronidazole, a 5-nitroimidazole drug, has been the gold standard for several decades in the treatment of infections with microaerophilic protist parasites, including Entamoeba histolytica. For activation, the drug must be chemically reduced, but little is known about the targets of the active metabolites. Applying two-dimensional gel electrophoresis and mass spectrometry, we searched for protein targets in E. histolytica. Of all proteins visualized, only five were found to form adducts with metronidazole metabolites: thioredoxin, thioredoxin reductase, superoxide dismutase, purine nucleoside phosphorylase, and a previously unknown protein. Recombinant thioredoxin reductase carrying the modification displayed reduced enzymatic activity. In treated cells, essential non-protein thiols such as free cysteine were also affected by covalent adduct formation, their levels being drastically reduced. Accordingly, addition of cysteine allowed E. histolytica to survive in the presence of otherwise lethal metronidazole concentrations and reduced protein adduct formation. Finally, we discovered that thioredoxin reductase reduces metronidazole and other nitro compounds, suggesting a new model of metronidazole activation in E. histolytica with a central role for thioredoxin reductase. By reducing metronidazole, the enzyme renders itself and associated thiol-containing proteins vulnerable to adduct formation. Because thioredoxin reductase is a ubiquitous enzyme, similar processes could occur in other eukaryotic or prokaryotic organisms.
| The protist parasites Entamoeba histolytica, Trichomonas vaginalis, and Giardia intestinalis grow in environments with low oxygen concentration. Infections with these parasites are commonly treated with metronidazole, a nitroimidazole drug that must be reduced for activation, resulting in several toxic metabolites. We examined the soluble proteome of metronidazole-treated E. histolytica cells for target proteins of these metabolites, applying two-dimensional gel electrophoresis and mass spectrometry. Of about 1,500 proteins visualized, only five formed covalent adducts with metronidazole metabolites, including thioredoxin, thioredoxin reductase, and superoxide dismutase. Metronidazole-bound thioredoxin reductase displayed diminished activity. In addition to these proteins, small thiol molecules, including cysteine, formed adducts with metronidazole. Supplementation with cysteine allowed the cells to survive otherwise lethal metronidazole concentrations. Finally, we discovered that one of the modified proteins, thioredoxin reductase, reduces metronidazole, suggesting a central role for this enzyme with regard to metronidazole toxicity. Taken together, our work reveals a new area of molecular interactions of activated metronidazole with cellular components. Because thioredoxin reductase is a ubiquitous enzyme, similar processes could also occur in other eukaryotic or prokaryotic organisms.
| Entamoeba histolytica is a microaerophilic protozoan parasite and the causative agent of amoebiasis, a disease that affects millions of people worldwide and claims up to 100,000 casualties per annum [1]. As is the case with other microaerophilic parasitic infections, such as giardiasis (caused by Giardia intestinalis) and trichomoniasis (caused by Trichomonas vaginalis), the 5-nitroimidazole drug metronidazole has established itself as the most effective treatment of amoebiasis. Due to the high prevalence of these infections [2] and due to its role as a second-line defense against Helicobacter pylori infections [3], metronidazole has been included in the “essential medicines” list by the World Health Organization [4]. Metronidazole, like other nitroimidazoles, requires reduction at the nitro group in order to be transformed into its cytotoxic form, the nitroradical anion [5]. The activated nitro group undergoes further reduction so that a nitrosoimidazole is generated [6] which can react with sulfhydryl groups [7] and with DNA [8] while being further reduced to an amine via a hydroxylamine intermediate. In the presence of oxygen, however, the nitroradical anion is suggested to be rapidly reoxidized to its respective parent drug before nitroso intermediates can be formed, i.e., a redox cycling effect also termed “futile cycle” [9]. Despite the resulting oxidative stress, this futile cycle is believed to render metronidazole treatment safe in man. However, there are still concerns regarding its potential carcinogenicity [10].
Since reduction of the nitro group is essential for nitroimidazole toxicity, extensive research has been dedicated to enzymes that can act as metronidazole-activating nitroreductases. In rat liver extracts, the microsomal enzyme NADPH-cytochrome P450 reductase was found to be responsible for nitroimidazole activation [11]. The microaerophilic parasites G. intestinalis, T. vaginalis, and E. histolytica, however, lack mitochondria [12] but depend on substrate-level phosphorylation [13]. In these organisms, ferredoxin, which is being reduced by pyruvate:ferredoxin oxidoreductase (PFOR), has been suggested to activate metronidazole [14]. Indeed, in T. vaginalis, metronidazole activation was found to take place in the hydrogenosome [15], a hydrogen-producing organelle in which PFOR and ferredoxin are localized [16]. Moreover, purified ferredoxin was shown to be able to reduce various nitroimidazoles in vitro [17]; and in some highly metronidazole-resistant laboratory T. vaginalis strains, PFOR and ferredoxin were absent, stressing a direct relationship between ferredoxin and metronidazole activation in vivo [18]. Likewise, PFOR activity [19,20] and ferredoxin levels were reduced in metronidazole-resistant G. intestinalis strains [21]. In partially metronidazole-resistant E. histolytica, expression of ferredoxin 1 was sharply decreased [22] although PFOR levels remained unaltered [23]. In contrast to T. vaginalis [24], metronidazole resistance in E. histolytica could be mainly attributed to the increased expression of the antioxidant enzymes peroxiredoxin [22] and superoxide dismutase [23], rather than to loss of PFOR activity, as observed in the other two parasites.
For several decades, great efforts have been undertaken to deepen the understanding of metronidazole activation in the parasitic cell, but the exact mode of action in vivo of this pivotal drug has remained rather understudied. DNA is suggested to be the major target of metronidazole [14,25], as implied by several in vitro studies addressing metronidazole's mutagenicity and DNA-binding capability [8]. In addition, in vitro adduct formation of nitroimidazoles with proteins and thiols, e.g., cysteine, was also demonstrated [11], but specific targets in the treated parasites were never defined because nitroimidazole action was assumed to be indiscriminate. As a contribution to fill this gap, it was our goal to elucidate the processes that occur in the E. histolytica cell during metronidazole treatment at concentrations that are applied during the treatment of amoebiasis, and, if existent, identify specific targets of metronidazole. After completion of the E. histolytica genome project [26], application of proteomic methods such as two-dimensional gel electrophoresis (2DE) was greatly facilitated, permitting a comprehensive and rapid identification of proteins affected by metronidazole in treated E. histolytica cells.
In this study, we show that, in E. histolytica, activated metronidazole does not bind to protein indiscriminately, but reproducibly forms covalent adducts with a small and defined number of proteins, including enzymes such as thioredoxin reductase, superoxide dismutase, and purine nucleoside phosphorylase, as well as the multiple-role reductant protein thioredoxin. When recombinantly expressed in Escherichia coli BL21 (DE3) in the presence of metronidazole, the capability of thioredoxin reductase to reduce thioredoxin was significantly diminished. Moreover, levels of non-protein thiols, e.g., cysteine, were found to be drastically lowered in metronidazole-treated E. histolytica cells due to adduct formation between activated metronidazole and accessible sulfhydryl groups. In accordance with this finding, addition of cysteine to the growth medium allowed the cells to survive otherwise lethal metronidazole concentrations and significantly reduced protein adduct formation. Finally, we propose an alternative mode of metronidazole activation by thioredoxin reductase, because it showed nitroreductase activity in enzymatic assays.
After having treated E. histolytica trophozoites for different time periods (1 h, 2 h, 3 h, and 6 h) and with varying metronidazole concentrations (10 μM–1 mM), cell lysates were prepared for 2DE experiments. Metronidazole concentrations between 50 μM and 100 μM proved to be the most suitable because cells were viable for more than 5 h, a time span which is sufficient for the cell to react to stress by expression of mRNA and proteins. In addition, therapeutic levels lie within this range. Because higher metronidazole concentrations led to rapid disintegration of the cells, and incubation periods for more than 2 h with 50 μM metronidazole did not reveal any additional changes in the protein profile, we chose exposure to 50 μM metronidazole for 2 h as our standard condition when challenging cells with metronidazole. We reproducibly found seven new protein spots on the gels that were isolated and analyzed by mass spectrometric tryptic peptide fingerprinting in combination with additional verification of selected peptide sequences by tandem mass spectrometry (MS/MS) (identified peptides and tandem mass spectra are listed in Figures S1–S5). These seven spots corresponded to five proteins (Figure 1A and 1B), identified as superoxide dismutase, purine nucleoside phosphorylase, thioredoxin, thioredoxin reductase, and a protein designated as “hypothetical protein XP_650662” in the E. histolytica protein database (Table 1). The last will further be referred to as “metronidazole target protein 1” (Mtp1), since its presence on 2D gels abolishes its hypothetical status. Surprisingly, the new spots did not correspond to newly synthesized protein, but appeared at the expense of other neighboring spots as shifted isoforms at a more basic isoelectric point (pI) (Figure 1A and 1B). The widths of the shifts in pI differed with each of the proteins. Thus, it was hypothesized that metronidazole exposure leads to modification of these five proteins, and that the newly appearing spots correspond to isoforms of pre-existing protein in the cell. Concomitant treatment of the cells with 100 μM cycloheximide to block protein synthesis did not prevent the appearance of the shifts on the gel when cells were exposed to metronidazole (unpublished data). This supported our notion that the observed spots do not correspond to newly synthesized protein. Moreover, we did not find any proteins significantly up-regulated or down-regulated in expression during metronidazole exposure; mRNA expression was also not found to be altered (M. Tazreiter, unpublished data). Thus, presumably, E. histolytica does not react to short-term metronidazole treatment by mRNA synthesis or by synthesis of proteins that are involved in stress response or antioxidant defense. This indicated that the protein shifts might be due to metronidazole adduct formation with the five proteins rather than to a general stress response of the cell.
Unfortunately, mass spectrometric analysis of tryptic fragments did not reveal any metronidazole-bound peptides. Thus, in order to confirm metronidazole binding to the proteins, we treated the cells with other nitroimidazoles, including the 2-nitroimidazole azomycin, and the two 5-nitroimidazoles tinidazole and ornidazole (Figure 2A). Since cross-resistances were reported for most of the nitroimidazoles [2], we expected all nitroimidazoles to give similar results as compared to metronidazole. Again, treatment regimens with 50 μM of the respective nitroimidazole for 2 h were chosen as the experimental conditions. Indeed, two-dimensional (2D) gels revealed that the same proteins were affected, but the width of the pI shifts to the basic differed according to the varying pKa's of the nitroimidazoles, which can be attributed to the different side chains at the N1 position. Protein shifts upon ornidazole treatment were slightly narrower, whereas protein shifts with tinidazole were considerably narrower than those observed with metronidazole (Figure 2B). Interestingly, the 2-nitroimidazole azomycin also shifted the same proteins. With regard to pI interval, the shifts by azomycin were wider than shifts by the other nitroimidazoles tested, but the amount of the respective proteins shifted was smaller. The modification of proteins seemed to be independent of the position of the nitro group in the ring, indicating a generalized pattern of nitroimidazole action in E. histolytica.
Quantitative analysis with Melanie 2DE imaging software indicated that, even after prolonged incubation, only a defined fraction of each protein was shifted until a certain maximum was reached (Table 1). Thus, higher nitroimidazole concentrations only allowed this maximum to be more rapidly attained (unpublished data). The proteins were also modified to the same extent in partially metronidazole-resistant E. histolytica cells that had been continuously cultured in the presence of 10 μM metronidazole (unpublished data). Like others before [22,23], we were not able to obtain highly metronidazole-resistant E. histolytica strains. Such would have been very helpful for an assessment of the impact of the modifications on metronidazole-mediated toxicity.
Because nitroimidazoles have been found to form adducts with sulfhydryl group containing compounds [7], we tested whether metronidazole treatment could reduce the levels of non-protein thiols in the cell, e.g., cysteine, which constitutes the major reductant in E. histolytica [27,28]. Treatment with 50 μM metronidazole for 2 h decreased non-protein thiol levels to 49% (Figure 3A) of those in untreated cells (11 fmol/cell), whereas additional 50 mM cysteine (17 mM cysteine constitutes the standard concentration in TYI-S-33 medium), led to accumulation of cysteine in the cell and raised non-protein thiol levels by about 180%, to approximately 31 fmol/cell. When metronidazole and cysteine were used in combination, thiol levels were also sharply decreased (54% of the untreated state). Since cysteine could not accumulate intracellularly in the presence of metronidazole (the observed 5% difference between total free-thiol levels of cells treated with metronidazole alone and those treated with metronidazole in combination with cysteine, is statistically not significant), these results indicate that cysteine levels in the cell are diminished by metronidazole.
As the observed decrease in non-protein thiol levels could also be the consequence of oxidative stress, e.g., oxidation of sulfhydryl groups by hydrogen peroxide, we attempted to verify covalent adduct formation of metronidazole with sulfhydryl groups by exposing the cells to 50 μM metronidazole under anaerobic (Merck Anaerocult A, 0% O2 and 18% CO2), microaerophilic (Merck Anaerocult C, 5% O2 and 8% CO2), and aerobic conditions (aerated vials), respectively. We reasoned that, if the decrease in thiol levels were to be attributed to oxidative stress, exposure of the cells to metronidazole under aerobic conditions would result in even more strongly reduced thiol levels, whereas, under anaerobic conditions, the observed effect should be significantly mitigated or even absent. However, our experiment showed that the reduction of thiol levels, to 36% of the level of untreated cells, was the most strongly pronounced after metronidazole exposure under anaerobic conditions (Figure 3B), followed by exposure to metronidazole under microaerophilic conditions, with 51%, and then under aerobic conditions, with 54%. Thus, oxygen did not promote, but conversely, counteracted the reduction of thiol levels upon metronidazole exposure, which strongly suggests that covalent adduct formation of activated metronidazole with thiol groups is the reason for the decrease in non-protein thiol levels. Moreover, these results are in line with the finding that oxygen can detoxify nitroradical anions and regenerate the parent drug by snatching the electron from the nitro group [9]. However, this effect was not as pronounced as anticipated, because reoxidation of the metronidazole radical anion by oxygen was incomplete, leading to almost halved non-protein thiol levels, even under aerobic conditions. Interestingly, adduct formation with the five proteins found was not affected in the presence of higher oxygen concentrations when cells were treated with 50 μM metronidazole for 2 h (unpublished data).
In order to evaluate the extent to which the observed decrease in non-protein thiol levels contributes to metronidazole toxicity, we treated cells with azomycin, which had been found to be by far less toxic to E. histolytica than metronidazole (D. Leitsch, unpublished data). Indeed, non-protein thiol levels were found to be affected by azomycin, even if considerably less than by metronidazole, paralleling the observation that azomycin shifts the same proteins as metronidazole, albeit to a smaller extent. Under anaerobic conditions, a drop to 65% of the original level when treating cells with 50 μM azomycin for 2 h, as compared to a drop to 36% when treating the cells with metronidazole, was observed (Figure 3B). In the presence of oxygen, non-protein thiol levels after exposure to azomycin were only slightly reduced: 84% (microaerophilic) and 82% (aerobic), respectively, of the non-protein thiol levels were retained as compared to untreated cells. Interestingly, higher oxygen levels counteracted adduct formation of azomycin with non-protein thiol groups more strongly than was the case with metronidazole. As compared to anaerobic conditions, 49% less adduct formation with azomycin (a drop of non-protein thiols of only 18%—to 82% of the untreated sample—instead of a drop of 35%—to 65%) and 28% less adduct formation with metronidazole (a drop of 46%—to 54% of the untreated sample—instead of a drop of 64%—to 36%) were observed when treating cells under aerophilic conditions. Differences between microaerophilic and anaerobic conditions with regard to their impact on the decrease of non-protein thiol levels in the presence of metronidazole or azomycin were hardly significant, if existent.
Although non-protein thiol levels were only insignificantly higher in cells that were treated with metronidazole and cysteine in combination, as compared to non-protein thiol levels in cells treated with metronidazole alone, we observed that the cells were not rounding off and disintegrating in the presence of higher cysteine levels. Therefore, we tested whether raised cysteine levels in the medium could protect E. histolytica during metronidazole treatment. Cells were treated with 30 μM or 50 μM metronidazole either in the presence or absence of additional 50 mM cysteine (67 mM in total) (Figure 3C). Addition of 50 mM cysteine slightly impaired viability; only 83% of the cells were still viable after 20 h of incubation. In the cultures treated with 30 μM metronidazole alone, only 23% of the cells were still viable after the same time span, whereas 50 μM of metronidazole was sufficient to kill almost all cells in the culture (2% viable cells). However, 86% of the cells in the culture treated with 30 μM metronidazole and 71% of the cells treated with 50 μM metronidazole were still viable after 20 h when 67 mM cysteine was present in the growth medium. These results clearly indicate that cysteine strongly counteracts metronidazole toxicity in E. histolytica.
We speculated that the protective effect of cysteine might be due to less metronidazole adduct formation with proteins, because we expected free cysteine to compete with proteins for activated metronidazole. Thus, we assessed the influence of raised cysteine levels on adduct formation of metronidazole with the five proteins identified. Additional 50 mM cysteine markedly reduced metronidazole adduct formation with protein when cells were treated with 50 μM metronidazole for 2 h (Figure 3D). The shifts in pI of superoxide dismutase, thioredoxin, and Mtp1 were clearly diminished in this case, whereas thioredoxin reductase was only subject to putatively one modification and purine nucleoside phosphorylase remained unmodified. The latter two observations were not always made when repeating the experiment, but a very distinct decrease of adduct formation with all five proteins was perfectly reproducible. These results suggest that non-protein cysteine competes with proteins in the formation of adducts with activated metronidazole, and that the observed protective effect of cysteine during metronidazole exposure might be due to fewer protein adducts formed.
As a result of our observation that metronidazole adducts are also formed in Es. coli under microaerophilic conditions, albeit at much higher metronidazole concentrations than is the case in E. histolytica (unpublished data), we decided to recombinantly express E. histolytica thioredoxin reductase and E. histolytica thioredoxin in Es. coli either in the presence or in the absence of metronidazole. We speculated that this strategy would allow us to obtain a large quantity of metronidazole-bound protein that could aid in the mass spectrometric identification of the observed modifications. In addition, the influence of metronidazole adduct formation on protein function could be assayed in vitro with the modified and unmodified recombinant proteins at hand. As thioredoxin and thioredoxin reductase are of central importance for the cell's physiology [29], these two proteins were of particular interest to us. Moreover, human [30] and Arabidopsis thaliana [31] thioredoxin reductases were shown, in addition to their intrinsic disulfide reductase activity, to reduce nitro compounds such as tetryl or 1-chloro-2,4-dinitrobenzene (CDNB). Therefore, we also wanted to determine in vitro whether E. histolytica thioredoxin reductase can, in addition to its role as a disulfide reductase, reduce nitroimidazoles.
Recombinant E. histolytica thioredoxin reductase (recEh TrxR) and recombinant E. histolytica thioredoxin (recEh Trx) were produced in Es. coli BL21 (DE3) cells and purified on Ni-NTA columns via their carboxy-terminal hexahistidine tags. RecEh TrxR had a deep yellowish color that can be attributed to its FAD or FMN cofactor [32].
When Es. coli BL21 (DE3) cells were treated with 1 mM metronidazole during recEh TrxR and recEh Trx expression, the recombinant proteins were efficiently modified by metronidazole as verified by 2DE (unpublished data). In contrast to our failed attempts to identify metronidazole on peptides that were directly isolated from 2D gels, we observed, by liquid chromatography electrospray ionization quadrupole time-of-flight tandem mass spectrometry (LC-ESI-QTOF-MS) of intact proteins, that a high proportion of recEh TrxR and recEh Trx displayed a shift in molecular mass (Figure 4A). The shift in the deconvoluted spectra of both recEh Trx and recEh TrxR (but here only illustrated with recEh Trx), corresponded to a mass gain of 141 Da. This is in good agreement with the in vitro reaction scheme for 5-nitroimidazoles as proposed by Wislocki and colleagues [33] (Figure 4B). In this scheme, the activated 5-nitroimidazole is first reduced to an electrophilic nitrosoimidazole, which is subsequently attacked at its C4 atom by a sulfhydryl group. This is accompanied by further reduction to a hydroxylamine group, which is finally reduced to an amino group. Thus, the nitroimidazole loses two oxygen atoms from the nitro group and one proton from C4, and gains two hydrogens, resulting in a decrease in mass of 31 Da. At physiological pH and under the conditions applied during LC-ESI-QTOF-MS, however, the amino group is protonated, leading to a total mass decrease of only 30 Da. Since metronidazole has a mass of 171.16 Da, a protein that binds metronidazole can be expected to increase in mass by about 140 Da because it gains 141 Da from the bound 5-aminoimidazole derivate of metronidazole but loses the proton of a sulfhydryl group.
When allowing for the methodological restraints of the LC-ESI-QTOF-MS instrument, which has a mass deviation of ±4 Da when analyzing a protein of the size of recEH Trx, the calculated theoretical mass gain of 140 Da after modification by activated metronidazole corresponds well with the observed shift of 141 Da. In order to check whether the proposed model also applies for other 5-nitroimidazoles than metronidazole, we expressed recEh TrxR and recEh Trx in presence of tinidazole, which has a molecular mass of 247.3 Da. The result obtained with metronidazole was paralleled by that with tinidazole (Figure 4A); the shifts in the deconvoluted spectra amounted to 217 Da (molecular weight of the nitroimidazole less 30 Da). Moreover, it is important to add that the proposed model for nitroimidazole binding is also supported by the fact that, on 2D gels, the pI values of all five proteins identified were shifted to the basic, probably due to the reduction of the nitro group of the nitroimidazoles to a basic amino group.
Additional peaks that can be observed on the mass spectra of metronidazole- and tinidazole-bound recEh Trx correspond to mass increments of approximately 16 Da or 32 Da, and very likely can be attributed to oxidation (one or two oxygen atoms, respectively). Not unexpectedly, reactive oxygen species that are generated during nitroimidazole treatment led to oxidation of proteins (e.g., recEh Trx) and, probably, other cell constituents.
Unfortunately, our attempts so far to pinpoint the modifications to specific tryptic peptide fragments from recEh Trx and recEh TrxR have been as unsuccessful as had been our attempts to identify the modifications on E. histolytica proteins directly isolated from 2D gels. Obviously, the nitroimidazole adducts were not stable under the experimental conditions applied, possibly because incubation periods with trypsin (overnight at 37 °C) were too long or the conditions during LC-ESI-QTOF-MS were too harsh. We will, therefore, intensify our efforts in the future and modify the standard protocols accordingly.
Thioredoxin reductase activity of recEh TrxR (applied at a concentration of 148 nM) was verified (Figure 5) by reduction of the disulfide 5,5′-dithiobis-(2-nitrobenzoic acid) (DTNB) to 2-nitro-5-thiobenzoic acid (TNB) via recEh Trx (applied at a concentration of 174 nM). Specific reduction of recEh Trx was determined by subtracting the ground-level reduction of DTNB (206 nmol min−1 mg−1) by recEh TrxR and was found to amount to 559 nmol min−1 mg −1 (which equals a turnover of approximately 23.5 min−1). RecEh TrxR and its substrate, recEh Trx, were used in roughly equimolar amounts because our 2D gels suggested that both proteins are about equally abundant in the cell. Because we wanted to stay as close to physiological conditions as possible, we did not apply such a high excess of the substrate to the enzyme as would be necessary for the exact determination of the kinetic constants of recEh TrxR. Nevertheless, the activity of recEh TrxR, determined by us, is in good accordance with the thioredoxin reducing activity of E. histolytica thioredoxin reductase as determined just recently by Arias and colleagues [34].
We used metronidazole-modified recEh TrxR and recEh Trx for an estimation of the influence of metronidazole on protein function. Thioredoxin reducing activity of metronidazole-modified recEh TrxR dropped by more than 50% to 265 nmol min−1 mg−1 (Figure 5), and when used in combination with metronidazole-modified recEh Trx, the efficiency of recEh Trx reduction by recEh TrxR was even further diminished (220 nmol min−1 mg−1). When assaying metronidazole-modified recEh Trx alone, disulfide reduction lay also clearly below (427 nmol min−1 mg−1) the reduction rate as compared to using unmodified forms of both recEh TrxR and recEh Trx (559 nmol min−1 mg−1).
Using a recEH TrxR concentration of 118 nM (equal to 4 μg/ml), CDNB reduction was determined by measuring NADPH consumption at 340 nm (Table 2). Nitroreductase activity amounted to 233 nmol min−1 mg−1 (7.8 reduction events per minute per molecule of recEh TrxR) when using CDNB as the substrate at a concentration of 100 μM. Unfortunately, the assay was heavily disturbed by the absorbances of the nitroimidazoles. As an alternative method (Table 2), nitroreductase activity of recEh TrxR was indirectly determined by measuring reduction of cytochrome c by reduced nitro compounds [30], or by superoxide radical anions generated by the transfer of an electron from nitroradical anions to oxygen, respectively. In both cases, reduction of cytochrome c could be attributed to the previous reduction of the assayed nitro compounds, i.e., CDNB, the 2-nitroimidazole azomycin, and the 5-nitroimidazole metronidazole, by recEh TrxR (4 μg/ml). As expected, CDNB was readily reduced at a rate of 171 nmol−1 mg−1 at concentrations as low as 10 μM (Table 2). This rate lies in the range of the determined nitroreductase activity of recEh TrxR when measuring NADPH consumption. The 36% higher reduction of CDNB in the first assay is likely due to the 10-fold higher CDNB concentration used. The 2-nitroimidazole azomycin was also readily reduced at a concentration of 100 μM (63 nmol min−1 mg−1). The reduction rate of metronidazole (1 mM) amounted to 31 nmol min−1 mg−1. All values given above have been corrected for the ground-level activity of recEh TrxR, i.e., the reduction of molecular oxygen in the assay buffer, resulting in the formation of superoxide radical anions that, in turn, can reduce cytochrome c. In the absence of recEh TrxR, no reduction of cytochrome c was observed. In contrast to thioredoxin reductase activity, nitroreductase activity was not impaired with metronidazole-bound recEh Trx (unpublished data). This is possibly due to the fact that the flavin cofactor rather than the enzymatic site of recEh TrxR is responsible for nitroreduction [31].
In order to distinguish between direct reduction of cytochrome c by nitroradical anions and between reduction of cytochrome c by superoxide radical anions that had previously been formed by the transfer of the electron of the nitroradical anion to molecular oxygen, superoxide dismutase was added to the reactions in about 20-fold excess (2.5 μM) (Table 2). The addition of superoxide dismutase completely abolished ground-level reduction of cytochrome c by EhTrxR and decreased cytochrome c reduction in the presence of 10 μM CDNB to 47% of the original value. Cytochrome c reduction in the presence of 100 μM azomycin was diminished to 53%, whereas 75% of the original value was retained with 1 mM metronidazole. Higher concentrations of superoxide dismutase did not lead to further decreases in cytochrome c reduction. Thus, in the case of CDNB and azomycin, roughly half of the cytochrome c reduction is mediated by superoxide radical anions, whereas the metronidazole nitroradical anions directly transferred most of the electrons to cytochrome c. These results provide direct evidence for the formation of superoxide radical anions upon nitroimidazole treatment in general and within limits upon metronidazole treatment, and thereby for the generation of oxidative stress in the microaerophilic cell.
In this study, we show for the first time that metronidazole forms adducts with proteins and non-protein thiols in an in vivo model, i.e., a parasite that is commonly treated with this drug. The shifts in masses, found on the mass spectra of metronidazole-bound recEh Trx and recEh TrxR, confirmed the in vitro model for 5-nitroimidazole adduct formation by Wislocki and colleagues [33] (Figure 4B). In contrast to the in vitro data from the late 1970s and early 1980s, however, we found discrete changes in the protein profile of E. histolytica after metronidazole treatment, i.e., modification of five proteins, rather than indiscriminate protein adduct formation. It is possible that there are some more proteins affected, because 2DE does not cover the whole proteome of a given organism. Very large proteins, low-abundance proteins such as transcription factors, and highly hydrophobic membrane proteins that could also potentially form adducts with nitroimidazoles cannot be identified by our approach. Nevertheless, even if allowing for these restraints, the number of affected proteins can be expected to remain small.
Because we made similar observations with Entamoeba dispar (the nonpathogenic relative of E. histolytica), G. intestinalis, T. vaginalis, and Es. coli (unpublished data), we suggest that adduct formation with a defined subset of proteins can take place in any organism that is treated with nitroimidazoles (unpublished data): the presumed reduction of the nitro group to an amino group during adduct formation with the proteins leads to easily discernable shifts on 2D gels to more basic pI values. It is, therefore, interesting to speculate that nitroimidazoles could be an invaluable tool in proteomics, because our data suggest that they allow identification of nitroreductases and associated proteins by shifting them to more basic pI values that can easily be detected on 2D gels.
Apart from forming covalent adducts with proteins, metronidazole also diminishes non-protein thiol levels in the cell (Figure 3A), including that of cysteine. The observed decrease in non-protein thiol levels is due to covalent adduct formation with metronidazole and not due to oxidative stress generated by the activated drug, because the drop in thiol levels was most pronounced in the total absence of oxygen and least pronounced under aerobic conditions (Figure 3B).
Since cysteine is a compound of essential importance to E. histolytica cell physiology [35], and because it is assumed to function as the major reductant in the cell [27], its depletion could contribute to metronidazole toxicity in E. histolytica. On the other hand, cysteine could also predominantly have a protective role because other essential thiols in the cell, such as coenzyme A, might also form adducts with metronidazole. Interestingly, after 2 h of incubation, non-protein thiol levels in cells treated with 50 μM metronidazole and 50 mM cysteine were almost diminished to the same extent as non-protein thiol levels in cells treated with metronidazole alone (Figure 3A). Longer incubation periods with additional 50 mM cysteine in the growth medium, however, might lead to consumption and, consequently, detoxification of metronidazole. This is indicated by the observation that the toxic effect of metronidazole was drastically reduced after addition of 50 mM cysteine, because more than 70% of the cells in a culture survived 20-h exposure to (otherwise lethal) 50 μM metronidazole (Figure 3C). In addition, raised cysteine levels also led to less protein adduct formation after metronidazole treatment (Figure 3D), which indicates the interdependence of non-protein and protein sulfhydryl groups with regard to metronidazole toxicity and which could be an explanation for the protective effect of cysteine during metronidazole exposure.
Treatment of E. histolytica cells with the clearly less toxic 2-nitroimidazole azomycin also led to the modification of the five proteins found (Figure 2B) and a decrease in non-protein thiol levels (Figure 3B), albeit to a far smaller extent. Interestingly, the capability of azomycin to diminish non-protein thiol levels in the cell was also much more decreased by oxygen than was the case with metronidazole. Thus, the lower toxicity of azomycin could be based on its reduced tendency to form adducts with sulfhydryl groups and on its higher reactivity with oxygen (Table 2).
A potentially detrimental effect of metronidazole binding to protein function is indicated by the significant decrease of the thioredoxin reductase activity of recEh TrxR after metronidazole treatment (Figure 5). However, it is also conceivable that the observed oxidation of several amino acids concomitant with nitroimidazole binding, as observed with recEh Trx (Figure 4A), contributes to a diminished enzymatic function.
E. histolytica thioredoxin reductase was found to be a nitroreductase that is able in vitro to reduce CDNB, as well as the nitroimidazoles azomycin and metronidazole (Table 2). In contrast to thioredoxin reductase activity of recEh TrxR, nitroreductase activity of metronidazole-bound recEh TrxR was not decreased (unpublished data), suggesting that nitroreduction is directly exerted by the FAD or FMN cofactor [31,32]. Azomycin (2.4 min−1 at a concentration of 100 μM) was more effectively reduced in the nitroreductase assay than metronidazole (1.2 min−1 at a concentration of 1 mM), possibly due to the higher redox potential of azomycin (E17 = −418 mV) as compared to metronidazole (E17 = −486 mV) [36]. Interestingly, CDNB, which was the compound tested to be most efficiently reduced (6 min−1 at a concentration of 10 μM) by recEh TrxR, exerted only a mildly toxic effect on E. histolytica in our experimental setting—arguably because it does not form adducts with the same proteins as observed with nitroimidazoles and because it does not lead to a reduction of non-protein thiol levels in the cell (D, Leitsch, unpublished data). It is also possible, however, that CDNB does not enter the cell as readily as metronidazole.
When superoxide dismutase was added to the reactions, cytochrome c reduction rates in the presence of CDNB and azomycin were approximately halved (47% and 53%, respectively), whereas in the presence of metronidazole, 75% of the original rate was retained. These findings provide direct evidence for the generation of oxidative stress upon nitroimidazole treatment, because superoxide radical formation is evident. However, reoxidation of metronidazole by oxygen, with only 25% of the metronidazole nitroradical anions reoxidized, was by no means as complete and as rapid as anticipated [9]. Thus, it is doubtful whether the futile-cycle effect is really as influential on metronidazole toxicity as has been suggested.
Although the nitroreductase activity of recEH TrxR is rather low, it is comparable to the nitroreductase activities determined for A. thaliana and mammalian thioredoxin reductases [30,31]. According to our quantitative evaluations of 2D gels from E. histolytica cell extracts, thioredoxin reductase amounts to approximately 0.2% of the total protein in the cell, equaling 10–20 million copies per cell. Our estimate of the concentration of non-protein sulfhydryl groups amounts to 11 fmol/cell, i.e., only a 400-fold excess of non-protein sulfhydryl group levels over those of thioredoxin reductase. It is therefore conceivable that nitroimidazole reduction by thioredoxin reductase plays an important role in the decrease of non-protein thiol concentrations in the treated cell. In this context, it is interesting to note that studies in G. intestinalis have shown that the turnover of metronidazole reduction by purified ferredoxin was also not very pronounced (4 min−1), when purified PFOR instead of cell extract was used as the electron donor for ferredoxin [20].
We suggest that, due to spatial proximity to the reactive nitroimidazole species generated, reduction of nitroimidazoles by thioredoxin reductase renders this enzyme vulnerable to nitroimidazole adduct formation. Thioredoxin, superoxide dismutase, and Mtp1, in turn, can be expected to be localized in proximity to thioredoxin reductase or to interact with thioredoxin reductase. This could render these proteins prone to nitroimidazole modification as well. Proteins that do not interact with thioredoxin reductase are likely to be less affected, because activated nitroimidazoles react with non-protein thiols or other compounds before they can react with proteins that are more distant to the site of nitroimidazole reduction. Thioredoxin needs to be reduced by thioredoxin reductase in order to fulfill its multiple purpose as a reductant protein [29], whereas superoxide dismutase could be required to remove superoxide anion radicals that are indirectly generated by the nitroreductase activity of thioredoxin reductase. Superoxide dismutase might minimize the damage caused by superoxide when being positioned near thioredoxin reductase. Metronidazole target protein 1 (Mtp1) has no close homolog in any other organism whose genome has been sequenced so far, but it contains an O-glycosyl hydrolase domain and displays extended homology to an α-amylase in E. histolytica. Recent research suggests that thioredoxins are also of decisive importance for starch degradation in plants. Very strong evidence has been presented for thioredoxin-mediated regulation of α-amylases in barley grain [37]. Reduction of intramolecular disulfide bonds in amylases renders these enzymes more soluble, which is a prerequisite for amylase activity. Since Mtp1, as a protein of about 14 kDa, contains as many as six cysteines, it is conceivable that it requires thioredoxin in order to be functional. In the case of purine nucleoside phosphorylase, the situation could be different because arsenate reductase activity has been observed with human purine nucleoside phosphorylase [38]. This gives reason to speculate about the potential nitroreductase activity of the corresponding enzyme in E. histolytica. Possibly, purine nucleoside phosphorylase binds to the imidazole ring of nitroimidazole compounds,because purines have an imidazole moiety. However, in the same study, the authors found arsenate reductase activity to strongly depend on reductants, especially DTT, which suggests that purine nucleoside phosphorylase could require reduction by thioredoxin as well. Very surprisingly, we did not find PFOR or ferredoxin among the proteins forming adducts with metronidazole, although PFOR can be readily found on 2D gels when analyzing E. histolytica cell extracts [39]. High percentage (20%) acrylamide gels did not show any shifted proteins in the range of ferredoxin (approximately 6 kDa). It has been reported that, in contrast to T. vaginalis or G. intestinalis, PFOR was not found to be down-regulated in metronidazole-resistant E. histolytica [22,23], which supports the assumption that PFOR might not be involved in nitroimidazole activation in this organism. However, this is questioned by the fact that ferredoxin 1 levels were found to be decreased in metronidazole-resistant E. histolytica [22]. At the moment, we do not have a conclusive explanation for these contradictory results, but it is conceivable that down-regulation of ferredoxin 1 in metronidazole-resistant E. histolytica could also be an accompanying effect of down-regulation of thioredoxin and thioredoxin reductase. As a potential parallel, thioredoxin has been shown to regulate a large number of proteins in plants, including enzymes involved in glycolysis such as aldolase, enolase, glyceraldehyde 3-phosphate dehydrogenase, and triose phosphate isomerase [40]. Thus, loss of thioredoxin activity or diminished thioredoxin levels in the cell could also impair the cellular metabolism, consequently leading to a down-regulation of ferredoxin.
Interestingly, a thioredoxin reductase originally with different annotation had been sequenced before the E. histolytica genome project had started. It was called disulphide oxidoreductase (Eh34) [41] or later, flavin reductase, [22], and was found to have reduced expression in metronidazole-resistant E. histolytica [22]. The very slight differences in the sequences of thioredoxin reductase from the genome project and Eh34 (2% on the DNA level) and the absence of an exact copy of the Eh34 gene in the genome database have not been resolved conclusively (I. Bruchhaus, personal communication). Eh34 was hypothesized by the authors to be involved in metronidazole activation because recombinant overexpression of Eh34 in E. histolytica rendered cells more vulnerable to metronidazole [22]. These unexpected findings of our colleagues strengthen our argument that thioredoxin reductase is involved in metronidazole activation in E. histolytica because down-regulation of thioredoxin reductase, as an enzyme known to be involved in oxidative stress response [42], would otherwise be highly counterproductive during metronidazole exposure that leads to the formation of reactive oxygen species, at least under microaerophilic conditions.
The data gathered prompted us to propose a model of metronidazole action in E. histolytica that implies, apart from generation of reactive oxygen species in the presence of oxygen, that toxicity of metronidazole could be attributed to covalent adduct formation with essential thiols and the proteins described, leading to impaired protein function (Figure 6). Arguably, the formation of covalent adducts and oxidative stress could even intertwine, because metronidazole toxicity was shown to be exacerbated in E. histolytica under microaerophilic conditions as compared to metronidazole treatment in the complete absence of oxygen [43]. At first glance, this is counterintuitive, because our results show that higher oxygen levels lead to less non-protein thiol depletion (Figure 3A). Moreover, azomycin is by far less toxic to E. histolytica than metronidazole although it gives rise to more superoxide anion radicals (Table 2). However, azomycin forms fewer adducts with the five proteins identified, of which three, i.e., superoxide dismutase, thioredoxin reductase, and thioredoxin, are known to be involved in antioxidant defense. We show here that metronidazole-modified recEh TrxR displays a considerably reduced thioredoxin reductase activity. Thus, it is conceivable that the higher sensitivity of E. histolytica to metronidazole in the presence of oxygen is not due to a increased toxicity of metronidazole itself, but due to a reduced tolerance to oxygen because the cell's capability to remove harmful oxidants is impaired. Thioredoxin, for example, requires prior reduction by thioredoxin reductase in order to reduce peroxiredoxin [44,45], an enzyme that oxidizes hydrogen peroxide to water and oxygen. Since, in contrast to thioredoxin reductase activity, the nitroreductase activity of thioredoxin reductase is not diminished by metronidazole binding, superoxide radicals might be continuously generated after reduction of oxygen by activated nitroimidazoles. Superoxide dismutase breaks down superoxide radical anions to oxygen and hydrogen peroxide. The latter would then accumulate due to a reduced peroxiredoxin activity.
In any case, even in the complete absence of oxygen, metronidazole is a potent drug, suggesting that adduct formation of nitroimidazoles with the proteins identified and with non-protein thiol compounds can be very effective in killing E. histolytica cells (Figure 6). In addition, because mRNA expression was unchanged during metronidazole-exposure with 50 μM metronidazole (M. Tazreiter, unpublished data) and higher doses (e.g., 1 mM) of metronidazole rapidly led to the disintegration of cells, we do not believe that DNA damage plays a decisive role in short-term metronidazole-mediated toxicity.
Finally, because thioredoxin reductase is a ubiquitous enzyme, we suggest that our proposed model of metronidazole action might also, at least partly, apply for other organisms. Our preliminary data from 2D gels of, for example, T. vaginalis extracts corroborate the findings presented and discussed in this study, because thioredoxin reductase was identified among the proteins modified in this parasite.
E. histolytica HM-1:IMSS cells were grown axenically at 36.5 °C in culture flasks that were completely filled with TYI-S-33 medium [46] and carefully sealed in order to ensure low oxygen tension. Culture medium was changed every 3 d. Es. coli BL21 (DE3) was grown in LB medium with appropriate antibiotics. Agar plates contained 15 g/l of agar.
When non-protein thiol levels were measured in E. histolytica while applying defined oxygen tensions, culture flasks were only half filled with growth medium, sealed with a vented plug, and then preincubated for 1 d either in tightly sealed jars containing Merck Anaerocult A (Merck, http://www.merck.de/) for anaerobic conditions (0% O2 and 18% CO2) or Merck Anaerocult C for microaerophilic conditions (5% O2 and 8% CO2) or in a normal incubator in the presence of air (21% O2 and 0.4% CO2) for aerobic conditions. Cell cultures were then harvested, followed by the resuspension of the cell pellets in the respective preincubated media and the exposure to metronidazole or to azomycin as stated below.
Cells were harvested by centrifugation at 500g at room temperature for 5 min and then washed twice with PBS to remove residual serum components. Cell lysates for 2DE were prepared as described previously [39]. Briefly, proteins were precipitated with trichloroacetic acid and acetone, and solubilized in a classical buffer containing urea, thiourea, CHAPS, and DTT.
2DE was performed as described previously [39]. Analytical gels were silver stained [47], whereas preparative gels were stained with Coomassie Brilliant Blue R-250. After staining, gels were scanned with an Epson 1680 Pro scanner (Epson, http://www.epson.com/) and analyzed with the Melanie 2D gel analysis software (GeneBio, http://www.genebio.com/).
The excised 2DE spots were destained, digested with trypsin, and analyzed by LC-ESI-QTOF-MS as described previously [48]. For protein identification, the MS/MS data were subjected to database search against the SwissProt database using the Mascot search engine (http://www.matrixscience.com/) and Protein Global Server 2.1 (Waters-Micromass, http://www.waters.com/). A compilation of mass data is given in Figures S1–S5.
The mass of intact proteins with and without metronidazole or tinidazole treatment (see above) was determined using LC-ESI-QTOF-MS. An aliquot corresponding to approx 500–1,000 pmol of purified protein (recombinantly produced in Es. coli) was subjected to liquid chromatography (LC) using a BioBasic C4 column (30 × 0.32 mm; Thermo Electron Corporation, http://www.thermo.com/) using a CAP-LC (Waters-Micromass). Proteins were loaded onto the column in solvent A (water containing 0.1% formic acid) and eluted using a gradient from 0%–70% solvent B (acetonitrile containing 0.1% formic acid) in 30 min. The flow rate was held at 5 μl/min throughout the analysis. Instrument calibration and tuning in the mass range from 500–4,000 Da was achieved using a 2 mg/ml solution of sodium iodide in 50% isopropanol. Spectra were deconvoluted using MaxEnt1 function of the MassLynx 4.0 SP4 software (Waters-Micromass).
The levels of non-protein thiols in metronidazole-treated, azomycin-treated, or in untreated cells were determined as described [49]. Briefly, cells were harvested and washed in 20 mM EDTA. Pellets were resuspended in 20 mM EDTA, and cells were disrupted by repeated freezing and thawing. Cell debris was removed by centrifugation at 20,000g for 15 min. After measuring protein levels with Bradford reagent, equal amounts of protein were precipitated in 5% TCA at room temperature, followed by centrifugation at 20,000g for 15 min. Supernatants containing the non-protein thiols were removed, and the double volume of 0.4 M Tris/HCl (pH 8.9) was added. A total of 1.7 μl 100 mM 5,5′-dithiobis-(2-nitrobenzoic acid) (DTNB) per ml reaction mixture was added. Reduction of DTNB was measured at λ = 412 nm in a Jenway 6505 UV/Vis spectrometer (Δε412 = 13.6 mM−1 cm−1; http://www.jenway.com/en/index.php).
Before addition of various amounts of metronidazole and/or cysteine, total and viable cell numbers were determined by trypan blue exclusion in a Bürker-Türk hemocytometer. After addition of reagents, cells were incubated for 20 h at 36.5 °C. After incubation, total and viable cell numbers were redetermined.
The genes for thioredoxin reductase and thioredoxin were amplified from genomic E. histolytica HM-1:IMSS DNA. Primers were 5′-TAC GTA CGC ATA TGA GTA ATA TTC ATG ATG TTG TGA TTA TCG GC-3′ (TrxR forward) and 5′-TCA TCC AGC TCG AGT TAG TGG TGA TGG TGA TGA TGA GTT TGA AGC CAT TTT TCA CAG-3′ (TrxR reverse) for thioredoxin reductase, and 5′-TAC GTA CGC ATA TGG CTG TAC TTC ATA TTA ACG CTC TTG ATC AA-3′ (Trx forward) and 5′-TCA TCC AGC TCG AGT TAG TGA TGG TGA TGG TGA TGT CGT GTT TCA ACC ATT TGT TTT AAG GCA-3′ (Trx reverse) for thioredoxin. Forward primers include an NdeI restriction site, whereas reverse primers bear an XhoI restriction site and a hexahistidine tag for convenient protein isolation. PCR fragments were ligated into the pET 17b vector (Novagen/VWR, http://www.emdbiosciences.com/html/NVG/home.html). The plasmid sequences were confirmed on both strands by using T7 and pET reverse primers (GATC Biotech, http://www.gatc-biotech.com/en/index.php). The confirmed plasmids were transfected into Es. coli BL21(DE3) cells. Transformants were selected on 20 μg/ml ampicillin. Expression of recombinant proteins was induced by addition of 0.5 mM IPTG. If proteins were to be modified with metronidazole, 1 mM metronidazole was added to the LB medium, and cells were grown in completely filled tissue culture flasks under exclusion of air. Three hours after induction, cells were harvested and then disrupted by vigorous grinding in a mortar. Subsequently, recombinant proteins were purified via Ni-NTA spin columns (Qiagen, http://www1.qiagen.com/). Recombinantly expressed E. histolytica thioredoxin and E. histolytica thioredoxin reductase are referred to as recEh Trx, and recEh TrxR, respectively.
The assay was performed as described elsewhere [50] using recEh Trx and recEh TrxR in combination. The reaction buffer contained 100 mM potassium phosphate (pH 7.5), 1 mM EDTA, 1 mM DTNB, and 0.5 mM NADPH. All reactions were done with 2 μg/ml recEh Trx and 5 μg/ml recEh TrxR. RecEh TrxR oxidized NADPH to NADP in order to reduce the disulfide bond at the active site of recEh Trx, which, in turn, reduced DTNB. Ongoing reduction of DTNB was measured at λ = 412 nm (Δε412 = 13.6 mM−1 cm−1) over a period of 5 min at 25 °C.
Reduction of CDNB by Eh TrxR was measured by determining NADPH consumption at λ = 340 nm (Δε340 = 6.2 mM−1 cm−1). The reaction buffer contained 100 mM potassium phosphate (pH 7.5), 1 mM EDTA, and 0.1 mM NADPH. Reduction of CDNB was determined at a concentration of 100 μM. For reasons of practicability, because nitroimidazoles display high absorbances at λ = 340 nm and thereby heavily disturb the assay, reduction of nitroimidazoles was measured in a modified assay [30] via reduction of cytochrome c at λ = 550 nm (Δε550 = 20 mM−1 cm−1), either directly by reduced nitro compounds (CDNB, azomycin, and metronidazole), or indirectly by superoxide radical anions that are generated when nitroradicals transfer an electron to molecular oxygen in the reaction buffer. In any case, nitro compounds had been previously reduced by recEh TrxR. Therefore, it was assumed that reduction of one nitro group by Eh TrxR subsequently resulted in one reduced cytochrome c molecule. Reaction mixtures contained 100 mM potassium phosphate (pH 7.5), 1 mM EDTA, 0.5 mM NADPH, 50 μM cytochrome c, 4 μg/ml recEh TrxR (equal to 118 nM) and different amounts of CDNB, azomycin, or metronidazole, respectively. Reduction of cytochrome c was measured over a time span of 5 min at 25 °C. In order to assess the proportion of electrons that are directly transferred from the reduced nitro compounds to cytochrome c, 2.5 μM bovine erythrocyte superoxide dismutase (i.e., an approximately 20-fold excess over Eh TrxR) were added to the reactions to remove all generated superoxide radical anions. Bovine erythrocyte superoxide dismutase was purchased from Sigma (http://www.sigmaaldrich.com/).
The National Center for Biotechnology Information (http://www.ncbi.nlm.nih.gov/) accession numbers for the proteins discussed in this paper are as follows: α-amylase (XP_652601), metronidazole target protein 1 (XP_650662), purine nucleoside phosphorylase (XP_655398), superoxide dismutase (XP_648827), thioredoxin (XP_656726), and thioredoxin reductase (XP_655748).
|
10.1371/journal.ppat.1002166 | Crystal Structure of Reovirus Attachment Protein σ1 in Complex with Sialylated Oligosaccharides | Many viruses attach to target cells by binding to cell-surface glycans. To gain a better understanding of strategies used by viruses to engage carbohydrate receptors, we determined the crystal structures of reovirus attachment protein σ1 in complex with α-2,3-sialyllactose, α-2,6-sialyllactose, and α-2,8-di-siallylactose. All three oligosaccharides terminate in sialic acid, which serves as a receptor for the reovirus serotype studied here. The overall structure of σ1 resembles an elongated, filamentous trimer. It contains a globular head featuring a compact β-barrel, and a fibrous extension formed by seven repeating units of a triple β-spiral that is interrupted near its midpoint by a short α -helical coiled coil. The carbohydrate-binding site is located between β-spiral repeats two and three, distal from the head. In all three complexes, the terminal sialic acid forms almost all of the contacts with σ1 in an identical manner, while the remaining components of the oligosaccharides make little or no contacts. We used this structural information to guide mutagenesis studies to identify residues in σ1 that functionally engage sialic acid by assessing hemagglutination capacity and growth in murine erythroleukemia cells, which require sialic acid binding for productive infection. Our studies using σ1 mutant viruses reveal that residues 198, 202, 203, 204, and 205 are required for functional binding to sialic acid by reovirus. These findings provide insight into mechanisms of reovirus attachment to cell-surface glycans and contribute to an understanding of carbohydrate binding by viruses. They also establish a filamentous, trimeric carbohydrate-binding module that could potentially be used to endow other trimeric proteins with carbohydrate-binding properties.
| Human reoviruses bind first with low affinity to a carbohydrate receptor that brings the virus in close proximity to the host cell. This interaction then facilitates high-affinity binding to a second receptor, the tight junction component junctional adhesion molecule-A (JAM-A). While all human reoviruses bind JAM-A, they differ in carbohydrate receptor specificity, and this difference may influence the distinct disease patterns of reovirus serotypes. We present here the structure of the attachment protein of type 3 reovirus in complex with carbohydrates that naturally occur on human cells. Our results show that the protein forms an elongated trimer, with the carbohydrate binding site being located close to the midpoint of the molecule in a fiber-like region. Our findings provide insights into mechanisms of reovirus attachment to cell-surface glycans and contribute to an understanding of carbohydrate binding by viruses. They also establish a filamentous, trimeric carbohydrate-binding module that could potentially be used to introduce carbohydrate-binding properties into other trimeric proteins.
| Viral infections are initiated by specific attachment of a virus particle to receptors at the surface of the host cell. This process, which serves to firmly adhere the virus to its cellular target, is rarely a bimolecular interaction between one viral attachment protein and one receptor. In most cases, several receptors are employed, and recognition events are frequently accompanied by substantial structural rearrangements that serve to expose new binding sites, strengthen the initial interaction, and prime the virus for cell entry. Structure-function analyses of virus-receptor interactions have provided detailed insights into the attachment strategies of viruses belonging to several different families [1]–[18]. However, much less is known about structure-function interrelationships between different binding sites for distinct receptors on the same viral attachment molecule.
Reoviruses are useful experimental models for studies of virus-receptor interactions and viral pathogenesis. Moreover, the recent development of plasmid-based reverse genetics for reovirus provides an opportunity to manipulate these viruses for oncolytic and vaccine applications. Reoviruses form icosahedral particles approximately 850 Å in diameter. At the virion five-fold symmetry axes, the trimeric attachment protein, σ1, extends from pentameric turrets formed by the λ2 protein. A similar arrangement of a trimeric attachment protein inserted into a pentameric base is also observed for the adenovirus attachment protein, fiber. The σ1 protein is about 400 Å long and consists of three discrete domains, termed tail, body, and head [19]. Residues 1 to 160 encompass the tail domain, which partially inserts into the virion capsid [20]–[22]. This region of the molecule is predicted to form an α-helical coiled-coil structure. The body domain encompasses residues 170 to 309 and contains β-spiral repeat motifs [22]. Lastly, the globular head domain incorporates residues 310 to 455 and folds into an 8-stranded β-barrel [22], [23].
Reovirus attachment is thought to proceed via a two-step adhesion-strengthening mechanism, in which σ1 first engages widely distributed carbohydrate receptors with lower affinity. The three prototype reovirus strains, type 1 Lang (T1L), type 2 Jones (T2J), and type 3 Dearing (T3D) recognize different carbohydrate structures, which may account for the serotype-specific differences in routes of spread in the host and end-organ tropism. In the case of serotype 3 (T3) reoviruses, the carbohydrate bound is α-linked sialic acid [24]–[26]. This initial contact, which has lower affinity and may allow for lateral diffusion of the particle at the membrane [27], is followed by high-affinity interactions with junctional adhesion molecule-A (JAM-A) [28], a component of tight junctions [29]–[31]. All reoviruses, including prototype and field-isolate strains, use JAM-A as a high-affinity receptor [28], [32], [33]. Firm adherence to the cell triggers uptake of the particle, which is dependent on β1 integrins [34], [35].
Discrete regions of σ1 mediate binding to its cell-surface receptors. Structural and functional analyses show that the σ1 head, which projects farthest from the virus capsid, engages JAM-A [33], [36], [37]. In contrast, sequences in the σ1 body bind to carbohydrates [38]. Sequence analysis of reovirus variants identified three residues, Asn198, Arg202, and Pro204, as likely critical for the interaction of T3 σ1 with sialic acid. These residues lie near the midpoint of the protein, at the lower end of the body domain, about 100 Å away from residues in the head that interact with JAM-A. Earlier structural analyses of T3D σ1 [22], [23], [36] were based on constructs that did not include this putative carbohydrate-binding site. It is therefore currently unclear how σ1 achieves its specificity for sialic acid, whether the large distance between the two receptor-binding sites on σ1 is relevant for binding, or whether σ1 undergoes rearrangements after engaging its carbohydrate receptor.
To enhance an understanding of mechanisms by which viral attachment proteins engage cell-surface glycans, we determined the crystal structure of T3D σ1 in complex with α-2,3-sialyllactose, α-2,6-sialyllactose, and α-2,8-disiallylactose. All three carbohydrates terminate in sialic acid but feature different linkages that are present in various physiologic glycans. In addition, we used plasmid-based reverse genetics to engineer reoviruses that express mutagenized forms of σ1 to define residues required for functional binding to sialic acid. These studies shed light on the structural basis of σ1-sialic acid interactions and define a new carbohydrate-binding structural motif in a viral attachment protein.
The σ1 protein belongs to a class of fiber proteins constructed from triple β-spirals, a motif that was first identified in the adenovirus fiber [39]. In a previous study, we crystallized a smaller region of σ1, spanning residues 246 to 455 and containing three β-spiral repeats as well as the globular head domain [22]. While this structure provided no insights into the carbohydrate-binding region of σ1, it served as a basis to predict that β-spiral repeats form the entire body domain of the protein (residues 167–309) [22]. Near residue 170, the body domain transitions into a long α-helical coiled-coil region that forms the N-terminal tail domain (residues 1–156).
To determine the structure of a longer fragment of σ1 including the predicted sialic-acid binding residues 198, 202, and 204, we designed a construct for the expression of residues 170–455. This construct excluded the long α-helical coiled-coil region to simplify protein expression, purification, and crystallization. Prototype strain T3D σ1 is sensitive to trypsin-mediated cleavage after Arg245 [40]. However, a sequence polymorphism occurring in the majority of T3 field-isolate strains, Thr249Ile, renders the protein resistant to trypsin [40]. A construct containing Ile249 was therefore used in our study. Trimerization was promoted by using a hexahistidine-tagged trimerization domain, a modified GCN4 sequence [41], at the N-terminus of the expressed protein. This domain was proteolytically removed before final purification and crystallization.
The structure of σ1 residues 170 to 455 reveals a highly elongated, symmetric trimer that measures about 200 Å in length (Table 1 and Figure 1A,B). Tail residues N-terminal to amino acid 170, which were not included in the crystallized protein, are predicted to form an α-helical coiled-coil structure that adds another 200 Å in length to the protein (Figure 1C). As expected, the structure of the globular head domain (residues 310 to 455) is essentially identical to that described previously [22]. However, the body domain displays a number of unusual features. Although sequence-based predictions suggested that this region would be composed of eight consecutive triple β-spiral repeats [22], we find that the body domain contains a mixture of α-helical coiled-coil and β-spiral repeats (Figure 1). Four β-spiral repeats at the N-terminus (β1–β4, residues 170 to 235) are followed by a short α-helical coiled-coil (cc, residues 236 to 251) and three additional β-spiral repeats (β5-β7, residues 252 to 309) (Figure 2). Inspection of the sequence indicates a likely reason for the deviation from the β-spiral fold at the center of the body (Figure 2B). Three hydrophilic residues (Thr236, Ser244, and Ser252) are located at positions that are typically occupied by hydrophobic side chains in β-spirals. Moreover, Ser241 replaces a characteristic proline or glycine at the turn in a β-spiral repeat. While some deviations from the β-spiral consensus sequence can be tolerated, even residues replacing the glycine or proline (e.g., residues Gln224 or Thr278), the cumulative effect of the four non-consensus residues results in a β-spiral no longer being the optimal fold. The α-helical coiled-coil structure contains two heptad-repeat sequences, starting with Phe239 and ending with Gln251 (Figure 2A,C).
To elucidate the structural basis of the interaction of the reovirus attachment protein σ1 with its carbohydrate coreceptor, we prepared a complex by soaking crystals of σ1 with 10 mM α-2,3-sialyllactose, a compound that terminates in α-linked sialic acid. The subsequent structure, determined at 2.25 Å resolution (Table 1), unambiguously demonstrated the location of the carbohydrate in an unbiased difference electron-density map (Figure 3A). The oligosaccharide binds in a shallow groove next to the loop connecting the second and third β-spiral repeats. The σ1 protein contains three identical binding sites, one on each chain, and all three are occupied by α-2,3-sialyllactose molecules, with the sialic acid making identical and extensive contacts in each chain (Figure 3B,C). The lactose moieties face different directions, probably as a result of internal flexibility and participation in crystal contacts (Figure 3C).
Sialic acid contains four characteristic functional groups: a carboxylate at C1, a hydroxyl group at C4, an N-acetyl group at C5, and a glycerol chain at C6. All four groups are recognized by σ1 (Figure 3B). Arg202 forms a bidentate salt bridge with the carboxyl group. A single hydrogen bond links the hydroxyl group at C4 to the carbonyl of Gly205. The amide of the N-acetyl group is engaged in a hydrogen bond with the backbone carbonyl of Leu203, and the N-acetyl methyl group is facing into a partially hydrophobic cavity. The glycerol chain lies parallel to the peptide backbone, forming direct hydrogen bonds with the backbone carbonyl of Ile201 and the amide nitrogen of Leu203 and in some of the binding sites water-mediated hydrogen bonds with the Asn210 side chain and the amide nitrogen of Ile211. We note that Arg202, which was previously shown to influence sialic acid binding [42], provides a key contact to the ligand. Moreover, Pro204, which also had been implicated in sialic acid binding [42], is part of a structure that shapes the ligand-binding site.
As contacts in the complex of σ1 with α-2,3-sialyllactose exclusively involve the sialic acid moiety, we hypothesized that σ1 should be capable of binding sialic acid in different naturally occurring linkages, including α-2,6- and α-2,8-linked sialic acid. We therefore determined crystal structures of σ1 in complex with α-2,6-sialyllactose (Figure 4A) and α-2,8-disialyllactose (Figure 4B). Refinement statistics for both structures are provided in Table 1. In each case, only two of the binding sites are occupied, as the third is partially blocked by crystal contacts. For the α-2,6-sialyllactose complex, the electron density allowed us to unambiguously identify all three sugar residues (Figure 4A). The electron density for the α-2,8-disialyllactose complex did not allow us to model the terminal glucose. Comparison of these structures with each other and with the α-2,3-sialyllactose complex shows that the terminal sialic acid is bound in the same conformation and with identical contacts in all three cases. However, the remaining moieties of the glycans differ in conformation and contacts with σ1. The α-2,3-sialyllactose and α-2,8-disialyllactose ligands assume an elongated shape in which the lactose groups face away from the protein (Figure 3C, Figure 4B). Inspection of the α-2,8-disialyllactose complex shows that the N-acetyl group of the second sialic acid forms a hydrogen bond to the side chain of Ser195. In contrast, σ1 binds α-2,6-sialyllactose in a folded-back conformation (Figure 4A). This conformation is stabilized by an intramolecular hydrogen bond and the galactose O2 and O3 hydroxyl groups, which form hydrogen bonds to the backbone carbonyl atoms of Ser195 and Leu194, respectively.
To identify sequences that influence sialic acid binding, we used plasmid-based reverse genetics [43], [44] to introduce point mutations into the σ1 protein of reovirus strain T3D. Mutant viruses were isolated following co-transfection of murine L929 cells with RNA-encoding plasmids corresponding to the T3D L1-L3, M1-M3, and S2-S4 genes and a plasmid corresponding to the σ1-encoding S1 gene incorporating site-specific mutations. Thus, each recombinant virus is isogenic, with the exception of the S1 gene and its protein product, σ1. Guided by the structure of the σ1-sialic acid complexes, we engineered individual alanine substitutions of amino acids ranging from Asn189 to Asn210. By their location in the structure, we hypothesized that these residues would be required for functional sialic acid binding. In addition, substitutions N198D, R202W, and P204L, which have been implicated in sialic acid binding by sequence comparisons of reovirus strains that differ in sialic acid utilization [26], [45] and genetic analysis of reovirus mutants adapted to growth in murine erythroleukemia (MEL) cells [42], were engineered to define the effect of these polymorphisms in an otherwise isogenic background.
After confirming the σ1-encoding S1 gene nucleotide sequences, the mutant viruses were tested for hemagglutination (HA) capacity (Figure 5) and growth in L929 cells and MEL cells (Figure 6). In comparison to rsT3D, rsT3D-σ1N198D, rsT3D-σ1R202A, rsT3D-σ1R202W, rsT3D-σ1L203A, rsT3D-σ1P204A, rsT3D-σ1P204L, and rsT3D-σ1G205A produced little or no agglutination of calf erythrocytes, a sensitive assay for sialic acid binding [26]. However, rsT3D-σ1N189A, rsT3D-σ1S195A, and rsT3D-σ1N210A produced HA titers that were comparable to those of wild-type rsT3D. Each of the point-mutant viruses produced approximately 1000-fold yields of viral progeny after growth in L929 cells (Figure 6), a cell line that does not require sialic acid binding for reovirus to replicate [45]. In contrast, those containing mutations N198D, R202A, R202W, L203A, P204A, P204L, and G205A displayed attenuated growth in MEL cells (Figure 6), a cell line permissive only to sialic acid binding reovirus strains [45]. These findings indicate that viruses with mutations of residues 198, 202, 203, 204, and 205 are altered in sialic acid binding efficiency, suggesting that these residues serve a functional role in T3D σ1-sialic acid interactions.
Although all known reovirus strains engage cells by binding to the tight junction protein JAM-A [33], the major reovirus serotypes differ in the routes of dissemination in the host and tropism for host tissues [46]–[48]. These differences are linked to the σ1-encoding S1 gene segment and most likely attributable to serotype-specific interactions of σ1 with different cell-surface receptors. T3 reoviruses require sialic acid as a coreceptor, but the context in which sialic acid is bound is unknown. To define this interaction, we determined crystal structures of reovirus σ1 in complex with three sialylated glycans that incorporate a terminal sialic acid moiety in different linkages. These structural analyses were complemented with mutagenesis experiments that establish the physiologic relevance of the observed interactions.
The σ1 protein uses a complex network of contacts to engage terminal sialic acid, which is a common feature of all three glycans studied here. The interactions involve σ1 residues at the lower end of the body domain, between β-spirals 2 and 3. At this location, the sialic acid moiety docks into a shallow pocket that is formed mainly by residues in the third β-spiral. All four functional groups of sialic acid make contacts with σ1 through an elaborate network of hydrogen bonds and van der Waals interactions. Mutations that alter these contacts lead to significantly reduced sialic acid binding as assessed by HA profiles and diminished infection of MEL cells. Although all three ligands used for complex formation with σ1 contain additional carbohydrates, these make very few interactions. The complex with α-2,8-disialyllactose identified a hydrogen bond between the N-acetyl group of the second sialic acid and the side chain of Ser195 (Figure 4B). However, the results from mutagenesis experiments demonstrate that a Ser195A mutation has no effect on either HA capacity or viral growth. Therefore, the observed contact is unlikely to have physiologic relevance. The interactions between σ1 and α-2,6-sialyllactose identified two hydrogen bonds that link the galactose to the protein and may help to stabilize the folded-back conformation of the ligand (Figure 4A). As both contacts involve main chain atoms of σ1, their functional significance cannot be easily probed by site-directed mutagenesis. Nevertheless, it is likely that the observed contacts lead to a modest increase in the affinity of σ1 for compounds terminating in α-2,6-linked sialic acid. It is unclear if such an increase is biologically significant.
Naturally occurring sequence variability at three amino acid positions (residues 198, 202, and 204) has been linked to the sialic acid-binding capacity of T3 σ1 [26], [42]. Our structures readily identify two of these residues, Arg202 and Pro204, as key determinants of sialic acid binding. The side chain of Arg202 forms a salt bridge with the sialic acid carboxylate group, while the Pro204 side chain stacks against the Arg202 guanidinium group. Moreover, the carbonyl oxygen in the peptide bond linking Leu203 and Pro204 forms a hydrogen bond with the sialic acid. Substitutions of either Arg202 or Pro204, as seen in the R202W and P204L variants, would decrease the affinity for sialic acid, and this is confirmed by the mutagenesis data. In contrast, the critical role of residue 198 in ligand recognition is not apparent from the crystal structures. Our mutagenesis data (Figure 5 and Figure 6), in conjunction with previous results [42], clearly demonstrate that Asn198 is required for successful sialic acid-dependent infection, with viruses carrying an N198D mutation having substantially reduced infectivity in MEL cells. However, the crystal structures show that Asn198 is not involved in direct or water-mediated contacts to any of the three oligosaccharides. Furthermore, the Asn198 side chain is solvent-exposed, forming a single hydrogen bond with the Asn189 side chain. Mutation of Asn189 to alanine does not affect sialic acid binding (Figure 5 and Figure 6), suggesting that the observed Asn198-Asn189 hydrogen bond is not relevant for ligand recognition. It is possible that the introduction of a negatively charged side chain at position 198, as is the case with the N198D mutation, leads to long-range electrostatic effects or structural rearrangements that indirectly affect receptor binding. However, given the distance of Asn198 from the binding site and its surface-exposed location, this possibility appears remote. We think it more likely that Asn198 serves as a contact point with a part of the functionally relevant glycan, which has not been included in the structural analysis. Although our results define the interactions of σ1 with terminal sialic acid, the actual receptor may be a more complex sialylated glycan, perhaps carrying several branches. Such complex receptor structures, which can be attached to proteins or lipids, have recently been identified as the true ligands for several adenovirus and polyomavirus capsid proteins [16]–[18]. Therefore, Asn198 may well define a second receptor contact point for reovirus σ1.
A large collection of structures of viruses or viral attachment proteins in complex with sialylated oligosaccharide receptors is available, and these have produced significant insights into mechanisms of sialic acid binding, receptor specificity, and viral pathogenesis [1]–[3], [5], [9], [11], [14], [16]–[18], [49]–[52]. However, the interactions observed between T3D σ1 and sialic acid differ in important ways from those found in all other virus-receptor complexes, offering new insights into the parameters that guide viral attachment and specificity. In all cases in which structures are available, the receptors are bound by a globular domain in a region that projects farthest from the viral capsid and is easily accessible for interactions with the cell surface. In contrast, the highly elongated T3D σ1 protein engages its carbohydrate ligand at its midpoint, about 150 Å away from the region that projects farthest from the virion. Although the σ1 protein possesses some flexibility at defined regions [19], [22], the location of the sialic acid-binding site would not appear optimal for engagement of membrane-bound receptors that feature sialylated ligands close to the membrane. The region of JAM-A that is engaged by the σ1 head domain is fairly close to the membrane [36]. Even when allowing for considerable flexibility between the σ1 head and body, it is difficult to envision a conformation in which the tail of σ1 is still inserted into the virus and the sialic acid binding site can closely approach the membrane. However, σ1 could more easily engage sialic acid that projects far above the membrane, perhaps by being located on a large protein or projecting from prominent loops.
Prior to this study, structural information had been available only for the C-terminal portion of the σ1 protein [22]. Based on analysis of that structure, as well as sequence comparisons with the related adenovirus fiber protein, full-length σ1 was predicted to fold into three distinct regions: an N-terminal α-helical coiled coil (termed the tail), a region containing eight consecutive β-spiral repeats (the body), and a globular β-barrel (the head). Our structural analysis of a fragment comprising the body and head domains show that this model must be revised, as we find an insertion of a short α-helical coiled coil that interrupts the β-spiral sequence in the body, replacing one β-spiral repeat with a helical structure. Thus, it is clear that the structure of σ1 features several transitions between α-helical and β-spiral regions. This topological relationship differs from that of the adenovirus fiber, in which the shaft domain is thought to consist entirely of β-spiral repeats [39]. Examination of the T3D body domain sequence shows that it contains a nearly perfect heptad repeat pattern, which is typical for α-helical coiled coils, in a short stretch of 14 residues (Figure 2). A similar pattern is observed in the T1L and T2J σ1 sequences, but a proline residue within the consensus makes it unlikely that these proteins also feature a continous α-helical coiled coil at the equivalent location.
To our knowledge, the structures presented here are the first examples of any fibrous viral protein engaging a ligand via its repetitive fiber region. Other viral attachment proteins contain fibrous- or stalk-like structures, but they usually engage receptors with globular head domains placed on top of these structural elements, as observed in complexes of adenovirus fiber proteins with their receptors [7], [15], [18]. Globular head domains offer higher variability in engaging ligands and can more easily create recessed binding pockets suitable for high-affinity binding. Instead, fiber-like structures generally feature short connections between their repeating units and a relatively flat surface, limiting binding options. However, inspection of the β-spirals in σ1 reveals subtle modifications in a single repeat that allow it to create a shallow binding site for sialic acid. One of the hallmarks of β-spirals is a highly conserved β-turn between two strands, involving residues at positions g, h, i, and j (Figure 2). The residue at position j is usually a proline or glycine. This turn is enlarged by two amino acids in the σ1 repeat that engages sialic acid, transforming the turn into a small loop (Figure 7). Interestingly, Pro204 introduces a kink after a β-strand, causing the chain to deviate from the β-spiral motif at this position to provide a pocket for the ligand. Thus, alteration of the typical repeating motif identifies a ligand-binding site in the case of σ1. It is conceivable that similar aberrations in other fibrous protein sequences might also indicate binding sites. The location of a sialic acid binding site in an elongated fiber-like structure also raises the possibility of creating a small sialic acid binding cassette that could be transferred into a variety of trimeric fiber-like proteins constructed from α-helical coiled coils or β-spirals. Our work thus enhances an understanding of reovirus-glycan interactions and may also guide the construction of new sialic acid binding platforms to facilitate structure-function analyses and sialic acid-mediated cell targeting.
The expression of soluble and properly folded T3D σ1 trimers was facilitated by appending a trypsin-cleavable trimerization domain based on the GCN4 leucine zipper [41] N-terminally to a cDNA encoding the entire σ1 body and head domains (amino acids 170–455). The construct was cloned into the pQE-80L expression vector, which encodes a non-cleavable N-terminal hexahistidine-tag. The protein was expressed in E. coli Rosetta 2 DE3 (Novagen) at 20°C for 16 h post-induction or by autoinduction at 20°C for 48–72 h. Bacteria were lysed by two passages through an EmulsiFlex (Avestin) homogenizer and purified by Ni-IMAC using His-Trap-FF columns (GE-Healthcare). The immobilized protein was eluted by on-column digestion with 0.1 mg/ml trypsin at a flowrate of 0.1 ml/min for 12 h. Size-exclusion chromatography (Superdex-200, GE-Healthcare) was used as the final purification step.
Crystals were grown using 15% PEG200, 0.1 M MES (pH 6.5) as a precipitant. The crystals belong to space group P21212 and contain one trimer in the asymmetric unit. Complexes with carbohydrate ligands were prepared by soaking crystals with the respective carbohydrate prior to data collection. The crystals were transferred into mother liquor supplemented with 10 mM carbohydrate, incubated for 5 min, and cryoprotected by incubation for 15 s in 35% PEG200, 0.1 M MES, 10 mM carbohydrate (pH 6.5).
Diffraction data were collected at the beamlines PXI (SLS) and ID14-4 (ESRF). Diffraction data were integrated and scaled using XDS [53], and the structure was solved by molecular replacement with AMoRe [54] using the structure of the T3D σ1 head (PDB ID 1KKE) as a search model. Refinement was performed with Refmac5 [55] and Phenix [56], and model building was done in Coot [57]. Ligands were fitted into weighted Fo-Fc difference density maps at a contour level of 3σ and refined using the CCP4 library and user-defined restraints. Coordinates and structure factors for all three complexes have been deposited in the PDB data bank (www.rcsb.org) with accession codes 3S6X (complex with α-2,3-sialyllactose), 3S6Y (complex with α-2,6-sialyllactose) and 3S6Z (complex with α-2,8-di-sialyllactose).
L929 cells [58] were maintained in Joklik's minimum essential medium (Sigma-Aldrich) supplemented to contain 5% fetal bovine serum, 2 mM L-glutamine, 100 U/ml of penicillin, 100 µg/ml of streptomycin, and 25 ng/ml of amphotericin B. MEL cells, previously designated T3cl.2 cells [59], were maintained in Ham's F-12 medium (CellGro) supplemented to contain 10% fetal bovine serum, 2 mM L-glutamine, 100 U/ml penicillin, 100 µg/ml streptomycin, and 25 ng/ml amphotericin B.
Recombinant reoviruses were generated by plasmid-based reverse genetics [43], [44]. Reovirus strains rsT3D (wild type), rsT3D-σ1N198D, rsT3D-σ1R202W, and rsT3D-σ1P204L were recovered using monolayers of L929 cells at approximately 90% confluence (3×106 cells) in 60-mm dishes (Costar) infected with rDIs-T7pol [60] at an MOI of ∼0.5 TCID50. At 1 h post-infection, cells were co-transfected with ten plasmid constructs representing the cloned T3D genome using 3 µl of TransIT-LT1 transfection reagent (Mirus) per µg of plasmid DNA [43]. Reovirus strains rsT3D-σ1N189A, rsT3D-σ1S195A, rsT3D-σ1R202A, rsT3D-σ1L203A, rsT3D-σ1P204A, rsT3D-σ1G205A, and rsT3D-σ1N210A were recovered using BHK-T7 cells at 90% confluence (approximately 3×106 cells) seeded in 60-mm dishes. Cells were co-transfected with five plasmids representing the cloned T3D genome using 3 µl of TransIT-LT1 transfection reagent (Mirus) per µg of plasmid DNA [44]. The amount of each plasmid used for transfection was identical to that described for L929 cell transfections. Following 3 to 5 days of incubation, recombinant viruses were isolated from transfected cells by plaque purification using monolayers of L929 cells [61]. For the generation of σ1 mutant viruses, pT7-S1T3D [43] was altered by QuikChange (Stratagene) site-directed mutagenesis. To confirm sequences of the mutant viruses, viral RNA was extracted from purified virions and subjected to Onestep RT-PCR (Qiagen) using L1- or S1-specific primers. (Primer sequences are available from the corresponding authors upon request.) The purified PCR products were subjected to sequence analysis for the presence of the introduced mutation in the S1 gene segment and the noncoding signature mutation in the L1 gene segment [43].
Purified reovirus virions were prepared using second-passage L929-cell lysate stocks of twice plaque-purified reovirus as described [20]. Viral particles were Freon-extracted from infected cell lysates, layered onto CsCl gradients, and centrifuged at 62,000 × g for 18 h. Bands corresponding to virions (1.36 g/cm3) [62] were collected and dialyzed in virion-storage buffer (150 mM NaCl, 15 mM MgCl2, 10 mM Tris-HCl pH 7.4). The concentration of reovirus virions in purified preparations was determined from an equivalence of one OD unit at 260 nm equals 2.1×1012 virions [62]. Viral titers were determined by plaque assay using L929 cells [61].
Purified reovirus virions (1011 particles) were distributed into 96-well U-bottom microtiter plates (Costar) and serially diluted twofold in 0.05 ml of PBS. Calf erythrocytes (Colorado Serum Co.) were washed twice with PBS and resuspended at a concentration of 1% (vol/vol). Erythrocytes (0.05 ml) were added to wells containing virus particles and incubated at 4°C for at least 2 h. A partial or complete shield of erythrocytes on the well bottom was interpreted as a positive HA result; a smooth, round button of erythrocytes was interpreted as a negative result. HA titer is expressed as 1011 particles divided by the number of particles/HA unit. One HA unit equals the number of particles sufficient to produce HA. HA titers from three independent experiments were compared using an unpaired Student's t test as applied in Microsoft Excel. P values of less than 0.05 were considered statistically significant.
L929 cells or MEL cells (2×105 cells/well) were plated in 24-well plates (Costar) and incubated at 37°C for at least 2 h. Cells were adsorbed with reovirus strains at an MOI of 1 PFU/cell. Following incubation at room temperature for 1 h, cells were washed three times with PBS and incubated at 37°C for 24 or 48 h. Samples were frozen and thawed twice, and viral titers were determined by plaque assay [61]. For each experiment, samples were infected in triplicate. Mean values from three independent experiments were compared using an unpaired Student's t test as applied in Microsoft Excel. P values of less than 0.05 were considered statistically significant.
|
10.1371/journal.pntd.0006157 | Discovery of novel, orally bioavailable, antileishmanial compounds using phenotypic screening | Leishmaniasis is a parasitic infection that afflicts approximately 12 million people worldwide. There are several limitations to the approved drug therapies for leishmaniasis, including moderate to severe toxicity, growing drug resistance, and the need for extended dosing. Moreover, miltefosine is currently the only orally available drug therapy for this infection. We addressed the pressing need for new therapies by pursuing a two-step phenotypic screen to discover novel, potent, and orally bioavailable antileishmanials. First, we conducted a high-throughput screen (HTS) of roughly 600,000 small molecules for growth inhibition against the promastigote form of the parasite life cycle using the nucleic acid binding dye SYBR Green I. This screen identified approximately 2,700 compounds that inhibited growth by over 65% at a single point concentration of 10 μM. We next used this 2700 compound focused library to identify compounds that were highly potent against the disease-causing intra-macrophage amastigote form and exhibited limited toxicity toward the host macrophages. This two-step screening strategy uncovered nine unique chemical scaffolds within our collection, including two previously described antileishmanials. We further profiled two of the novel compounds for in vitro absorption, distribution, metabolism, excretion, and in vivo pharmacokinetics. Both compounds proved orally bioavailable, affording plasma exposures above the half-maximal effective concentration (EC50) concentration for at least 12 hours. Both compounds were efficacious when administered orally in a murine model of cutaneous leishmaniasis. One of the two compounds exerted potent activity against trypanosomes, which are kinetoplastid parasites related to Leishmania species. Therefore, this compound could help control multiple parasitic diseases. The promising pharmacokinetic profile and significant in vivo efficacy observed from our HTS hits highlight the utility of our two-step phenotypic screening strategy and strongly suggest that medicinal chemistry optimization of these newly identified scaffolds will lead to promising candidates for an orally available anti-parasitic drug.
| Leishmaniasis, caused by the protozoa of the Leishmania species, represents a spectrum of diseases that afflicts roughly 12 million individuals worldwide. Current drug therapies for this parasitic disease are suboptimal because they are toxic, expensive, difficult to administer, and subject to drug resistance. In order to identify new and improved drug candidates, we screened a large library of small molecules for compounds that inhibit parasitic growth inside mammalian host macrophages, and have low toxicity toward the macrophages. We discovered two compounds that significantly impaired disease progression when administered orally in an animal model of cutaneous leishmaniasis. The promising pharmacokinetic and in vivo efficacy profile of the compounds make them attractive starting points for pharmaceutical development.
| Leishmaniasis constitutes a spectrum of diseases that range in severity from self-healing to fatal. The disease can present as self-healing but potentially disfiguring cutaneous leishmaniasis [1]; metastatic and highly disfiguring mucocutaneous leishmaniasis [2]; or fatal visceral leishmaniasis [3], where the parasite targets internal organs such as the liver, spleen, and bone marrow. Different species and strains of Leishmania parasites cause these distinct pathologies. The severity of the disease also depends upon host factors such as immune status [4]. An estimated 12 million individuals are infected with leishmaniasis worldwide, with a widespread geographic range that spans from India to the Mediterranean countries, to North and South America [5]. All Leishmania species have a life cycle that includes motile promastigotes that reside in the gut of the sand fly vector and non-motile amastigotes that live in the phagolysosomal vesicles of mammalian host macrophages [5].
Despite the disease’s prevalence, the current antileishmanial drug therapies are inadequate [6]. Since the 1940s, standard therapies for leishmaniasis include pentavalent antimonials, such as sodium stibogluconate (Pentostam) and meglumine antimonate (Glucantime), which are administered daily over the course of 20–30 days. Both drugs are subject to widespread resistance and are highly toxic such that treatment alone can lead to mortality [7]. The diamidine pentamidine, which has similar disadvantages, has been another drug of choice to treat cutaneous leishmaniasis for several decades. Newer drugs include amphotericin B, especially in liposomal formulation (AmBisome), the aminoglycoside paromomycin, and the phospholipid miltefosine [8, 9], which received FDA approval in 2014. However, none of these drugs is even close to optimal. They all have moderate to high toxicity, need to be administered over multiple weeks, and suffer from increasing drug resistance. Only miltefosine, a known teratogen that is unsuitable for pregnant patients, can be administered orally [10]. Leishmaniasis has been characterized as ‘a major health problem, and there is no satisfactory treatment so far’ [6]. Hence there is an urgent need for novel therapies that are safe, potent, orally bioavailable, have a low cost of goods, and are effective against drug-resistant strains of Leishmania parasites. Although a major bottleneck in progress had been the paucity of lead compounds [11] that offer the potential of becoming new antileishmanial drugs, the situation has improved recently with the application of phenotypic screening and the associated identification of multiple lead series [12].
Phenotypic screens measure the effects of a compound on intact cells rather than an isolated target (i.e., biochemical enzymatic assay) [13, 14]. Active compounds generated from whole cell-based phenotypic screens generally offer favorable cell permeability and solubility that can facilitate compound development. One limitation with this approach is that the mechanism of action of new compounds is typically unknown. Nonetheless, phenotypic screens have the complementary advantage that they can identify compounds that act therapeutically against pathways that were previously not known to be critical for parasite viability [15].
Prior phenotypic screens have predominantly used the promastigote form of the parasite, which can be readily cultured in vitro but is not the disease-causing form of the parasite. This approach has the advantage of being able to accommodate large numbers of compounds, such as the 200,000-compound library that Sharlow and colleagues screened [16]. Investigators have also used axenic amastigotes [17, 18], which are more relevant to the disease but are nevertheless a host cell-free system that only imperfectly approximates intracellular amastigotes. Most scientists agree that assays that use intramacrophage amastigotes are the most physiologically relevant assays even though they offer lower throughput. Researchers have started employing a two-stage approach involving an initial screen of promastigotes or axenic amastigotes and a secondary step to confirm the hits by screening them against intracellular amastigotes [19–21]. This approach allows the screen to be carried out with a facile high throughput approach followed by a second, more stringent, test of the primary hits for efficacy against the disease-causing intra-macrophage parasites.
The advent of high-content microscopic approaches has enabled the direct screening of compounds against amastigotes growing inside cultured mammalian macrophages [22–26]. This method can eliminate compounds that act against promastigotes while leaving amastigotes unaffected. This method is also useful for identifying compounds that target amastigotes but not promastigotes. However, this assay is technically much more complicated to undertake than assays that use promastigotes or axenic amastigotes [19]. Although one can screen large libraries with sufficient time and effort, the screens published to date have all employed smaller libraries, such as the 26,500-compound library used in Siqueira-Neto et al.’s report [26], or the focused libraries of Medicines for Malaria Box [27], and the microbial extracts collection [28].
Although many of the hits identified in the above screens have not yet been advanced to testing in animal models of leishmaniasis [29, 30], some promising leads have been identified, and various organizations are currently conducting medicinal chemistry programs. For example, the Drugs for Neglected Diseases Initiative (DNDi) is subjecting several chemotypes such as the nitroimidazoles and oxaboroles [31] to both in vitro and in vivo evaluation as orally deliverable antileishmanials in mice and hamsters. Furthermore, the Genome Institute of the Novartis Research Foundation (GNF) has identified a selective inhibitor of the kinetoplastid proteasome, GNF6702, which is active against several parasite species [32]. In addition to these advances, there is substantial benefit to providing a continued robust pipeline of lead compounds for the development of safe, potent, and orally bioavailable antileishmanials that could considerably improve the current sub-optimal armamentarium for leishmaniasis.
In this paper we report a screen of roughly 600,000 compounds for growth inhibition of L. mexicana promastigotes from several libraries, namely the St. Jude Children’s Research Hospital Chemical Biology & Therapeutics (CBT) library [33] and the Tres Cantos Antimalarial Set [34]. Two top hits from this screen, compounds 4 and 5, exhibited promising pharmacokinetic profiles that were substantially efficacious in a L. mexicana murine model of cutaneous leishmaniasis when delivered by oral gavage at a dose of 25–30 mg/kg over 10 days. Together these results suggest that compounds 4 and 5 are promising new starting points for the development of orally bioavailable antileishmanial drugs.
Animal work was approved by the Oregon Health & Science Institutional Animal Care and Use Committee under protocol #IS00002639 under adherence to the Animal Welfare Act regulations and Public Health Service Policy for the Humane Care and Use of Laboratory Animals or by the St. Jude Children’s Research Hospital Institutional Animal Care and Use Committee under protocol #477 in compliance with the Animal Welfare Act and rules articulated by the Public Health Service Policy for the Humane Care and Use of Laboratory Animals.
The current CBT library consists of roughly 600,000 unique molecules purchased from a variety of commercial sources. The library breaks into four major sets: approved drugs (~1,100 compounds); other known bioactives (~2,500 compounds); focused sets directed at defined targets, including G protein coupled receptors, kinases, proteases, and phosphatases (~45,000 compounds); and the diversity collection, which is the largest component of this library. All samples in the CBT library were carefully chosen to provide a balanced, functionally diverse collection suitable for discovery of chemical matter active against a wide variety of targets and for phenotypic screening [35, 36]. In particular, the diversity subset has been designed using a maximally diverse cluster philosophy so that the population is made up of multiple clusters, each containing a series of related compounds, where the clusters are diverse with respect to one another.
First, commercially available compounds were filtered using a combination of physiochemical metrics to improve bioavailability, and functional group metrics to reduce the probability of non-specific or artifact effects. The former is guided by the correlation of physiochemical parameters with bioactivity, as opposed to oral availability [36]. The latter is guided by implementation of the Vertex ‘Rapid Elimination of Swill’ model [37–39], which utilizes a numeric scoring method with each functional group being assigned a score from –5 (always excluded) to 0 (never excluded) and allowing an aggregate score of –2 before elimination. Next, the filtered compound list was used to generate maximally diverse clusters. In order to do this, the compounds were reduced to core fragments (or ‘scaffolds) using the method of Bemis and Murcko [40], and the compound clusters were then prioritized for purchase based on the balance of cluster diversity from the existing library as assessed by Tanimoto similarity and the presence of a reasonable number of analogs within a cluster. From 5 to 20 compounds per cluster were required, with preference for clusters of more than 20 compounds, from which a maximum of only 20 representative compounds were purchased.
All materials were purchased from commercial suppliers and used without further purification. All hits subjected to further study were repurchased and identity and purity were assessed by ultra-performance liquid chromatography (UPLC) using an H-class Waters Acquity system. Data were acquired using Masslynx v.4.1 and analyzed using the Openlynx software suite. The total flow rate was 1.0 mL/min and gradient program started at 90% A (0.1% formic acid in H2O) and was changed to 95% B (0.1% formic acid in acetonitrile) and then to 90% A. A full scan ranging from m/z 110 to 1000 in 0.2 s was used to acquire MS data. Compound identity was confirmed by low-resolution mass spectrometry and purity was assessed by ultraviolet spectroscopy and evaporative light scattering. All samples were required to exhibit > 85% purity.
Into each well of 384-well microplates (black polystyrene, clear bottom, tissue culture treated, Corning), 15 μl of medium (DME-L [41] plus 100 μM xanthine and 10% heat inactivated fetal calf serum) was dispensed with a liquid dispenser (Matrix Wellmate, Thermo Scientific). Stock compounds, dissolved in DMSO at a fixed concentration of 10 mM, were pin-transferred (V&P Scientific) into the microplate to the desired final concentration using an automated robot arm. To each well of the plates, 15 μl of L. mexicana promastigotes (strain MNYZ/BZ/62/M379, 2 x 106/mL) was added with the Wellmate dispenser. Microplates were incubated (Liconic) at 28°C and 5% CO2 for 72 h. After incubation, 10 μl of lysis/dye solution (5X SYBR Green I, 5% Triton X-100 in PBS) was added to each well. Plates were shaken at 1000 rpm, incubated at room temperature for 20 min, and fluorescence signal measured (excitation 485 nm, emission 535 nm) with the Envision plate reader (PerkinElmer).
L. mexicana (MNYZ/BZ/62/M379) or L. donovani (LdBob strain) [42] parasites expressing the Renilla luciferase gene from a rRNA gene locus were used to infect J774A.1 macrophages. Growth of intracellular amastigotes was measured using a luminescence assay, as detailed previously [43]. The growth inhibitor activities of compounds were tested against bloodstream form Lister 427 T. brucei in 96-well plates containing 1 X 105 parasites per well in 0.2 ml HMI-11 medium (Gibco/Thermo Fisher) [44]. Compounds (2 μl volumes in DMSO) were added to the parasites using serial 3-fold dilutions to cover a range of concentrations from about 10 μM to 1 nM. After 48 h incubation at 37°C under a humidified 5% CO2 atmosphere, 10 μl of 10% Triton X-100 and 100 X stock SYBR Green I (Sigma-Aldrich) in PBS was added and florescence measured (excitation 497 nm; emission 520 nm) after 1 h incubation in the dark using a Spectra Max Gemini XPS fluorimeter (Molecular Devices). Data were log transformed and EC50 values were determined using GraphPad Prism 6 (GraphPad Software). In the absence of growth inhibiting compounds, the parasites grew from an initial density of 5 X 105 cells/mL to ~3 X 106 cells/mL.
Trypanosoma cruzi CL Brener epimastigotes were obtained from Dr. Fred Buckner of the Department of Medicine at the University of Washington. T. cruzi epimastigotes were grown in liver infusion tryptose medium and seeded in a 96-well plate at 105 epimastigotes per well in 100 μl medium. For each well, 1 μl of compound in DMSO at 100X the desired final concentration was added. Epimastigotes were exposed to a range of compound concentrations from 10 μM to 1 nM to determine EC50. Plates were incubated at 26°C for 72 h, then 10 μl 50X SYBR Green in 1% Triton X-100 was added to each well followed by incubation with shaking at room temperature for 30 minutes. Fluorescence was read (excitation 485nm, emission 535nm) with the Victor2 multiplate reader (PerkinElmer). All data processing and visualization were performed using GraphPad Prism 6 software.
Methods for determination of liver microsomal stability, solubility, permeability of artificial membranes, Caco-2 cell permeability, stability in simulated gastric fluid, binding to mouse serum proteins, and in vivo pharmacokinetic studies are reported in Supporting Information.
The BJ cell line was purchased from the American Type Culture Collection (ATCC, Manassas, VA) and cultured according to recommendations. Cell culture media were purchased from ATCC. Cells were routinely tested for mycoplasma contamination using the MycoAlert Mycoplasma Detection Kit (Lonza). Cells were grown to 80% confluence, collected, and plated in 25 μL of medium per well in 384-well plates (Costar 3712). Compounds were diluted as described above and transferred to cells using a pin tool (V&P Scientific) equipped with FP1S50 pins resulting in final compound concentrations of 25 μM, and the plates incubated for 72 h at 37°C in 5% CO2. CellTiter-Glo (Promega) detection reagent was added following the manufacturer’s instructions, and luminescence was measured using an EnVision (PerkinElmer) plate reader. Data were log transformed and EC50 values were determined using GraphPad Prism 6 (GraphPad Software).
Cytotoxicity of compounds to J774.A1 macrophages was determined by dose-response curves as described previously [43]. In the absence of growth inhibitors or DMSO, the macrophages increased in number ~6-fold over 96 h in Minimum Essential Medium, employed for both macrophage infections and the toxicity assays.
Drug doses were chosen based on pilot toxicology and pharmacokinetic studies. Female BALB/C mice (compound 4) or C57BL6 (compound 5) of 17–21 grams were purchased from Charles River Laboratories (Wilmington, MA). Food and water were provided ad libitum. Two mice were used as control and another 5 mice were dosed daily via oral gavage (25 mg/kg with compound 4 and 50 mg/kg with compound 5). Every day blood was collected by retro-orbital bleed from one animal from the treatment group for pharmacokinetics. Because compound 5 induced seizures when delivered at 50 mg/kg, blood glucose was simultaneously measured with a glucose meter (Alpha track) to determine whether reduced sugar levels could be a cause of this toxicity. Each mouse received two blood collections and glucose measurements over the 10-day course of treatment.
Female BALB/c mice (~20 g) were injected in one hind footpad with 1 x 106 stationary phase promastigotes suspended in 25 μl of phosphate buffered saline (PBS). Four weeks after infection, when a small cutaneous lesion was visible in the injected footpad, cohorts of five mice were treated with either compound or vehicle alone (90 μl), delivered daily for 10 consecutive days by oral gavage using a 20-gauge x 30 mm disposable plastic feeding needle. Vehicle consisted of 10/10/40/39 mixture of ethanol/(PG)/PEG400/PBS plus 1% (weight/volume) 2HβCD (PG is propylene glycol, PEG is polyethylene glycol, 2HβCD is 2-hydroxy-β-cyclodextran). The daily dose for each compound was: compound 4, 25 mg/kg; compound 5, 30 mg/kg; miltefosine, 20 mg/kg. The width of the footpad (top to bottom) was measured with calipers before injection of parasites (day 0) and weekly from weeks 4–12. The width of the uninfected contralateral footpad was also measured each week, and its width was subtracted from that of the infected footpad to determine lesion size.
All research involving animals was carried out with the approval of the Institutional Animal Care and Use Committee of either St. Jude Children’s Research Hospital or the Oregon Health & Sciences University. The study was conducted adhering to the guidelines for animal husbandry of each institution.
A summary of the HTS workflow and quality control data is shown in Fig 1. The initial promastigote screen was performed with the St. Jude Chemical Biology & Therapeutics (CBT) library consisting of 596,414 compounds. Library compounds were filtered by several computational methods [35, 36] to remove those likely to have undesirable physical or biological properties and biased towards oral bioavailability. In this way we focused the collection on compounds most likely to be effective in cellular models of activity, without structural features that would pose a challenge to drug development [33]. In the primary screen, compounds were applied to promastigotes of L. mexicana at a fixed concentration of 10 μM, and parasite proliferation was monitored, following a 72 h incubation, by quantifying total DNA content after lysis using the nucleic acid binding dye SYBR Green I [45].
The raw data for the HTS campaign are summarized in Fig 1B as a scatterplot of normalized percent growth inhibition relative to the control drug pentamidine, which gives 100% inhibition of proliferation under these conditions. The scatterplot demonstrated ample signal separation between the positive (green) and negative (red) controls throughout the HTS campaign and a well-defined activity distribution of test compounds (blue and black). The fidelity and quality of the HTS assay were assessed using two metrics: Z-prime and EC50 of the control (pentamidine) that were calculated for each screening plate. The entire screen produced a median Z-prime value of 0.81 (interquartile range: 0.75–0.85, Fig 1C) and a consistent EC50 value of pentamidine (median 2.3 μM, interquartile range: 1.7–3.1) indicating the assay was consistent throughout the screen.
The assay’s discriminatory power was assessed using Receiver Operator Characteristic (ROC) analysis [46] as described [33]. This method helped define an optimal cutoff for designating primary hits by balancing the likelihood of a true positive with acquiring a reasonable total number of hits. Briefly, compounds were stochastically selected from the HTS screening set to sample the primary assay results according to the distribution of observed activities (ranging from 0 to 100% activity). The selected compounds were plated in a 10-point dose-response and re-evaluated in the HTS assay. True positives were defined as any compound yielding a well-behaved, saturating sigmoidal curve in the dose-response assay. The ROC curve, shown in Fig 1D, demonstrated that the assay has good discriminatory power, with an area under the curve (AUC) of 0.89 (a perfect assay would have an AUC 1.0, whereas a random assay has an AUC of 0.5). Based on this analysis, a cut-off value of > 65% inhibition was chosen, resulting in 2,703 primary hits with an expected true positive rate of 85%. It is worth noting that a significant number of true hits likely remain in the group of compounds exhibiting growth inhibition of lower the 65% cut-off activity, and these compounds were not considered in this manuscript.
To confirm the activity of the primary hits and improve confidence that they would be reasonable starting points for drug development, a variety of secondary screens and analyses were employed (Fig 1A). First, EC50 values were determined against the promastigotes using a 10-point dose-response, run in triplicate, with concentrations ranging from 0.0005–25 μM. Compounds that reproducibly exhibited EC50 activity lower than 2 μM were considered validated hits. In parallel, mammalian cell growth inhibition was determined using in vitro proliferation assays with normal human fibroblasts (BJ cells). Compounds inhibiting proliferation of BJ cells at concentrations lower than 20 μM were deprioritized. To further triage the hits, we carried out a chemical structure analysis for the 2,703 primary hit compounds utilizing topology mapping and clustering methodology [16]. We identified a wide range of chemotypes, including several scaffolds with potential structure-activity relationships (SARs), based on their dose response activity (Fig 2). Validated hits were then culled by eliminating scaffolds with less favorable drug development properties such as charged planar structures, reactive electrophilic warheads, known pan-assay interference motifs (PAINS) [47], and compounds displaying gross rule of five noncompliance [48]. Finally, we prioritized scaffolds with the possibility of facile chemical modification to generate a substantial number of structural analogs for future SAR and structure-property relationship (SPR) studies.
Among the 2703 hits, 230 compounds exhibited both an EC50 of < 2 μM for L. mexicana promastigotes and a TI > 5 (based on mammalian fibroblast toxicity). These were chosen as candidates for further study. From these 230 compounds, we were able to repurchase 113 from commercial vendors. These compounds were characterized for purity by ultra-performance liquid chromatography using ultraviolet spectroscopy and evaporative light scattering detection [49] and identity by mass spectrometry. All validated compounds were profiled for activity against intracellular amastigotes, the disease-causing stage of the life cycle. Intracellular amastigote activity was determined using a strain of L. mexicana in which the Renilla luciferase gene was integrated into the rRNA locus [43], allowing robust expression for measuring amastigote growth within cultured macrophages [50]. All 113 compounds were applied at 1 μM concentration for 96 h to J774A.1 macrophages infected with L. mexicana luciferase-expressing parasites. Of the compounds tested 55 inhibited amastigote growth by > 70%.
Next, we generated dose-response curves for these 55 compounds against intracellular amastigotes and independently against J774A.1 macrophages to establish the relative potency of each compound against the pathogen and its host cell. Those that had EC50 values < 1 μM and TI values > 10 for macrophages were selected from the 55 compounds, as suggested for lead identification for leishmaniasis [12], and several compounds were then removed due to known biological liabilities of scaffolds (manual curation, Fig 1). The nine remaining compounds, each representing a unique chemical scaffold (compounds 1–9, Fig 3), were designated top hits (Fig 1).
Notably, this screening strategy successfully identified several known antileishmanial scaffolds, including compounds 1, 2, 3, and 4, thus providing further validation of the screen. The alkaloid cephaeline (1), a known irritant of gastric mucosa and component of ipecac, has been shown to be potent against L. mexicana and L. donovani intracellular amastigotes [51]. Another known scaffold, the quinazoline-2,4-diaminoquinazolines, represented by compound 2, has been studied extensively and shown to have activity against L. donovani, and L. amazonensis [52, 53]. Compound 2 is also present in the malaria box of compounds active against Plasmodium falciparum and has been shown to have activity against L. infantum [54]. We also found a member of the 2,4-diaminopyrimidine scaffold, compound 3, some of which are selective against L. major amastigotes, with EC50 values in the low μM range and in once case with a therapeutic index (TI) as high as 130 [55]. The compounds in that study share the 2,4-diaminopyrimidine scaffold with compound 3, but they differ in having a benzyl substitution at the 5 position of the pyrimidine ring rather than modifications on the 2- and 4-amino substituents that are present in compound 3. Finally, various 4H-chromen-4-ones, similar to compound 4, are active against L. major [56]. Of the validated scaffolds included in the HTS campaign, the three that exhibited the widest SAR range (7–88 fold) were the 2,4-diaminoquinazolines, 2,4-diaminopyrimidines, and 4H-chromen-4-ones (Fig 2).
Potency of the nine compounds against intracellular amastigotes of L. donovani was also quantified to assess each compound’s potential to control this agent of fatal visceral leishmaniasis (Fig 3). We have recently reported that compound 5 is also potent against another kinetoplastid parasite, the bloodstream form of Trypanosoma brucei (EC50 value of 0.027 μM) [58] and active in vivo in a murine model of African trypanosomiasis (manuscript in preparation). Thus, we also evaluated the other compounds for activity against the related pathogen Trypanosoma brucei. As noted for the broad spectrum kinetoplastid proteasome inhibitor GNF6702 [32], compounds exhibiting activity against multiple parasites are especially interesting, as such scaffolds can be explored for therapies against multiple neglected parasitic diseases.
All nine compounds were potent (EC50 < 0.6 μM) against both the L. mexicana and L. donovani intracellular amastigotes. Often, potency correlated well between the two species, although there were significant differences for some compounds (e.g., compounds 2, 4, 8, and 9). While none of the compounds affected the proliferation of BJ cells at concentrations as high as 20 μM, most of the compounds reduced viability of macrophages with half-maximal lethal dose (LD50) values around 1–10 μM. Only compounds 2 and 4 demonstrated no reduction in viability in dose-response studies against the host macrophage J774A.1, suggesting these compounds may afford the best selectivity for inhibiting parasite growth relative to toxicity toward the host macrophage or other mammalian cells. Thus, all of the nine compounds tested afforded favorable therapeutic indices (> 50), except compound 3. Notably, compounds 5, 8, and 9 exhibited good potency (< 0.3 μM) against bloodstream form T. brucei. To determine whether any of the top hits might also be effective against the related kinetoplastid parasite T. cruzi, we performed dose-response curves with compounds 4, 5, 8, and 9 against epimastigotes and found either no inhibition (4, n = 3) or EC50 values of 0.086 ± 0.03 μM (5, n = 4), 0.33 μM (8, n = 1), and 2.1 ± 0.07 μM (9, n = 2), respectively. Hence, each of these latter scaffolds is of potentially high interest for development of drugs against multiple species of kinetoplastid parasites.
Together, these data suggest the seven compounds not previously reported to possess antileishmanial activity (only 1 and 2 have been documented previously) can be good starting points for discovering new antileishmanials. Herein, we chose to further profile compounds 4 and 5, representing the 4H-chromen-4-ones and p-chloronitrobenzamides scaffolds, respectively. Compound 4 was chosen for its distinct lack of toxicity against host macrophages and compound 5 was chosen for its cross-species potency. We suggest that similar studies could be undertaken using the other validated compounds from our two-stage phenotypic screening campaign.
In order to evaluate compounds 4 and 5 for in vivo studies, we measured the in vitro ADME physiochemical properties likely to be predictive of oral bioavailability (Table 1). First, we looked at solubility in an aqueous buffer (pH = 7.4) and ability to cross an artificial (parallel artificial membrane permeability, PAMPA) or cellular (Caco-2) membrane. Compound 4 exhibited good solubility (67 μM) and moderate membrane permeability (Table 1) suggesting a high predicted absorption across the intestinal epithelium (~85%), and low probability of being a substrate of the drug resistance pumps expressed by Caco-2 cells (efflux ratio < 2). Compound 5 showed moderate permeability in both the PAMPA and Caco-2 assays as well as an acceptable efflux ratio of 1.92 (Table 1). Compound 5 exhibits low aqueous solubility (0.3 μM) but we anticipated that this could be compensated by formulation for delivery [59]. Next, we investigated the stability of both compounds in simulated gastric fluid and in microsomal models of oxidative metabolism. Both compounds exhibited high stability in simulated gastric fluid (t1/2 > 24 h) and demonstrated good metabolic stability (t1/2 > 4 h for all species) in liver microsome preparations from mouse, rat, and human. Compounds 4 and 5 also showed modest (<50%) binding to mouse plasma proteins, below the level of the positive control drug propranolol (Table 2). Criteria that have been suggested as promising for an orally bioavailable compound include: aqueous solubility > 1 μM but ideally > 100 μM [60], PAMPA permeability coefficient of > 1x10-5 cm/sec represents high permeability, Caco-2 cell permeability coefficient of > 1x10-6 cm/sec [60], Caco-2 cell efflux ratio < 2 represents no efflux, gastric stability > 24 h, t1/2 in microsomes > 30 min [12]. However, these values only represent broad guidelines, and many efficacious drugs violate them. Overall, the in vitro ADME data suggest that the scaffolds of compounds 4 and 5 would be appropriate for development into orally bioavailable antileishmanial compounds.
To further evaluate the potential of 4 and 5 in vivo, we performed preliminary single oral dose pharmacokinetic studies in mice. Following a single oral gavage (PO) of 4 in mice at 25 mg/kg (Fig 4) the plasma concentration remained above its EC50 of 0.08 μM for approximately 20 h. Compound 4 reached a peak plasma concentration (Cmax) of 3.2 μM within 1 h (tmax) of dosing, afforded an AUC of 16.7 μM.h, and an elimination half-life (t1/2) of 3 h (Table 3). Following PO dosing of 5 at 50 mg/kg (Fig 4, Table 3), the plasma concentration remained above the EC50 of 0.022 μM for roughly 48 h, with a Cmax of 6.49 μM, a tmax of 4 h, an AUC of 83.2 μM*h, and a t1/2 of 7.1 h. Thus, both compounds exhibited good plasma exposure and sustained plasma concentrations above an efficacious dose (EC50) for more than 12 h following a single oral gavage dosing using our standard formulation (10/10/40/39, EtOH/PG/PEG/PBS (7.4) (v/v) and 1% (w/v) HβCD). These results strongly suggested that both compounds were appropriate candidates for efficacy evaluation in the murine model of cutaneous leishmaniosis.
Next, we sought to determine the allowable dosing range for our efficacy model by carrying out dose-ranging tolerability studies. When compound 4 was dosed by oral gavage at 50 mg/kg, half of the animals exhibited seizure-like behavior. Blood chemistries revealed a very low glucose level in plasma (23–40 mg/dl for treated mice compared to 185–251 mg/dl for untreated mice). This observation might suggest blockage of a kidney and/or an intestinal glucose transporter. When we repeated the same experiments at 25 mg/kg, no seizures were seen and the glucose level of each animal remained within normal limits at both Cmax and Cmin. Daily oral administrations of 4 at 25 mg/kg were well-tolerated in all study animals, no significant changes in either clinical chemistry or complete blood counts were observed, and there were no other test article-related effects noted in the liver or any other tissues. For compound 5 dosed at 50 mg/kg in mice, a 10-day toxicity study revealed that animals reduced food intake and lost more than 10% of their weight overtime. The observed suppressed appetite was resolved by dose reduction to 30 mg/kg. Thus, we employed 25 mg/kg of 4 and 30 mg/kg of 5 in the efficacy model. No weight loss or other toxicity was observed at these doses.
Next we assessed the potential of compounds 4 and 5 to control disease in a murine model of cutaneous leishmaniasis [30]. We infected BALB/c mice with L. mexicana via footpad injections on day zero, allowed incipient lesions to develop for four weeks, and then treated cohorts of five animals with each compound for 10 consecutive days by oral gavage. In addition, five mice were treated with 20 mg/kg of the only orally available approved antileishmanial drug, miltefosine, as a positive control and with vehicle alone as a negative control. Footpad widths were measured from 4–12 weeks post-infection (Fig 5). For vehicle-treated mice, the lesions grew steadily up to 2 mm width, at which time mice were euthanized. Miltefosine reduced lesion size from the initial dimension and was able to maintain growth inhibition for eight weeks following treatment. Both compounds 4 and 5 controlled lesion size at dimensions similar to that at the time of compound dosing (4 weeks) until week 9, well after stopping oral administration. After week 9, the footpad lesions began to increase in size. Hence, oral dosing of both compound 4 and 5 controlled disease progression during the dosing period and for a significant period of time after dosing stopped but neither was as efficacious as miltefosine. The partial control of virulence exhibited by our HTS hits, without any optimization, strongly suggests both compounds are novel early leads for the development of orally available antileishmanials. Given the synthetic tractability of these scaffolds [58, 61], we envision a rapid timeline for the development of optimized leads with enhanced therapeutic properties.
High throughput phenotypic screening offers a powerful tool to discover therapeutically relevant leads for drug discovery [62]. The HTS campaign described in this paper represented part of a larger effort to identify selective inhibitors of hexose transporters from various parasitic protozoa [45], including a transgenic strain of L. mexicana, Δlmxgt1-3[pLmxGT2] [63]. The results reported here began as a second, adventitious outcome of that screen, where we identified 2,703 compounds that significantly inhibited the growth of promastigotes of the Δlmxgt1-3 parasites employed as the cellular expression vehicle for the hexose transporters. In this study, we leveraged this secondary outcome to identify novel orally available antileishmanials. We emphasize that since most of the top hits against L. mexicana are also potent against an agent of lethal visceral leishmaniasis, L. donovani (Fig 3), this screen is of potential therapeutic value for both cutaneous and visceral leishmaniasis.
The initial screen against the L. mexicana promastigote form of the parasite was highly robust with a median Z value of 0.81 and an AUC of 0.893 for the ROC curve. In addition to the antileishmanial compounds that were carried through the secondary validation assays (vide supra), the HTS identified a variety of inhibitors known to be active against various Leishmania strains: crystal violet (EC50: 0.29 μM) [64], disulfiram (EC50: 0.50 μM) [65], thiram (EC50: 1.77 μM) [65], actinomycin D (EC50: 0.36 μM) [28], anisomycin (EC50: 0.58 μM) [66], and avicin (EC50: 1.50 μM) [16]. The rediscovery of these known inhibitors provided another level of validation and confirmed the screen’s ability to identify active antileishmanial compounds.
Our motive behind the sequential screening of promastigotes followed by amastigotes was to eliminate promastigote-specific hits. In the process we identified multiple hits that were potent inhibitors of both promastigote and amastigote growth and removed compounds that inhibited growth of either host macrophages (J774A.1) or normal fibroblasts (BJ cells) (Fig 3). Hence, while there has been much discussion about the relative merits of screens employing promastigotes, axenic amastigotes, and intracellular amastigotes (see Introduction), the sequential approach employed here sidesteps that debate and identified multiple scaffolds with potential for further development toward orally bioavailable antileishmanial drugs.
Three scaffolds from the sequential screen stood out as promising candidates due to the wide range of SAR inherent in our screening data set (Fig 2), the high potency of certain exemplars, and good TI values (Fig 3): 2,4-diaminoquinazolines (2), 2,4-diaminopyrimidines (3), and 4H-chromen-4-ones (4). The 2,4-diaminoquinazolines have been disclosed previously as potential antileishmanials [52, 53, 67]. This scaffold has been explored by medicinal chemistry, and one compound was identified that exhibited an EC50 of 0.15 μM against L. donovani amastigotes and a TI of 100 [52, 53, 67]. These studies also demonstrated that Leishmania dihydrofolate reductase (DHFR) is inhibited by 2,4-diaminoquinazolines, highlighting this essential enzyme as one target for this class of antileishmanials.
Similarly, 2,4-diaminopyrimidines have been shown to have μM potency against Leishmania amastigotes [55]. Compound 3 is more potent than the previously studied 2,4-diaminopyrimidines [55], and it has substitutions on the 2,4-amino groups, unlike previously characterized 2,4-diaminopyrimidines. These results suggest that substitution at these positions may be important for potency and imply that additional modifications at these sites may be worth exploring. Furthermore, 2,4-diaminopyrimidines are structurally related to classical DHFR inhibitors such as pyrimethamine and trimethoprim [55] that also selectively inhibit the essential Leishmania DHFR, providing a potential molecular target for this family of antileishmanials. However, two distinct enzymes in L. major, DHFR and pteridine reductase 1 (PTR1), can reduce folate, and amplification of the PTR1 gene can confer methotrexate resistance upon the parasite by metabolically circumventing inhibition of DHFR by this antifolate [68]. Hence, effective inhibitors of DHFR may also need to inhibit PTR1, thus complicating chemotherapy against this target.
4H-chromen-4-ones [56], and related chroman-4-ones [69], have been demonstrated to have activity against both T. brucei and L. major, and they bind to and inhibit the critical [70] enzyme PTR1 that is present in kinetoplastid parasites, but not in mammals. These results with structurally related compounds suggest that PTR1 may be a principle target of compound 4. However, the compounds tested in this previous work had an aromatic substituent at the 2 position of the chromen-4-one ring rather than at the 3 position. Those compounds exhibited much lower potency against T. brucei and L. major, with EC50 values in the micromolar range, compared to compound 4 against L. mexicana or L. donovani amastigotes (Fig 3). Furthermore, compound 4 has lower toxicity, a higher TI (Fig 3, no inhibition of J774A.1 macrophages up to 10 μM concentration), and greatly superior pharmacokinetic properties (Fig 4, Table 3) compared to compounds tested by Borsari et al. [56], where the top hit exhibited a half-life of 7.6 min in mice. These observations suggest that structural features present in compound 4 may provide a route for developing this scaffold toward more optimal lead compounds against Leishmania parasites. Variants of the isoflavone scaffold present in 4 have been employed as dietary supplements and are known to have phytoestrogen and antioxidant properties [71]. Thus, there has been a long-standing interest in the development of synthetic routes to access these desirable properties [72, 73]. More recently, rapid synthetic routes to access highly substituted hydroxylated isoflavones such as 4 have been published [61].
Compound 5 has been identified previously [58] by our laboratory as active against all Trypanosoma species in vitro and efficacious against T. congolense and T. b. rhodesiense in vivo (manuscript in preparation). This scaffold is especially interesting, since it may act on a common cellular target found among kinetoplastids and could be developed as both antileishmanial and antitrypanosomal drugs. In addition, the in-house experience with the scaffold in in vivo models inspired confidence regarding its oral activity. Therefore, we chose to progress compounds 4 and 5 for further in vitro ADME and in vivo pharmacokinetic and pharmacodynamic testing. We note that compounds 8 and 9 also exhibit significant potency toward T. brucei and may therefore be of special interest for future investigations.
Compounds 4 and 5 were both able to partially control the size of an incipient cutaneous lesion when delivered orally at 25–30 mg/kg for 10 days, compared to mice that received vehicle alone. However, they were not as efficacious as the currently employed oral drug miltefosine. This efficacy in vivo indicates that improvements will be required to further address the potential of these scaffolds for drug development. In particular, analogs that exhibit higher potency and/or lower toxicity in animals may achieve greater efficacy or allow higher dosing. The ability to extensively modify both scaffolds by medicinal chemistry offers the potential to generate libraries of analogs of each lead whose members can then be tested for improved ADME, PK, toxicity, and in vivo efficacy in animal models of both cutaneous and visceral leishmaniasis. Additionally, the combination of potency, selectivity, and identification of DHFR as a potential molecular target all suggest that further exploration of the 2,4-diaminoquinazoline and 2,4-diaminopyrimidine scaffolds, represented by compounds 2 and 3 respectively, may be warranted. Overall, this work demonstrates how sequential screening of promastigotes, which are especially amenable to HTS assay development, followed by hit validation in the disease causing intramacrophage amastigotes can be used to successfully identify novel antileishmanial scaffolds. The promising pharmacokinetic profile and significant in vivo efficacy of our newly identified scaffolds strongly suggests that additional medicinal chemistry optimization may yield orally available anti-parasitic drugs.
|
10.1371/journal.ppat.0030051 | Bacterial Ligands Generated in a Phagosome Are Targets of the Cytosolic Innate Immune System | Macrophages are permissive hosts to intracellular pathogens, but upon activation become microbiocidal effectors of innate and cell-mediated immunity. How the fate of internalized microorganisms is monitored by macrophages, and how that information is integrated to stimulate specific immune responses is not understood. Activation of macrophages with interferon (IFN)–γ leads to rapid killing and degradation of Listeria monocytogenes in a phagosome, thus preventing escape of bacteria to the cytosol. Here, we show that activated macrophages induce a specific gene expression program to L. monocytogenes degraded in the phago-lysosome. In addition to activation of Toll-like receptor (TLR) signaling pathways, degraded bacteria also activated a TLR-independent transcriptional response that was similar to the response induced by cytosolic L. monocytogenes. More specifically, degraded bacteria induced a TLR-independent IFN-β response that was previously shown to be specific to cytosolic bacteria and not to intact bacteria localized to the phagosome. This response required the generation of bacterial ligands in the phago-lysosome and was largely dependent on nucleotide-binding oligomerization domain 2 (NOD2), a cytosolic receptor known to respond to bacterial peptidoglycan fragments. The NOD2-dependent response to degraded bacteria required the phagosomal membrane potential and the activity of lysosomal proteases. The NOD2-dependent IFN-β production resulted from synergism with other cytosolic microbial sensors. This study supports the hypothesis that in activated macrophages, cytosolic innate immune receptors are activated by bacterial ligands generated in the phagosome and transported to the cytosol.
| Innate immune recognition of microorganisms has a direct impact on the type and the magnitude of the immune response elicited. While recognition of microorganisms relies on receptors that sense pathogen-associated molecular patterns, (PAMPs), it was reasonable to suspect that immune cells could discriminate between live and dead bacteria. Listeria monocytogenes is an intracellular pathogenic bacterium used extensively as a model system for studying basic aspects of innate and acquired immunity. L. monocytogenes is internalized by macrophages, escapes from a vacuole, multiplies within the cytosol, and spreads from cell to cell without lysing the cells. We used wild-type and bacterial mutants of L. monocytogenes to demonstrate that macrophages not only respond differently to bacteria that are growing in the cytosol and to non-growing bacteria that are trapped in a vacuole, but that they also can discriminate between intact or degraded bacteria in the vacuole. We showed that macrophages induce specific immune response when bacteria are killed and degraded. This response was directly correlated to the ability of macrophages to degrade bacteria and involved receptors that were located in the host cell cytosol. These observations led us to suggest that bacterial degradation products may serve as messengers that inform immune cells that bacteria were killed and degraded. This information might affect directly the immune response, for example, by down-regulating inflammatory responses that can be deleterious. We call these bacterial degradation products PAMP-PM (PAMP–post-mortem).
| Macrophages are highly phagocytic cells that can act as benign scavengers, sentinels of microbial infection, and hosts to intracellular pathogens [1]. However, a key property of macrophages is their capacity to be immunologically activated by cytokines such as interferon (IFN)–γ. Subsequent to phagocytosis of microorganisms, activation is manifested as an enhanced microbiocidal, degradative, and secretory capacity concomitant with maturation of phagosomes into acidic hydrolytic compartments [2]. How macrophages couple microbiocidal and degradative activity with the development of an appropriate immune response is critical to understanding the regulation of inflammation.
Recognition of microorganisms by the innate immune system is mediated by invariable pattern recognition receptors (PRRs) that bind conserved molecules present on microorganisms, referred to as pathogen-associated molecular patterns (PAMPs). Among PRRs are the Toll-like receptors (TLRs), type I integral membrane proteins located at the cytoplasmic membrane and internal membrane-bound compartments, and nucleotide-binding oligomerization domain (NOD) proteins located in the cell cytosol [3,4]. Microbial structures exposed on the bacterial cell surface, such as lipopolysacaccharide (LPS), peptidoglycan (PGN), and flagellin, are recognized by TLR4, 2, and 5, respectively, which are localized to the host cell surface. In contrast, microbial nucleic acids are recognized by TLR3, 7, and 9, which are located within intracellular membrane-bound compartments that can fuse with phagosomes during their maturation. Treatments of cells with agents that block vacuolar acidification abrogate responses mediated by TLR3, 7, and 9 [5,6]. Whereas TLRs detect microorganisms extracellularly or within the luminal side of the phagosome, the NOD-like receptor family may comprise a surveillance system that recognizes intracellular pathogens, leading to both transcriptional responses and activation of the inflammasome [7]. Among the cytosolic innate immune receptors, RIG-I and MDA5 recognize double-stranded RNA, whereas NOD1 and NOD2 recognize bacterial PGN degradation products [7–12].
Engagement of innate immune receptors with specific microbial ligands results in signaling pathways that culminate in host transcriptional responses associated with inflammation. Signaling pathways are characterized by their shared adaptor molecules. For example, MyD88 is a major adaptor that mediates immune responses downstream of all TLRs except TLR3, leading primarily to activation of nuclear factor (NF)–κB [13]. Interaction of MyD88 with TLR7 and 9 can also lead to the induction of type I IFN response through activation of interferon regulatory factor (IRF) 3/7. TLR3 and TLR4 can both induce type I IFN responses via another adaptor molecule, called Trif (Lps2), again through the activation of IRF3 [13,14]. The cytosolic receptors that recognize RNA and DNA signal by interacting with the mitochondrial membrane adaptor MAVS/VISA/IPS-1/Cardif [15], leading to activation of IRF3 and production of IFN-β. Less is known about the signaling pathways downstream of the NOD proteins, although NOD1 and NOD2 interact with the adaptor molecule Rip2 (RICK) to activate NF-κB [16].
Type I IFNs have been studied extensively with regard to their role as anti-viral cytokines, but their role in response to bacterial infection has been less studied, although bacterial LPS derived from Gram-negative bacteria is clearly an inducer of type I IFN [17]. Recently, it was shown that a Gram-positive pathogen, Listeria monocytogenes, induces type I IFN, but only upon entry into the host cell cytosol [18–21]. Mutants lacking the secreted pore-forming protein listeriolysin O (LLO) fail to escape from a phagosome and fail to induce IFN-β. Recognition of LLO-minus L. monocytogenes in the phagosome is largely MyD88-dependent, while recognition of cytosolic bacteria is MyD88-independent and IRF3-dependent [18–21]. The nature of the bacterial ligand(s) and the host PRRs responsible for the activation of IRF3 in response to cytosolic L. monocytogenes are not known, although bacterial DNA can recapitulate this response [22]. The production of type I IFN in response to L. monocytogenes is enigmatic, as mice lacking the IFNα/β receptor are more resistant to listeriosis [23–25].
A role of type I IFNs in the induction of acquired immunity has become increasingly recognized, and it has been suggested that a key feature of effective adjuvants is the capacity to induce type I IFN [26]. However, NOD2, the target of one of the most powerful adjuvants (muramyl dipeptide [MDP] derived from bacterial PGN), has not been associated with the expression of type I IFNs. In this report, we show that activated macrophages express IFN-β after phagocytosis and degradation of L. monocytogenes, and that NOD2 is necessary for full expression. IFN-β induction by LLO-minus bacteria was dependent on the activity of the macrophage vacuolar ATPase, not for acidification of the phagosome but for the generation of the phagosomal membrane potential which, we hypothesize, has a role in the active transport of bacterial ligands into the cytosol.
It was previously demonstrated that macrophages respond differently to bacteria located in their cytosol (e.g., wild-type [w.t.] L. monocytogenes), compared to bacteria trapped in a phagosome (e.g., LLO-minus L. monocytogenes) [18,19]. Cytosolic bacteria trigger an MyD88-independent production of IFN-β, whereas phagosomal bacteria do not [20,21]. However, these observations were based on the response of macrophages that were not activated and consequently weakly bacteriocidal. In vivo, during L. monocytogenes infection, cytokines such as IFN-γ act on macrophages to render them highly bacteriocidal and therefore less permissive for L. monocytogenes replication [27]. We reasoned that the bacteriocidal activity of macrophages, such as killing and degradation of bacteria, would directly affect the innate immune response to L. monocytogenes infection. In order to test this hypothesis, we studied the response of IFN-γ–activated macrophages to infection with w.t. L. monocytogenes and an LLO-minus mutant. The bacteriocidal activity of macrophages was best demonstrated when peritoneal macrophages were infected with L. monocytogenes (Figure 1A). Growth curves of L. monocytogenes in activated peritoneal macrophages showed dramatic killing of w.t. bacteria. The number of bacteria recovered from the activated peritoneal macrophages decreased during 6 h of infection, while in non-activated peritoneal macrophages, bacteria were subjected to initial killing, but survivors continued to grow (Figure 1A). Infection of peritoneal macrophages with the LLO-minus mutant resulted in killing of bacteria even without IFN-γ treatment (Figure 1A). Unlike in peritoneal macrophages, the bacteriocidal activity of bone marrow–derived (BMD) macrophages was less profound and was completely dependent on IFN-γ treatment. Growth curves of w.t. L. monocytogenes in IFN-γ–activated BMD macrophages showed that IFN-γ treatment initially restricted the growth of w.t. L. monocytogenes, although bacteria were still able to escape to the cytosol and replicate (Figure 1B). The bactericidal activity of BMD macrophages was best observed when the cells were infected with the LLO-minus mutant. Like in peritoneal macrophages, a one-log decrease in the number of LLO-minus bacteria was recovered after 6 h of infection (Figure 1B). Since BMD macrophages killed phagosomal-trapped bacteria only upon IFN-γ activation, and w.t. L. monocytogenes were still able to escape to the cytosol in activated BMD macrophages, we chose to use these cells to study further the effect of the bacteriocidal activity on the innate immune response to L. monocytogenes. To examine whether L. monocytogenes were subjected to lysis in phagosomes, activated BMD macrophages were infected with w.t. L. monocytogenes or an LLO-minus mutant expressing cytosolic–green fluorescent protein (GFP). Immunofluorescence microscopy revealed that at 6 hours post-infection (h.p.i.), most w.t. bacteria were cytosolic, as many of them were engaged with actin tails, showed by co-localization of the GFP bacteria with the actin marker rhodamine-phalloidin. Only a few w.t. bacteria were labeled just with GFP, suggesting that they were trapped in the phagosome (Figure 1C). Infection with a GFP-expressing LLO-minus mutant revealed that non-activated macrophages contained one to two intact GFP-expressing bacteria per cell, whereas in activated macrophages the GFP was released from the bacteria and distributed in multiple vacuoles around the cell (Figure 1C). While we don't know the precise composition of these vacuoles, co-localization of some of the GFP-labeled vacuoles with the pH-sensitive dye LysoTracker RED suggested that they might have originated from the primary phagosome during its maturation to a phago-lysosome (Figure 1C). As a control, w.t. L. monocytogenes did not localize with LysoTracker RED labeling (Figure 1C). These results indicated that macrophage activation led to an enhanced degradative activity and trafficking of host-generated bacterial ligands, which potentially can be sensed by the innate immune system.
In order to study the innate immune response of activated macrophages to L. monocytogenes infection, we used Mouse Exonic Evidence Based Oligonucleotide (MEEBO) microarrays [28]. Analysis of the gene expression profiles of IFN-γ–activated and non-activated BMD macrophages infected with w.t. L. monocytogenes and an LLO-minus mutant were performed. Consistent with previous studies, the response of non-activated macrophages to infection with w.t. L. monocytogenes and the LLO-minus mutant clustered into three groups of genes: i) genes that are largely induced by w.t. cytosolic bacteria (i.e. cytosolic-induced genes), ii) genes that are largely induced by phagosomal LLO-minus bacteria (i.e., phagosomal-induced genes), and iii) genes that are induced by both cytosolic and phagosomal bacteria (Figure 1D; 253 genes vary > 5.6-fold). Analysis of these genes in IFN-γ–activated macrophages infected with w.t. or LLO-minus bacteria reveled that many of the genes that were cytosolic or phagosomal-specific in non-activated macrophages were induced by both bacteria in IFN-γ–activated macrophages (Figure 1D). As expected, w.t. L. monocytogenes triggered expression of genes from the “phagosomal-specific genes” category in the activated macrophages. Under these conditions, some w.t. bacteria failed to escape to the cytosol and remained trapped in the phagosome; consequently, they induced phagosomal-specific genes as well (Figure 1B and 1C). Pro-inflammatory cytokines such as interleukin (IL)–12 and IL-1α, which are normally induced by LLO-minus bacteria, were highly induced by w.t. cytosolic bacteria in activated macrophages compared to non-activated macrophages (Figure 1E). Conversely, LLO-minus bacteria induced many “cytosolic-specific genes” upon activation of macrophages, including the most highly induced gene in the macrophages' response to w.t. L. monocytogenes, IFN-β (Figure 1E). Like IFN-β, other cytosolic-specific genes that are normally induced by w.t. bacteria, such as IL-15, chemokine CXC ligand 11, chemokine C-C receptor like 2, and type I IFN–related genes, were also induced by LLO-minus bacteria in activated macrophages (Figure 1E). Interestingly, IFN-β was among the 20 most induced genes in activated macrophages infected with LLO-minus mutant (out of 12,344 genes total), together with Nos2, ubiquitin D, and C-C receptor like 2 (Table S1). Since the IFN-β response to w.t. L. monocytogenes is well established, we were interested in whether the IFN-β response to LLO-minus bacteria shares common signaling pathways. Further validation of IFN-β induction by the LLO-minus mutant in activated macrophages was performed using quantative real-time PCR (Q-RT-PCR) analysis. A time course analysis of ifnβ expression during w.t. L. monocytogenes infection demonstrated that ifnβ induction increased 10-fold in activated macrophages compared to non-activated macrophages (Figure 2A). Upon infection with an LLO-minus mutant, activated macrophages induced ifnβ to the same level as in response to w.t. L. monocytogenes (Figure 2A). This was also the case when activated peritoneal macrophages were infected with the LLO-minus mutant (Figure S1). These results demonstrated that bacteria trapped in the degradative phago-lysosomes of activated macrophages trigger the induction of IFN-β, a response seen in non-activated macrophages only by bacteria able to access the cytosol. An obvious consequence of bacterial degradation is the release of bacterial ligands, such as nucleic acids, that are not normally exposed by either live bacteria or killed, but non-degraded, bacteria. It is possible that these bacterial degradation products triggered the induction of IFN-β by activated macrophages infected with LLO-minus mutant.
Since several TLRs are known to induce IFN-β through the adaptor molecules MyD88 and Trif, we used macrophages isolated from mice lacking each one of these adaptors to examine their role in the IFN-β response to degraded bacteria. Whereas ifnβ induction was slightly reduced in Trif-deficient macrophages, it was completely abolished in MyD88-deficient macrophages (Figure 2B). However, examination of IFN-γ–activated myd88−/− macrophages infected with LLO-minus GFP-expressing mutants revealed a defect in bacterial degradation. While in C57BL/6 macrophages, bacteria were lysed as shown by GFP distribution, in MyD88-deficient macrophages the LLO-minus mutant remained intact even 6 h.p.i (Figure 2C). Growth curves of the LLO-minus mutant in MyD88-deficient activated macrophages revealed a slight decrease in bacterial colony-forming units (CFUs) when compared to non-activated macrophages, suggesting that MyD88-deficient macrophages have a defect in killing of the LLO-minus mutant (Figure S1). This result, although striking, made it impossible to decipher the precise role played by MyD88 in the ifnβ induction by LLO-minus mutants (see discussion below).
Since IFN-β induction occurs downstream of TLR3, 7, and 9, which are involved in nucleic acid recognition, we examined the possibility that bacterial nucleic acids, possibly released upon bacterial degradation, trigger IFN-β production in activated macrophages. Macrophages isolated from mice lacking individual TLR3, 7, and 9 were infected with the LLO-minus mutant to test their involvement in bacterial nucleic acid recognition. Somewhat surprisingly, none of these TLRs had any detectable affect on the induction of ifnβ by LLO-minus L. monocytogenes (Figure 2D).
Since none of the TLRs seemed to be playing a role in the detection of LLO-minus L. monocytogenes, we considered other host receptors that recognize bacterial ligands. NOD1 and NOD2 are cytosolic proteins that are activated by muropeptides derived from bacterial PGN [29]. Surprisingly, NOD2 was involved in production of IFN-β in response to the LLO-minus mutant, as macrophages from NOD2-minus mice expressed and secreted 50% of IFN-β (Figure 3A). The induction of IFN-β by w.t L. monocytogenes or an LLO-minus mutant was independent of NOD1 (Figure 3A). Although nod2 gene expression was induced by both w.t L. monocytogenes and LLO-minus bacteria (Figure 1D), it affected only the IFN-β response to phagosomal bacteria (LLO-minus) and not to cytosolic w.t. bacteria (Figure 3A). Unlike NOD2, the transcriptional regulator IRF3 was required for IFN-β production by both vacuolar and cytosolic bacteria, suggesting that both pathways share common adaptors downstream of the signaling pathways leading to type I IFN response (Figure 3B).
This is the first report to our knowledge that links NOD2 activation with type I IFN responses. MDP, the well-studied ligand of NOD2, does not induce IFN-β when delivered to the cytosol (Figure 3C). However, as NOD2 synergizes with other receptors for cytokine production [30–34], we tested the possibility that the IFN-β production in response to degraded bacteria was a result of synergism between NOD2 and other innate immune receptors. MDP was delivered with poly(I:C) (a dsRNA analog that is sensed by phagosomal and cytosolic receptors) to the activated macrophage cytosol. Whereas poly(I:C) alone led to production of IFN-β, when combined with MDP, the amount of IFN-β secreted by macrophages was 40% higher than with poly(I:C) alone (Figure 3C). This increase in IFN-β production was dependent on NOD2, demonstrating that NOD2 can synergize with other receptors leading to an enhanced type I IFN response.
Since the IFN-β response to phagosomal-degraded bacteria required the cytosolic receptor NOD2, we hypothesized that bacterial ligands were generated and transported from the phagosome and detected in the cytosol. In order to test whether degradation of bacteria is a prerequisite for IFN-β induction by LLO-minus mutants, we treated activated macrophages with bafilomycin A, a specific inhibitor of the vacuolar ATPase proton pump (V-ATPase). Bafilomycin A inhibits phagosome acidification and blocks maturation of phagosomes to phago-lysosomes. Bafilomycin A–treated macrophages indeed failed to degrade internalized LLO-minus mutants (Figure 4A), and did not induce ifnβ (Figure 4B), whereas this treatment had no effect on induction of ifnβ by cytosolic L. monocytogenes (Figure 4B). To distinguish between the requirement of degradation of bacteria or of acidification of the phagosome for IFN-β signaling, we used alternative endosomal acidification inhibitors, monensin or nigericin, which act differently than bafilomycin A. Monensin and nigericin are electro-neutral monovalent cation exchangers that are widely used to exchange K+/H+ ions across biological membranes [35]. In the presence of active V-ATPase, treatment with monensin or nigericin will induce intra-phagosomal accumulation of K+ ions as a result of exchange with luminal H+ (Figure 4C). This will lead to neutralization of vacuolar pH without changing the vacuolar membrane potential [35]. Neither addition of nigericin nor monensin blocked the induction of ifnβ; in fact, both resulted in the enhanced induction of ifnβ (Figure 4C), while neutralizing the phagosomal pH as determined by Lyso-Tracker RED staining (unpublished data). Combining bafilomycin A with monensin (or nigericin) or bafilomycin A alone abrogated ifnβ induction, demonstrating that the enhanced induction originated from acidic (phago-lysosome) compartments (Figure 4C). Immunofluorescence microscopy revealed that monensin treatment did not block phagosome maturation and bacterial degradation in activated macrophages except when combined with bafilomycin A (Figure 4D). These results demonstrated that while acidification of the phagosome was not required for IFN-β signaling, phagosome maturation and bacterial degradation were necessary for this response to LLO-minus mutants.
Since the V-ATPase contributes to the electrochemical potential across the phagosomal membrane [36], bafilomycin A treatment also results in dissipation of phagosomal membrane potential, which is not the case with monensin and nigericin [37]. We asked whether the phagosomal membrane potential was important for IFN-β signaling. In order to address this question, we dissipated the phagosomal membrane potential of monensin-treated cells by using the K+-specific ionophor valinomycin. Valinomycin transports K+ ions in accordance with existing chemical gradients, which together with monensin treatment results in leakage of accumulated K+ ions from the phagosome to the cytosol and, consequently, depolarization of the phagosomal membrane (Figure 4C). Treatment of valinomycin resulted in lower levels of ifnβ induction, although it did not affect degradation of bacteria (Figure 4C and 4D). These results suggest that the phagosomal membrane potential is required for IFN-β signaling in response to the LLO-mutant.
Next, we tested the role of NOD2 in the enhanced production of IFN-β by monensin treatment; interestingly, the induction of IFN-β was reduced by 4-fold in NOD2-deficient macrophages (Figure 4E). Possibly, monensin treatment impaired the signaling from TLRs that require acidic pH [5], leading to a greater effect on IFN-β expression in the NOD2-deficient macrophages. Importantly, NOD2 dependency was specific for a subset of cytokines like IFN-β, as other cytokines like IL-12p40, which are induced by TLRs, were unaffected by the NOD2 mutation (Figure 4E).
Recently, it was demonstrated that potassium ion flux plays a role in the activation of several proteases in the phagocytic vacuole of neutrophils [38]. Since treatment with monensin generates an influx of potassium ions into the phagosome, we were interested in whether lysosomal proteases play a role in the enhancement of IFN-β signaling by monensin treatment. In order to test this hypothesis, we treated cells with the protease inhibitor, chymostatin, which inhibits lysosomal serine and cysteine proteinases and several cathepsins. Chymostatin treatment resulted in a 4-fold decrease in the ifnβ expression and had no additional effect in NOD2-deficient macrophages (Figure 4E). Moreover, chymostatin treatment was also specific to ifnβ induction and had no effect on il-12p40 induction (Figure 4E). These results suggest that bacterial lysis and further digestion by lysosomal proteases are required for generation of a NOD2 ligand, leading to ifnβ induction.
NOD2 is activated by MDP, but how MDP is generated and transported to the cytosol is unknown. The phago-lysosome contains enzymes that can potentially degrade PGN of bacteria, such as lysozyme [39]. However, while L. monocytogenes PGN is resistant to lysozyme cleavage [40], we asked whether it is still cleaved in the phago-lysosomes of activated macrophages. In order to address this question, we labeled the PGN of LLO-minus bacteria prior to infection with fluorescent vancomycin (FL-Van) that binds specifically to the terminal D-alanyl-D-alanine moieties of PGN [41]. Vancomycin labeling is localized to sites of nascent PGN synthesis, which results in polar bacterial staining (Figure 5A). Whereas in non-activated macrophages we could detect intact bacteria, in activated macrophages the FL-Van PGN labeling was distributed in large vacuoles, most likely due to bacterial cell wall (CW) breakdown (Figure 5A). The FL-Van PGN labeling was localized to acidic vacuoles determined by LysoTracker RED staining (not shown). While labeling of intracellular bacteria with FL-Van was not as efficient as the GFP labeling, we were able to detect large vacuoles (larger then a bacterial cell) in which the FL-Van labeling was equally distributed, suggesting that the bacterial PGN was released in these vacuoles.
Next, we examined whether L. monocytogenes PGN contains a NOD2 ligand and whether it can induce cytokines production in a NOD2-dependent manner. The CW fraction of L. monocytogenes was purified and treated with RNAse A and DNAse I to prevent nucleic acids contamination. While L. monocytogenes intact CW induces tumor necrosis factor (TNF) α and IL-6, that response was independent of NOD2 (Figure 5B). When the CW was degraded in vitro with the muramidase mutanolysin that specifically cleaves the PGN glycan backbone (NAG-NAM), the degradation products (like MDP) triggered TNFα and IL-6 induction, which was largely dependent on NOD2 (Figure 5B). The delivery of CW fragments and MDP to the macrophages cytosol was mediated by lipofectamine, which resulted in a 5-fold increase in the response to these ligands. Interestingly, only CW-derived fragments and not intact CW or MDP itself induce ifnβ. However, this induction was only partially (30%) dependent on NOD2 (Figure 5B). Since IFN-β expression was induced by purified bacterial CW fragments, we addressed the role of MyD88 and Trif in ifnβ induction. We found that neither of these adaptors was required for ifnβ induction, raising the possibility that cytosolic microbial sensors are involved in this response (Figure 5C). These results demonstrated that L. monocytogenes PGN contains a NOD2 ligand that becomes accessible only after degradation with a muramidase, and that it can induce a pro-inflammatory response when delivered to the cytosol. However, the exact nature of the PGN fragments generated in the phago-lysosome and the in vivo ligand recognized by NOD2 are not yet known.
We have investigated the relationship between two fundamental processes of macrophages, degradation of bacteria and induction of innate immune response. In vivo, IFN-γ has a major role in controlling L. monocytogenes infection. By activating immune cells such as macrophages, IFN-γ enhances the bacteriocidal activity of macrophages and renders them less permissive to L. monocytogenes replication. In this study, we used the LLO-minus mutant to ask whether the bacteriocidal activity of macrophages has a role in the innate immune response to L. monocytogenes infection. Here, we show that phagosomal-degraded bacteria induce a specific innate immune response that is different than the response to phagosomal intact bacteria. We found that phagosomal-degraded bacteria induced type I IFN, a response that was previously shown to be specific to intracellularly growing L. monocytogenes. This research report shows that the cytosolic receptor NOD2 enhances the induction of IFN-β by phagosomal-trapped L. monocytogenes, but only when these bacteria are killed and degraded in the phago-lysosomes of IFN-γ–activated macrophages. To our knowledge, this is the first study to demonstrate that bacterial breakdown products generated in the phago-lysosome are targets for intracellular innate immune sensors. This study suggests that induction of IFN-β in response to L. monocytogenes in vivo might result from two distinct signaling pathways, one of them largely dependent on NOD2 [20]. These results are consistent with a model in which bacterial breakdown products generated in the phagosome are transported to the cytosol, where they are detected by cytosolic microbial innate immune receptors.
NOD2 is activated by small muropeptides derived from bacterial PGN [29,42]. Although MDP (MurNAc-l-Ala-d-Gln) was shown to be the minimal motif recognized by NOD2, the natural ligand(s) sensed by NOD2 in vivo are not known. Mammals and bacteria contain muramidases, like lysozyme, or in the case of bacteria, lytic transglycosylases, that can generate muropeptide analogs of MDP such as GlcNAc-MurNAc-l-Ala-d-Gln [43]. However, only bacteria are known to possess the endopeptidases with the specificity to generate the NOD2 ligands. For example, L. monocytogenes secrets a highly expressed endopeptidase, p60, that cleaves the bond between D-Glu and meso-DAP, thus potentially generating NOD2 ligands [44]. In the phago-lysosome, L. monocytogenes is potentially subjected to degradation by both bacterial and host enzymes. We found that protease inhibitors blocked the NOD2-dependent response without affecting the TLR-dependent response. Therefore, lysosomal proteases might contribute to degradation of the bacterial CW, thereby facilitating further digestion of the PGN. A role for proteases was most evident upon H+/K+ ionophors treatment. Interestingly, a role for potassium influx and protease activation in neutrophils has been proposed by Reeve et al. [38]. These investigators have suggested that potassium influx into phagocytic vacuoles results in the release of cationic proteases from the anionic sulphated proteoglycan matrix, thus resulting in microbial degradation. During treatment with H+/K+ ionophors that cause a phagosomal potassium influx, we found that protease inhibitors blocked the IFN-β expression, but had no affect on expression of IL-12p40, suggesting that proteolysis is necessary to generate the ligands transported to the cytosol, but not those recognized in the phagosome. Taken together, these results show that phagosomal degradation of L. monocytogenes resulted in the NOD2-dependent production of IFN-β, presumably due to the production of the appropriate muramyl dipeptide(s) generated during bacterial degradation.
NOD1 and NOD2 are cytosolic proteins that are activated by small PGN fragments. Although the mechanism(s) that lead to PGN entry into the cytosol are not known, it was suggested that a specific transport system might be involved [45]. In the case of cytosolic pathogens such as Shigella flexneri or L. monocytogenes, it is possible that NOD1 and NOD2 recognize PGN fragments released during bacterial growth [46,47]. In addition, NOD1 and NOD2 may be activated by PGN fragments, introduced into host cells during infection by pathogens that possess auxiliary secretion systems. Indeed, activation of NOD1 occurred upon infection of epithelial cells by Helicobacter pylori [48]. A third hypothesis, consistent with our data, is that bacterial products leak or are actively transported from the phagosome to the cytosol [1]. An example for active transport of NOD2 ligands was demonstrated in epithelial cells where the peptide transporter hPepT1 was shown to transport MDP into the colonic epithelial cells, leading to NF-κB activation [45]. Although the mechanism of transport in macrophages is not known, here we present pharmacological evidence that the phagosomal membrane potential is crucial for NOD2-dependent IFN-β responses. While monensin and valinomycin treatments did not affect degradation of bacteria, addition of valinomycin to monensin-treated cells resulted in dissipation of the phagosomal membrane potential and reduction in IFN-β response. This observation leads to a hypothesis that the phagosomal membrane potential could have a direct role as a driving force for transport of ligands into the cytosol, or might be involved indirectly by affecting the function of membrane proteins involved in the transport process.
There are a number of well-characterized host signal–transduction pathways stimulated by viral and/or bacterial products that result in the production of type I IFN [17]. Intriguingly, DNA and RNA are recognized by both intracellular TLRs and by caspase recruitment domain (CARD)–containing cytosolic proteins, both leading to type I IFN production. Therefore, we initially hypothesized that the induction of IFN-β by degraded L. monocytogenes was likely associated with one of these TLRs, as bacterial nucleic acids are undoubtedly released into the phagosome upon bacteriolysis. However, none of the individual TLR knockouts had a measurable affect on the production of IFN-β, although it is still possible that there is an overlapping affect of individual TLRs. Recently, two cytosolic dsRNA helicases with CARD-domains, RIG-I and MDA5, have been identified and shown to lead to IFN-β production in response to viral nucleic acids by interacting with the mitochondrial membrane adaptor called MAVS/VISA/IPS-1/Cardif [15]. Interestingly, microarray data revealed that whereas MDA5 is induced only by w.t. L. monocytogenes in non-activated macrophages, it is induced by both w.t. and LLO-minus bacteria in activated macrophages (Figure 1D). Since L. monocytogenes DNA and RNA preparations were able to induce IFN-β when delivered to the cytosol (A. Herskovits, unpublished data; [22]), it is possible that cytosolic nucleic-acids sensors such as MDA5 and RIG-I are contributing to the expression of IFN-β in response to degraded L. monocytogenes.
This report has demonstrated that NOD2 activation enhances IFN-β production, and although we do not know the exact nature of this signaling pathway, since NOD2 contains two CARD domains, we speculate that it may be interacting with other CARD-containing adapters. Whereas purified MDP, the synthetic activator of NOD2, did not lead to IFN-β production, when combined with poly (I:C), it enhanced the induction of IFN-β in a NOD2-dependent manner. Our in vitro preparation of CW fragments recapitulated the induction of IFN-β when delivered to the macrophage's cytosol. However, this induction was only partially dependent on NOD2. Since NOD2 is known to act synergistically with other innate immune receptors [30–34], we suggest that the NOD2-dependent IFN-β response is a result of synergism between PGN fragments generated in phagosomes and other bacterial ligands exposed upon bacterial lysis. Interestingly, it was recently shown that delivery of L. monocytogenes DNA to the cytosol promotes IFN-β production [22], raising the possibility that release of bacterial nucleic acids in the cytosol might be involved in type I IFN response to degraded bacteria.
This study highlights the downstream consequences that result from the enhanced microbiocidal and degradative capacity of IFN-γ–activated macrophages. Whereas non-activated macrophages show some bacteriocidal capacity, only activated macrophages killed and degraded bacteria. Bacterial degradation has a number of potential immunological consequences. A well-established consequence of phagosomal degradation is the generation of bacterial peptides ligands, which leads to the development of major histocompatibility complex class II–dependent responses [49]. In this study, we show that digestion of bacteria in a phagosome results in induction of specific innate immune responses that differ from a response to intact bacteria. We demonstrated a direct correlation between bacterial degradation and macrophage expression of IFN-β. Macrophages that failed to degrade bacteria failed to express IFN-β. Blocking degradation of bacteria by bafilomycin A treatment resulted in loss of signaling. While the role of the V-ATPase in phago-lysosome maturation is well characterized, we found that the signaling adaptor MyD88 was also essential for bacterial degradation. MyD88-deficient macrophages failed to degrade bacteria or express IFN-β when presented with intact bacteria, but did express IFN-β when bacterial CW fragments were delivered directly to the cytosol. Although the role of MyD88 in phagosome maturation is controversial, it is possible that MyD88 is involved in the initial signaling pathways that promote macrophage activation and their capacity to kill and degrade bacteria. Indeed, it was shown that the induction of many genes by IFN-γ is dependent on MyD88 [50]. Recently, it was suggested that MyD88 is required for proper assembly of the phagosomal NADPH oxidase, thus affecting the killing of Gram-negative bacteria [51]. It is clear that immune signaling pathways affect cellular processes in specialized phagocytic cells, but how these processes, such as pathogen digestion and generation of new ligands, are involved in further shaping the immune response is less understood. We speculate that macrophages can discriminate between ligands that are presented by live bacteria or ligands that are generated after degradation of bacteria (post-mortem). We refer to these ligands as pathogen-associated molecular patterns post-mortem (PAMP-PM). The results of this study suggest that a fully active phagosome provides PAMP-PM for detection by the innate immune system. However, this “information” is not restricted to receptors that are located in the phagosome, but crosses the phagosomal membrane to activate intracellular immune receptors as well. Lastly, we suggest that bacterial degradation in the phagosome plays a major role in the development of innate and acquired immune responses.
The L. monocytogenes strains used were a w.t. strain, 10403S and 10403S expressing GFP, or strains containing in-frame deletions of the hly gene (LLO, DP-L2161) [52] and DP-L2161 expressing GFP [53]. Single colonies were inoculated into 2 ml of BHI broth (brain-heart infusion) and incubated overnight at 30 °C without shaking.
C57BL/6 mice were obtained from The Jackson Laboratory (http://www.jax.org). CD-1 mice were obtained from Charles River Laboratory (http://www.criver.com). All knockout mice used in this study were on the C57BL/6 background or backcrossed with C57BL/6 mice for at least eight generations. Femurs or mice were obtained from the following source: myd88–/– from R. Medzhitov, Yale University School of Medicine, New Haven, Connecticut; tlr3−/−, tlr7−/−, tlr9−/− from K. A. Fitzgerald and D. Golenbock, University of Massachusetts Medical School, Worcester, Massachusetts; trif−/− (lps2/lps2), myd88trif−/−, from B. Beutler, The Scripps Research Institute, La Jolla. California; nod2−/− [54] from V. Dixit, Genentech, South San Francisco, California; nod1−/− from Millennium, Cambridge, Massachusetts; irf3−/− from G. Cheng, Department of Microbiology, Immunology and Molecular Genetics, University of California, Los Angeles, California.
Primary cultures of resident peritoneal macrophages were prepared from CD-1 mice as previously described [27]. BMD macrophages were isolated from 6- to 8-wk-old female mice and cultured as described [55]. Activation of macrophages was done by incubating macrophage monolayer with 1 ng/ml recombinant murine IFNγ (Biosource, http://www.biosource.com) for 36 h before infection and during infection. Approximately 8 × 106 w.t. L. monocytogenes or 1 × 108 LLO-minus bacteria were used to infect 2 × 106 macrophages cells seeded on a 60-mm petri dish. These numbers resulted in infection of one to two bacteria per cell in the case of w.t. L. monocytogenes, and ~25–50 bacteria per cell in the case of the LLO-minus mutant. Thirty minutes after addition of bacteria, macrophage monolayers were washed three times with PBS, and fresh medium was added. At 1 h.p.i., gentamicin was added to 50 μg/ml to limit the growth of extracellular bacteria. Unless indicated otherwise, infections were completed at 6 h.p.i, and further analyzed. Activation of macrophages was confirmed by visual inspection after infection with GFP-expressing bacteria (Figure 1C). Where indicated, 0.5 μM bafilomycin A (Sigma, http://www.sigmaaldrich.com), 1 μM monensin (Sigma), 0.1 μM nigericin (Sigma), 1 μM valinomycin (Sigma), were added at 30 min post-infection. The protease inhibitor chymostatin (100 μM) (Sigma), was added at the time of infection. Growth curves of L. monocytogenes in macrophages cells were performed as described earlier [27].
MEEBO microarrays were printed at the Center for Advanced Technology at University of California San Francisco [28]. Each microarray experiment was done in triplicate. Macrophages were infected with w.t. and LLO-minus mutant for 5 h with and without IFN-γ treatment. Then, macrophages were washed with PBS, and total RNA was extracted using RNAqueous kit (Ambion, http://www.ambion.com). A half microgram of RNA was amplified using Amino Allyl MessageAmp II aRNA Amplification Kit (Ambion). A total of 5 μg of amplified RNA from each sample was labeled fluorescently with Cy5 (Amersham, http://www.gelifesciences.com) and mixed with a Cy3 (Amersham)–labeled reference pool. The common reference pool contained equal amounts (5 μg of each) of amplified RNA from all the samples analyzed in the experiment (including uninfected, hly, and w.t. samples from activated and non-activated macrophages). RNA was hybridized to the MEEBO microarrays for 48 h. Microarrays were gridded using SpotReader software (Niles Scientific, http://www.nilesscientific.com) and GenePix Pro 6 (Molecular Devices, http://www.moleculardevices.com), and analyzed using Acuity 4 software (Molecular Devices). Highest quality spots meeting extra-stringent criteria were identified. These highest quality spots were used to calculate normalization factors such that the median Cy5/Cy3 ratio was brought to 1. These factors were then applied to the entire dataset after removing low quality spots. Finally, all arrays were normalized to the uninfected non-activated macrophages array. Significant analysis of microarrays (SAM) algorithm was used with two-class unpaired designs to identify genes that were differentially expressed in w.t. versus LLO-minus infection of non-activated macrophages, and in LLO-minus mutant infections of activated versus non-activated macrophages. Genes that showed a 5.6-fold or greater difference were selected for further analysis. Pearson hierarchical clustering was applied on selected genes.
One liter of L. monocytogenes culture was grown at 37 °C to exponential phase and used for CW preparation. Bacterial cells were lysed by three passages through a French press at 12,000 PSI, and treated with DNAse I (Invitrogen, http://www.invitrogen.com). After removal of cell debris, supernatant was added drop by drop to 8% boiling SDS with stirring, and was boiled for an additional 30 min [56]. The mixture was cooled overnight at room temperature, and washed with hot water. CW was washed in 0.1% Triton X-100, and then washed six times with water and stored at −20 °C. Mutanolysin treatment was done overnight in 50 mM MES (pH 5.88), 1 mM MgCl2. Insoluble CW was pelleted and the supernatant pH was adjusted. CW preparation and CW fragments were treated with RNAse A (Fermentas, http://www.fermentas.com). L. monocytogenes CW fragments, 100 μg/ml MDP (Calbiochem, http://www.emdbiosciences.com/html/CBC/home.html) and 100 μg/ml poly(I:C) (Invivogen), were delivered to macrophage cytosol with Lipofectamine 2000 (Invitrogen).
RNA was harvested from macrophages at 6 h.p.i. using RNeasy Mini kit (Qiagen, http://www.qiagen.com). To synthesize cDNA, 1 μg of total RNA, M-MLV reverse transcriptase, Random Primers, and RNaseOUT ribonuclease inhibitor (Invitrogen) were used. For regular PCR analysis, 1 μg of cDNA was used with specific primers. SYBR Green−based quantitative PCR amplification was performed in 96-well plates using SYBR Green PCR core reagents (Applied Biosystems, http://www.appliedbiosystems.com), the Stratagene Mx3000P Real-Time PCR System (http://www.stratagene.com, and a 60 °C annealing temperature. For each indicated gene, as well as to the reference gene (actin), a standard curve was generated to calculate the quantity of mRNA as function of the Ct value. The level of expression of each gene was determined by normalizing its mRNA quantity to the quantity of the actin mRNA at the same sample. The following mouse primer sequences were designed using Applied Biosystems Primer Express software: ifnβ-F: 5′-ctggagcagctgaatggaaag; ifnβ-R: 5′-cttgaagtccgccctgtaggt; β-actin-F: 5′-aggtgtgatggtgggaatgg; β-actin-R: 5′-gcctcgtcacccacatagga; tnfα-F: 5′-gcaccaccatcaaggactcaa; and tnfα-R: 5′-tcgaggctccagtgaattcg; il-6-F: 5′-ttccatccagttgccttcttg; and il-6-R: 5′-gaaggccgtggttgtcacc; il-12p40-F: 5′- aaccatctcctggtttgcca; and il-12p40-R: 5′- cgggagtccagtccacctc.
IFN-β was measured by a mouse IFN-β enzyme-linked immunoassay (ELISA) kit (R&D Systems, http://www.rndsystems.com).
Bacterial PGN was labeled as described by Daniel and Errington [41]. A mixture of 1:1 FL-Van (Molecular Probes, http://probes.invitrogen.com) and unlabeled vancomycin (Sigma) was added to growing cultures to a final concentration of 1 μg/ml. The culture was then incubated for 30 min at 37 °C to allow absorption of the antibiotic and then washed four times in PBS. For immunofluorescence microscopy, FL-Van–labeled bacteria or GFP-expressing bacteria were used to infect macrophages on 18-mm2 glass cover slips. LysoTracker RED staining was performed according to manufacturer instructions (Molecular Probes). Then, 4 h.p.i., macrophages were washed once with PBS and fixed in 4% paraformaldehyde. Cover slips were incubated with coumarin-phalloidin or tetramethylrhodamine B isothiocyanate-phalloidin (Sigma, 1:1,000 dilution) for cytosolic F-actin staining and mounted with Vectashield mounting medium with DAPI (Vector Laboratories, http://www.vectorlabs.com). Samples were viewed at ×600 with a Nikon TE300 inverted microscope.
MEEBO (http://meebo.ucsf.edu:8080/meebo/meeboQuery.html) gene ID numbers for the genes and gene products discussed in this paper are 5′ nucleotidase (mMC016279); C-C receptor like 2 (mMC024156); C-X-C ligand 1 (mMC023634); CXC ligand 11 (mMC011370); dual specificity phosphatase 16 (mMA034830); IFN activated gene 205 (mMA032762); IFN-β (mMC16397); IL-1α (mMR001431); IL-1 receptor agonist (mMC015067); IL-4 (mMC019427); IL-12 (mMC018187); IL-15 (mMC009424); IL-15 receptor (mMA033400); Mda5 helicase (mMC010553); myxovirus resistance gene (mMC023295); NF-κB light polypeptide zeta (mMC002655); NOD2 receptor (mMR030202); TNFα-induced protein 2 (mMC011682); and TNF receptor super-family member 5 (mMR028074). |
10.1371/journal.pntd.0004512 | Unbiased Characterization of Anopheles Mosquito Blood Meals by Targeted High-Throughput Sequencing | Understanding mosquito host choice is important for assessing vector competence or identifying disease reservoirs. Unfortunately, the availability of an unbiased method for comprehensively evaluating the composition of insect blood meals is very limited, as most current molecular assays only test for the presence of a few pre-selected species. These approaches also have limited ability to identify the presence of multiple mammalian hosts in a single blood meal. Here, we describe a novel high-throughput sequencing method that enables analysis of 96 mosquitoes simultaneously and provides a comprehensive and quantitative perspective on the composition of each blood meal. We validated in silico that universal primers targeting the mammalian mitochondrial 16S ribosomal RNA genes (16S rRNA) should amplify more than 95% of the mammalian 16S rRNA sequences present in the NCBI nucleotide database. We applied this method to 442 female Anopheles punctulatus s. l. mosquitoes collected in Papua New Guinea (PNG). While human (52.9%), dog (15.8%) and pig (29.2%) were the most common hosts identified in our study, we also detected DNA from mice, one marsupial species and two bat species. Our analyses also revealed that 16.3% of the mosquitoes fed on more than one host. Analysis of the human mitochondrial hypervariable region I in 102 human blood meals showed that 5 (4.9%) of the mosquitoes unambiguously fed on more than one person. Overall, analysis of PNG mosquitoes illustrates the potential of this approach to identify unsuspected hosts and characterize mixed blood meals, and shows how this approach can be adapted to evaluate inter-individual variations among human blood meals. Furthermore, this approach can be applied to any disease-transmitting arthropod and can be easily customized to investigate non-mammalian host sources.
| Female mosquitoes require a blood meal to acquire the nutrients necessary for egg production. While feeding on host species, mosquitoes can transmit pathogens that cause several diseases including malaria, lymphatic filariasis and dengue. Understanding the mosquito host choice is important to better implement control strategies to reduce mosquito populations and therefore transmission of disease. Currently, the majority of methods for evaluating host species only test for the presence of pre-selected, expected hosts. Here, we describe an unbiased assay that combines amplification of any mammalian DNA with high-throughput sequencing to comprehensively characterize the composition of mosquito blood meals. We applied this approach to Anopheles mosquitoes collected in Papua New Guinea and observed that they fed on expected (humans, dogs and pigs) and unexpected hosts (mice, bats, marsupials). In addition, we show that 16.3% of mosquitoes fed on multiple hosts, from the same or different species. Overall, this approach enables unbiased characterization of mosquito blood meals and can be easily applied to significantly improve our understanding of the feeding behavior of any disease-transmitting insect.
| Many insects require a blood meal to complete their gonotrophic cycle. By feeding successively on different hosts, these insects can transmit blood borne pathogens that cause diseases responsible for significant burden on global health [1, 2]. In particular, insects that seek human blood meals are vectors of devastating diseases such as malaria, dengue fever, sleeping sickness, filariasis, leishmaniasis, typhus and plague. Understanding the complex blood feeding patterns of the insects transmitting these human diseases is crucial for developing and prioritizing vector-based control program activities and identifying potential unrecognized disease reservoirs, and thus for reducing disease transmission and burden.
The blood meals of arthropods have traditionally been analyzed using serological techniques such as ELISA or precipitin tests [3–5]. While these methods have provided valuable information, they have limited taxonomic resolution as they are generally only able to characterize hosts at the order or family levels [6]. In addition, since these approaches test for the presence of a protein from a specific organism, they only test for absence/presence of organisms that are a priori believed to be blood meal hosts. More recently, a number of PCR-based molecular techniques have been developed to characterize host blood meals ([7] and references within) and determine the blood feeding preference of mosquitoes [8–11], ticks [12–14], sandflies [15–17] and Tsetse flies [18, 19]. While these PCR-based approaches enable rigorous identification of the host species, they typically focus on species-specific amplification of putative hosts and therefore are not designed to identify novel, unanticipated host blood sources. In addition, the detection of mixed blood meals (i.e., when an insect feeds on more than one host) by these approaches is complicated as the dominant host signal can completely overwhelm signals from other minor hosts. These limitations may have biased our understanding of the transmission of many vector-borne diseases and have prevented identification of important disease reservoirs.
Beyond the identification of the host species, it may also be important to understand which individuals of a given species are being fed upon: for example, knowing whether an insect preferentially bites specific individuals or, in contrast, feeds on multiple individuals per night, could influence our assessment of disease transmission. A number of studies have used microsatellites or other polymorphic genetic markers to generate individual DNA fingerprints from human blood meals of mosquitoes [20–25] and lice [26, 27]. However, interpretation of these data can become complicated if DNA from more than one individual is present in a single blood meal.
Anopheles punctulatus sensu latu (s.l) mosquitoes are the principal vectors of malaria and lymphatic filariasis in Papua New Guinea (PNG) and the South Pacific [28]. There are 13 sibling species in An. punctulatus s.l, five of which are major disease vectors: An. punctulatus s.s., An. koliensis, An. farauti s.s., An. hinesorum and An. farauti 4. While these species have been little studied, they are generally characterized as unspecialized with regards to their feeding behaviors and ecological preferences [29] and shown to feed roughly indiscriminately on humans, dogs and pigs that are the most abundant species found in PNG villages [30, 31].
Here, we describe a novel approach using next-generation sequencing technology to analyze the blood meal composition of individual mosquitoes in an unbiased manner. We first amplify DNA extracted from a single female mosquito using universal primers targeting the mammalian mitochondrial (mt) 16S rRNA genes. Following individual barcoding, PCR products from up to 96 mosquitoes are pooled and simultaneously sequenced using Illumina high-throughput sequencing methods. We also use the same approach to interrogate whether individual mosquitos fed on more than one person by sequencing a highly polymorphic region of the human mt hypervariable region I. We applied this approach to 442 Anopheles punctulatus sensu lato (s.l) mosquitoes captured in five villages of the Madang Province of Papua New Guinea and provide evidence that (i) Anopheles punctulatus s.l. mosquitoes feed on a variety of mammalian species, including several unanticipated hosts, and (ii) Anopheles punctulatus s.l. mosquitoes frequently feed on multiple mammalian hosts. We also show how this assay can be easily customized to examine the number of individual hosts within a specific species. Overall, our results illustrate the potential of this approach to comprehensively characterize host species for any blood feeding arthropods, to identify reservoirs of pathogens and to provide opportunities for developing better evidence-based strategies to decrease transmission of important infectious diseases.
This study was approved by the Papua New Guinea Institute of Medical Research Institutional Review Board (1203) and PNG Medical Research Advisory Board (12.05).
We collected mosquitoes from the villages of Dimer, Wasab, Kokofine, Mirap and Matukar in the Madang province of Papua New Guinea (PNG) in June and August 2012. In each village, field technicians collected mosquitoes between 1800 and 0600 using barrier screens as described by Burkot et al [32]. These screens were manually searched every 20 minutes and resting mosquitoes were captured from the screens using an aspiration device. After collection, the sex and species of each mosquito were determined by morphology as previously described [33]. All male mosquitoes and non-Anopheles mosquitoes were discarded. We visually classified each female Anopheles mosquito as non-fed, partially-fed or fully-fed by examining the size and coloration of their abdomen. We individually stored each mosquito in an Eppendorf tube containing silica gel as desiccant.
We extracted DNA from individual mosquitoes using a 96 well Qiagen DNeasy blood and tissue kit as previously described [34]. Mosquito species identification was determined using a PCR-based assay that evaluates species-specific polymorphisms in the ribosomal RNA internal transcribed spacer unit 2 (ITS2) [35].
To test the range of mammals that should be amplified using mt 16S rRNA primers [36], we conducted an in silico analysis using the primerTree package. We also conducted in silico analyses for two other previously published primers, cytochrome oxidase I (COI) and cytochrome b (Cytb) that have been previously used for mosquito blood meal identification [37]. Since the 16S rRNA locus appeared to be the most informative for our purposes (S1 Fig), we restricted our further analyses to this locus. Briefly, we performed primer-BLAST searches using the mammalian mt 16S rRNA primer sequences against the NCBI nucleotide database using default parameters but allowing for up to three mismatches in the primer sequences. In our search, we set the maximum number of blast hits retrieved to 10,000 and retrieved the taxonomical information of each sequence retrieved. As this search can be biased by recent release of many DNA sequences from a specific taxon, we performed this search separately for each mammalian order. We then calculated how many different species were obtained from each order to calculate the total number of mammalian species likely to be amplified by this primer pair. Note that, when conducting the search without any taxonomic restrictions, these mammalian primers were also predicted to amplify amphibian and fish 16S rRNA genes.
To estimate the total number of mammalian species for which the targeted locus has been sequenced and deposited in NCBI, we randomly selected one DNA sequence from each mammalian family and used BLAST searches to identify similar DNA sequences in the NCBI nucleotide database (accessed on July 2015). We filtered out any DNA sequence from the database that did not contain the primer sequences (allowing for up to three mismatches). We then merged the results from the searches performed in each family and counted how many unique species were observed. These analyses provided us with the total number of mammalian species that should be amplified if the primers were truly universal.
We also evaluated whether the 16S rRNA primers amplified sufficiently informative DNA sequences to support rigorous species identification (i.e., whether related species could be distinguished). First, we retrieved the mammalian DNA sequence alignment from the primerTree analysis and calculated the number of nucleotide differences (including deletions) between every pair of DNA sequences using the dist.dna program of the Ape package [38]. We then calculated the average proportion of nucleotide differences between species belonging to the same mammalian order and between species belonging to different orders. Second, we used the same approach to determine, for each mammalian order, how often two different species (or genera) have the exact same DNA sequence for the targeted region of the 16S rRNA gene.
To interrogate the mammalian species composition of individual mosquito blood meals we amplified a 140 bp of the mammalian mt 16S rRNA gene using universal mammalian primers [36] modified to include a 5’-end tail complementary to the Illumina sequencing primers (S1 Table). We also attempted to amplify a subset of 192 mosquitoes with universal avian primers ([39] and S1 Fig) using a pooled approach but failed to detect any bird DNA in these samples. To identify individual differences among human blood meals, we designed PCR primers to amplify 300 bp of the human mt hypervariable region I. We first aligned 795 whole mt genomes of individuals from Oceania [40] using MAFFT version 7 [41] to evaluate the extent of sequence variation across the mt genome hypervariable region I and then designed primers positioned in conserved flanking sequences with Primer3 [42]. As described above, we added a 5’ tail to each primer for sample barcoding and high-throughput sequencing (S1 Table).
For each sample and amplicon, we performed two rounds of PCR amplification to prepare products for Illumina sequencing (Fig 1). First, we performed a locus-specific amplification (i.e., targeting either the mammalian mt 16S rRNA or the human mt hypervariable region) using the Promega GoTaq PCR kit protocol (50 μL reaction) with 1μL of genomic DNA, 0.2mM of each dNTP, 1.25 units of GoTaq DNA polymerase, 4mM of magnesium and 0.2 μM primers. PCR amplification was carried out under the following conditions: 3 minutes at 94°C followed by 30 cycles at 94°C for 45 seconds, 50°C for 45 seconds, 72°C for 30 seconds and a final elongation step at 72°C for 3 minutes. We then purified these PCR products using the QIAquick 96 PCR purification kit protocol (QIAGEN). Second, we incorporated Illumina adaptors, including unique 6-nucleotide sample identification barcode sequence through 10 additional PCR cycles, using barcoding primers complementary to the 5’-end tail incorporated during the first PCR amplification (Fig 1). For these reactions, we used the Promega GoTaq protocol as described above with 1uL of PCR product being added to each reaction. The same thermocycling conditions as described above were used but for an annealing temperature of 56°C. Predicted sizes for the mammalian mt 16S rRNA amplicons ranged from 265 to 343 bp; sizes for the human mt hypervariable region I amplicons ranged from 440 to 444 bp (amplicon sizes include Illumina sequencing primers, unique barcode sequence and Illumina adaptors, Fig 1). Finally, we pooled the barcoded amplification products from 96 individual mosquitoes and simultaneously sequenced them on an Illumina MiSeq instrument (Sequences deposited in NCBI SRA: SRP062959).
We discarded from further analyses any read that did not carry the exact barcode and primer sequences. After recording the read origin using the barcode sequence, we removed the primer and barcode sequences to only keep the amplified DNA sequences. We discarded any resulting read smaller than 50 bp as these likely represent primer dimers. Since each amplified molecule was sequenced in both directions using paired-end reads, we merged each pair of sequencing reads using PANDAseq [43] (Fig 1) keeping, at each position, the nucleotide with the highest sequencing quality. We then analyzed 16S RNA and human mtDNA sequences separately.
Using all 43,743,363 16S rRNA sequences generated from the 442 mosquitoes, we identified all unique DNA sequences using Mothur [44] and recorded the number of reads carrying each unique DNA sequence. We removed any DNA sequence that was observed less than 10 times across all samples, as these likely resulted from sequencing errors. We compared the remaining unique DNA sequences against all DNA sequences present in the NCBI nucleotide database using blastn. For each DNA sequence, we recorded the best match(es), only considering sequences with > 90% identity over the entire sequence length. We then retrieved the taxonomic information from each best-matched sequence using the ‘get_taxonomy’ function in PrimerTree. When an amplified sequence matched multiple species equally well, we recorded all species names associated with that sequence. We then summarized the blood meal of each mosquito by calculating the proportion of reads matching each species. As a small number of reads generated could reflect low level PCR contamination or an error in the sequence barcode identification, we only analyzed mosquito samples with at least 1,000 reads (S2 Fig). For the same reason, we considered a mosquito as having fed on a single mammalian host if >90% of the sequencing reads aligned to the 16S rRNA of that species. Alternatively, if >10% of the sequencing reads aligned to a second species, we considered the mosquito to have fed on multiple mammalian hosts.
For the human mt hypervariable region, we aligned the consensus reads to the human mitochondrial reference genome sequence (NC_012920.1) using bowtie2 [45] and calculated, for each sample, the number of reads supporting each haplotype. Only haplotypes supported by more than 500 reads were considered to avoid incorporating sequencing or PCR errors (i.e., rare haplotypes that differed from an abundant haplotype by one nucleotide difference) in the analyses. Finally, we reconstructed a phylogenetic tree with all identified human mt haplotypes using MEGA version 6 [46].
We first conducted extensive in silico analyses to confirm that the primer pair selected [36] could amplify DNA sequences from a wide range of mammalian orders including Primates, Rodentia (rodents), Artiodactyla (even-toed ungulates), Carnivora (carnivorans), Chiroptera (bats), Cetacea (cetaceans), Insectivora (insectivores) and Marsupials (Table 1). Overall, in silico analysis predicted that these primers should amplify 1,752 of the 1,779 mammalian species (98.5%) sequenced at this locus and present in the NCBI nucleotide database (Table 1). Besides mammals, these primers were predicted to also amplify Actinopteri (bony-fishes) and Amphibia (amphibians) (S3 Fig).
In addition to amplifying a wide range of targets, our approach requires primers to amplify DNA sequences containing enough information to identify each species specifically. We tested this parameter by comparing the DNA sequences predicted to be amplified by this primer pair (see Methods for details). Despite the short amplified DNA sequence (~140 bp), these primers enabled rigorous differentiation of most mammalian species as illustrated by the average proportion of nucleotide differences (including deletions) between sequences of species belonging to the same or different order (S2 Table). For example, 27% of the nucleotide positions at this locus differ, on average, between one Carnivora and one Primate species and 17% of the nucleotides differ between the sequences of two Carnivora species. This high discriminating ability is also shown by the long branch lengths displayed by the phylogenetic tree reconstructed using these sequences (S1C Fig). In fact, we found that one DNA sequence amplified by these primers typically matches a single genus and, in 86% of the cases, a single species (Table 1).
We analyzed mosquitoes collected in five villages in the Madang Province in PNG: Dimer (n = 45), Wasab (n = 81), Kokofine (n = 83), Mirap (n = 171) and Matukar (n = 62). These mosquitoes included several species of the Anopheles punctulatus group: An. punctulatus s.s., An. koliensis, An. farauti 4 and An. farauti s.s. We characterized the blood meal composition of a total of 442 female Anopheles by amplifying the mammalian mt 16S rRNA genes from DNA extracted from these mosquitoes, pooling the PCR products of 96 samples after individual barcoding, and simultaneously sequencing the samples on an Illumina MiSeq instrument (Fig 1). We generated a total of 43,743,363 paired-end reads of 150 bp (includes added primers). For 42,198,573 DNA sequences (96.5%), we were able to collapse the overlapping paired-ends (Fig 1) and thus correct many sequencing errors. After combining the reads generated from all samples together, we identified 2,436,277 unique DNA sequences. We discarded from further analyses 2,404,684 unique DNA sequences that were carried by less than 10 reads across all samples as these likely represent DNA sequences with rare sequencing errors (accounting for a total of 4,432,784 reads or 10.1% of the total number of reads generated). We then compared the remaining 31,593 unique DNA sequences, accounting for 39,310,579 reads (89.9%), to all DNA sequences deposited in the NCBI database. 28,999 of these DNA sequences (representing 38,375,616 reads) had > 90% nucleotide identity to at least one mammalian DNA sequence present in NCBI: 18,814 unique DNA sequences best matched a single mammalian species sequence while 10,185 unique DNA sequences matched equally well to multiple mammalian species sequences (S3 Table).
Overall, we generated an average of 82,528 reads per mosquito. The number of reads generated from each mosquito varied considerably (S2 Fig) as it depends on several factors including: the amount of starting template (i.e., quantity of mammalian DNA present in the mosquito), the amplification efficiency and uneven pooling or variations in sequencing output between MiSeq runs. For further analyses, we only considered mosquito samples with more than 1,000 reads. None of the 30 extraction controls (i.e., water samples that have been processed in parallel through DNA extraction, PCR and sequencing) reached this cutoff illustrating the low level of cross-contamination or read mis-assignment due to errors in the barcode sequence (if any). Overall, we analyzed mammalian DNA from 314 blood fed mosquitoes, including 258 out of the 337 mosquitoes characterized as fully-fed (76.6%) and 56 out of the 86 mosquitoes visually-classified as partially-fed (65.1%). Only 5 out of the 19 mosquitoes visually classified as non-fed yielded mammalian 16S rRNA sequences: four yielded exclusively human 16S rRNA sequences, the last one a mix of human and pig sequences. These DNA sequences could indicate possible contamination either during field collection or in the laboratory, or detection of DNA from a previous, partially digested, blood meal. There was no statistical difference between fully-fed and partially-fed mosquitoes, however the number of sequencing reads generated for mosquitoes visually classified as fully-fed or partially-fed were significantly different from those classified as non-fed (p<0.05, Wilcoxon Rank-Sum test, S4 Fig). In total we successfully amplified and sequenced mammalian DNA from 319 Anopheles mosquitoes.
We identified 201 Anopheles mosquitoes that carried human DNA, 111 carried pig DNA, 60 carried dog DNA and 5 carried mouse DNA (Table 2; further details in S3 and S4 Tables). In addition to these expected hosts, we identified one mosquito that carried DNA from two different bat species: 7.2% of the reads matched perfectly Dobsonia moluccenis, a fruit bat commonly found in PNG, while 5.1% of the reads were most similar (94.4% identity) to another megabat species, Dobsonia praedatrix, also endemic to PNG (Table 2). These bat DNA sequences were clearly distinct (8 nucleotide differences between them) and unlikely to have been derived from sequencing errors, indicating that the mosquito fed on two different bats (S5 Fig). Additionally, in another mosquito 13% of the total reads (7,599 reads) were most similar to the common spotted cuscus (Spilocuscus maculatus, 98% similarity), a marsupial found in the forests of PNG (S6 Fig). Note that, consistent with our in silico analysis, we were not always able to identify the exact species that was fed upon. For example, we could not differentiate Canis lupus from Canis aureus (S3 Table). Overall, these finding illustrate the unbiased nature of this sequencing approach to identify host species regardless of expectations for mosquito blood meal feeding (as long as a closely related species has been sequenced).
Out of 319 mosquitoes analyzed, 52 (16.3%) showed clear evidence of having fed on more than one host species (with >10% of the reads supporting the minor host): 44 mosquitoes carried DNA from two species and eight carried DNA from three species (Fig 2).
Within each village, we identified three major mammalian hosts—humans, dogs and pigs—accounting for 37 to 100% of each mosquito blood meal. However, the proportion of mosquitoes that fed on each host varied within and between villages (Fig 2). For example, in Mirap, only 31 of the 127 Anopheles mosquitoes (24%) fed on humans while 62 (49%) fed on pigs, 11 fed on dogs (9%) and 23 fed on two or three species (18%) including one mosquito that fed on two bat species and one mosquito that fed on a common spotted cuscus (Fig 2A). By contrast, in Kokofine, 52 out of the 62 mosquitoes fed on humans (84%) while the remaining 10 mosquitoes fed on dogs (n = 3), pigs (n = 3) or on multiple species (n = 4) (Fig 2B). The data for the three other villages are presented in Fig 2C–2E. Note that as host density information is not available for these villages, we were unable to test whether the observed differences in blood meal composition were caused by differences in mosquito feeding behavior among locations or species.
Since we observed that 16.3% of the mosquitoes analyzed had fed on multiple mammalian hosts, we hypothesized that mosquitoes could also be feeding on multiple human individuals. We therefore investigated the number of different human DNA sequences present in 157 human-fed mosquitoes, using the same approach to sequence ~300 bp of the human mt hypervariable region. We generated an average of 26,721 sequencing reads of 250 bp for each sample and successfully amplified 102 of the 157 mosquitoes for the human mt hypervariable regions yielding a total of 20 different human mtDNA sequences (S7 Fig). While a single DNA sequence was amplified from 78.5% (n = 80) of the human-fed mosquitoes analyzed, 21.5% (n = 22) mosquitoes carried two distinct DNA sequences (S7 Fig). One sequence, identified in 14 of these potential mixed human blood meal, was always present at low frequency (<8% of the reads) and was actually more similar to a region of human chromosome 11 (98% similarity) than to the mitochondrial genome sequence (91%). This DNA sequence likely resulted from the amplification of the nuclear insertion of the mitochondrial sequence (numt, [47]) and was excluded from further analyses. Nine mosquitoes, belonging to two species and collected in three locations, showed presence of two human mtDNA sequences (S5 Table). For four of these mosquitoes, only one substitution (out of the 300 bp amplified) differentiated the two sequences and these could possibly be caused by a PCR error occurring at an early cycle. However, for the remaining five mosquitoes, 5–14 nucleotide substitutions differentiated the two sequences amplified and indicated that the mosquito successively fed on multiple individuals (Fig 3 and S5 Table).
Vector-borne diseases such as dengue, malaria, Chagas disease or leishmaniasis, account for more than 17% of all human infectious diseases and cause more than one million deaths annually [48]. To control and eliminate these diseases, it is essential that we fully appreciate the diversity and relative importance of the disease hosts and vectors. For example, while birds are well known to be the primary reservoir host of Eastern Equine Encephalitis virus (EEEV), a virus transmitted by mosquitoes that can cause zoonotic infections, recent studies have shown that snakes constitute another, previously unsuspected, reservoir of EEEV [49].
Most molecular techniques used to investigate insects’ blood meal composition are specifically designed to identify one or a few specific host(s) and cannot characterize blood meal composition in an agnostic manner. Universal primer pairs have been used to circumvent this limitation and amplify any mammalian [10, 50], or vertebrate DNA [51, 52]. However, these former studies have relied on cloning the amplified products and sequencing a few clones from each insect and are consequently very expensive and labor intensive. In addition, the presence of multiple host species in a blood meal complicates the sequence analysis when the amplification product is sequenced directly (resulting in high background noise) or further increases the cost of the experiment if the PCR products are cloned and several clones sequenced per mosquito. These challenges have limited the number of studies that rigorously examined mixed blood meals from disease vectors and provided a potentially incomplete perspective on these vectors’ feeding patterns. Rigorous identification of mixed blood meals is however critical to understand disease transmission as it might reveal higher transmission rates, if a blood meal typically consist of the blood from multiple individuals, or, lower, if the insect often feeds on species not susceptible to infection.
By contrast, a unique strength of the assay described here is its ability to rigorously detect and quantify mixed blood meals by identifying, in a single mosquito, the presence of multiple species’ DNA even if they only contribute to a small fraction of the entire blood meal (down to 10% in the current study). We were able to accurately detect and quantify mixed blood meals due to the high sequencing coverage achieved by high-throughput sequencing: on average, mammalian mt 16S rRNA genes amplified from each mosquito was sequenced by 82,528 reads and, therefore, even minor host DNA present in 10% of the total mammalian DNA was represented by several thousand reads. Note that the DNA amplification might have different efficiency for different DNA sequences (e.g., amplify better pig than dog and human DNA). Consequently, the proportion of reads obtained from each species might not reflect the true proportions of these species in the blood meal (especially since the mtDNA content in blood might also vary among species). However, this possible bias will affect all samples similarly and will not interfere with comparisons of the blood meal composition across samples. In addition the host DNA is degraded after the blood meal and the time between the mosquito’s meal and sample collection could therefore influence the interpretation of the results. Note that in mosquitoes, host blood meals can typically be detected up to 24–30 hours post-feeding, but have been detected up to 48 hours post-feeding [8].
The second key feature of our approach is its ability to detect novel blood hosts that would not have been detected using traditional techniques. For example, here we report the first observation that Anopheles mosquitoes can feed on bats and marsupials. Importantly, all the hosts identified in our study are endemic to New Guinea where our samples were collected. For one of the bat sequences, we were not able to identify the exact species (as the most similar sequence in NCBI only had 94.4% identity) but our analyses revealed that it is likely closely related to the megabat Dobsonia praedatrix (1,916 sequence reads). This result also illustrates that, even if the actual host has not been sequenced for the locus of interest, our approach can still reveal its presence (and guide future studies to obtain more precise taxonomic information).
There are however some limitations to this approach. First, the primers may not allow the exact species to be identified: we estimated that 14% of the mammalian species do not have a unique DNA sequence at the locus amplified and the sequencing may therefore not enable differentiation among several closely related species. However, this limitation could easily be overcome by designing species-specific primers for a more variable region (e.g., the mt hypervariable region). Second, since we are comparing DNA sequences to the NCBI nt database there is the possibility of identifying incorrectly annotated sequences or pseudogenes, which could introduce spurious results. For example, one of the DNA sequences amplified that matched perfectly many pig DNA sequences (Sus scrofa, Sus barbatus, Sus philippensis, Sus celebensis and Sus verrucosus) was also identical to a thrip DNA sequence (Scolothrips takahashii). This instance likely represents a misannotation in NCBI but could be problematic without stringent quality controls. Similarly, several DNA sequences matched equally well human and gorilla, chimpanzee or orangutan DNA sequences and likely represent amplification of nuclear pseudogenes (numt) common in apes. Typically, these sequences were supported by a much lower number of reads (on average, 411) than DNA sequences that perfectly matched Homo sapiens mtDNA (on average represented by 70,405 reads) (S3 Table). Lastly, given the sensitivity of PCR and of the sequencing detection method, it is important that stringent controls are used to rule out human contamination. Here, we included 30 extraction (water) controls that were all negative suggesting very low levels of laboratory contamination (if any). An interesting complementary control, which would also control for field contamination, would be to analyze male mosquitoes collected at the same time.
Finally, our approach enables simultaneous processing of batches of 96 samples with minimum hands-on time (7–9 hours of laboratory work). This provides a unique throughput that is essential to analyze several hundred mosquitoes for well-powered comparisons. In addition, the high multiplexing of our approach dramatically reduces the cost of next-generation sequencing (to less than US$10 per sample), especially when combining the characterization of the blood meal composition with other data such as intra-species host characterization (see below), molecular species determination or genotyping.
Previous studies have used microsatellites to compare the attractiveness of different individuals or group of individuals [22, 24, 25], examine the blood feeding patterns of mosquitoes [20, 53, 54] or determine the effectiveness of insecticide treated bed nets [55–58]. DNA profiling with microsatellites allows for the identification of unique genetic profiles from human individuals fed on and can be a very powerful method to differentiate DNA from unrelated individuals. However, microsatellites can only detect the simultaneous presence of multiple individual DNAs (typically two) if their proportion in one sample is relatively similar. Otherwise, the signal from the less abundant DNA is typically obscured and not distinguishable from background noise. Rigorously identifying whether a disease vector feeds on a single or multiple individuals is however essential for disease control as vectors that feed on multiple individuals are more likely to rapidly spread the disease than those that only feed on a single individual.
As an alternative to microsatellites, our approach relies on identifying unique human mitochondrial haplotypes carried by a mosquito by analyzing 300 bp of the mt hypervariable region I. We showed that at least five (out of 102 mosquitoes analyzed) carried human mitochondrial DNA sequences from more than one person. It is important to emphasize here that the number of mixed human blood meals is clearly underestimated as only maternal lineages can be detected by this approach: all offspring will carry the same DNA sequence as their mother and therefore it would not be possible to distinguish between siblings (or cousins from mothers who are sisters). However, one could, at least partially, circumvent this limitation by including additional polymorphic nuclear loci in the assay and sequence them together with the mt hypervariable region locus (and the 16S rRNA). Overall, our approach allows for a rapid evaluation of the number of maternal lineages a mosquito has fed on that can be added to the characterization of the blood meal at no additional costs, and could be used to determine if mosquitoes preferentially feed on some individuals and avoid other individuals.
|
10.1371/journal.pgen.1003716 | Combining Quantitative Genetic Footprinting and Trait Enrichment Analysis to Identify Fitness Determinants of a Bacterial Pathogen | Strains of Extraintestinal Pathogenic Escherichia coli (ExPEC) exhibit an array of virulence strategies and are a major cause of urinary tract infections, sepsis and meningitis. Efforts to understand ExPEC pathogenesis are challenged by the high degree of genetic and phenotypic variation that exists among isolates. Determining which virulence traits are widespread and which are strain-specific will greatly benefit the design of more effective therapies. Towards this goal, we utilized a quantitative genetic footprinting technique known as transposon insertion sequencing (Tn-seq) in conjunction with comparative pathogenomics to functionally dissect the genetic repertoire of a reference ExPEC isolate. Using Tn-seq and high-throughput zebrafish infection models, we tracked changes in the abundance of ExPEC variants within saturated transposon mutant libraries following selection within distinct host niches. Nine hundred and seventy bacterial genes (18% of the genome) were found to promote pathogen fitness in either a niche-dependent or independent manner. To identify genes with the highest therapeutic and diagnostic potential, a novel Trait Enrichment Analysis (TEA) algorithm was developed to ascertain the phylogenetic distribution of candidate genes. TEA revealed that a significant portion of the 970 genes identified by Tn-seq have homologues more often contained within the genomes of ExPEC and other known pathogens, which, as suggested by the first axiom of molecular Koch's postulates, is considered to be a key feature of true virulence determinants. Three of these Tn-seq-derived pathogen-associated genes—a transcriptional repressor, a putative metalloendopeptidase toxin and a hypothetical DNA binding protein—were deleted and shown to independently affect ExPEC fitness in zebrafish and mouse models of infection. Together, the approaches and observations reported herein provide a resource for future pathogenomics-based research and highlight the diversity of factors required by a single ExPEC isolate to survive within varying host environments.
| Antibiotic resistance is an increasingly serious problem, especially among pathogenic strains of Escherichia coli that cause urinary tract infections, sepsis and meningitis. It is important to obtain a more comprehensive genome-wide understanding of bacterial virulence because it has the potential to uncover novel and alternative therapeutic targets. Therefore, we probed the genome of a pathogenic E. coli isolate using transposon mutagenesis, deep sequencing and comparative pathogenomics in an effort to define its virulence gene repertoire. Using this multilayered approach in combination with high-throughput zebrafish infection models, we identified hundreds of genes that affect pathogen fitness during localized and/or blood-borne infections. We also developed a bioinformatics-based method to systematically sift through our datasets for genes that are broadly conserved among an assortment of pathogenic species. Follow-up analysis of several pathogen-associated candidate genes using zebrafish and mouse infection models highlighted the capacity of our approach to identify novel fitness determinants. The results from this study are available via an interactive online data viewer (http://pathogenomics.path.utah.edu/F11_TnSeq/) so that investigators can more effectively search and utilize these findings.
| Within the Escherichia coli lineage there are several distinct virulent subgroups that are principally classified by an ability to cause a common set of diseases. One specific subgroup, Extraintestinal Pathogenic E. coli (ExPEC), is typically thought to be a benign inhabitant of the lower intestinal tract of warm-blooded vertebrates. However, outside this niche, ExPEC strains have the ability to persist in an array of secondary host-associated habitats where they can cause urinary tract infections (UTIs), meningitis, bacteremia and sepsis in both humans and domesticated animals [1], [2], [3], [4]. The combined medical, agricultural and economic burden of ExPEC-related diseases is likely to increase as antibiotic resistance spreads [1], [5], [6]. Delineation of the genetic elements utilized by ExPEC to infect such a diverse spectrum of host niches and cause disease promises to advance our understanding of pathogen evolution and behavior while also highlighting more effective strategies to combat these pervasive opportunistic pathogens.
Despite colonizing similar ecological niches, ExPEC isolates can differ by 20–30% of their respective gene inventories and to date, a specific and ubiquitous molecular trait or function exclusive to this cohort has not been characterized [7], [8]. Without knowledge of a unifying ExPEC-associated feature, the development of broad-spectrum therapeutic strategies is made exceptionally difficult [9]. Previous work indicates that individual ExPEC strains appear to have evolved distinct genetic repertoires that promote unique and at times subtle fitness advantages during colonization of specific host environments [10], [11]. It is becoming clear that the working definition of ExPEC is more multigenic in nature. Further complicating the search for ExPEC-associated traits is the 30–40% of genes within their genomes that still require functional annotation [7]. This problem will likely persist as genome sequencing continues to outstrip rates of experimental characterization. Consequently, there is a pressing need to develop methods that mesh high-throughput genomics with context-based functional observations—a notion that is becoming increasingly appreciated [12], [13]. To this end, we adapted a previously described transposon mutagenesis technique known as ‘Tn-seq’ to identify genes that promote pathogen fitness during colonization of a vertebrate host [14], [15].
The use of transposons to conduct unbiased forward genetic screens in bacteria has provided many avenues for investigation over the last two decades [16], [17], [18]. However, these approaches are, for the most part, time and labor intensive and often lack reliable quantitative metrics and sampling depth to identify genes of interest for retrospective follow-up. On the other hand, transposon mutagenesis as a tool is quite amenable to innovation [19], [20], [21]. Tn-seq, also known as ‘INSeq’, utilizes deep sequencing to monitor the composition of insertion variants within bulk mutant pools [14], [15]. This is accomplished through use of a modified mariner transposon that contains recognition sites within the two distal inverted repeats that are specific for the restriction enzyme MmeI. Cleavage of DNA by this enzyme occurs 18–20 base pairs from its recognition site, such that excision of the transposon by MmeI captures flanking genomic sequences. These sequences serve as tags that can be used to track the relative abundance of distinct insertion variants within heterogenetic mutant libraries. This technique ultimately allows the fitness of thousands of mutant variants to be quantified simultaneously following exposure to specific selective pressures. By comparing the composition and abundance of mutants before and after passage through selective conditions, such as within a host organism, a genome-wide map of loci that are important for bacterial fitness can be assembled.
The screening approach reported here involved injection of 48 h post-fertilization (hpf) zebrafish embryos with transposon-mutagenized pools of ExPEC derived from a single parent isolate. Embryonic zebrafish are particularly useful for this type of selection screen because their vertebrate physiology closely matches that of humans, and they have proven useful for the identification and characterization of virulence determinants that enable ExPEC to survive within host niches [10], [11], [22], [23]. Of particular importance, zebrafish embryos possess many of the same innate defenses that mammalian hosts rely upon to resist ExPEC (e.g. phagocytes, complement and antimicrobial peptides) [10], [24], [25], [26]. This attribute underscores the ability of the zebrafish host to mirror a set of overlapping selective pressures that ExPEC naturally encounter. Zebrafish embryos can be used to model both localized and systemic infections and the complete recovery and enumeration of bacteria from within the entire host is relatively straightforward [10], [11]. Through three independent biological experiments employing Tn-seq we identified 970 genes (∼18% of gene content) that promote the competitive fitness of ExPEC during either localized or systemic infections. A large number of these genes have unknown functions, while many others have reported roles in iron transport, bacterial secretion, two-component signaling and metabolism. To focus on genes that are likely important to ExPEC pathogenesis specifically, we devised a gene ontology-like method called ‘TEA’ (Trait Enrichment Analysis), that enabled us to organize candidate genes according to their association with certain bacterial lineages or groups of bacteria that are phenotypically similar. The data presented in this report indicate that TEA, in combination with Tn-seq, provides an effective, streamlined approach for identifying biologically relevant fitness and virulence determinants that are employed by ExPEC within various host environments. Datasets generated in this study have been made freely available through a curated and searchable web-based data viewer (http://pathogenomics.path.utah.edu/F11_TnSeq/).
We previously described the use of the embryonic zebrafish as a surrogate host to study genotype-phenotype relationships of a variety of human and non-human ExPEC isolates [10], [11]. During this initial work we observed that the human cystitis (bladder infection) isolate F11 is particularly adept at growing within and eliciting death of zebrafish during either localized or systemic infections. Elucidation of the virulence gene repertoire (i.e. genetic determinants that promote pathogenic behaviors such as colonization or destruction of host tissues) of this ExPEC strain using Tn-seq provides a starting point for future comparative pathogenomics studies. Figure 1 presents an outline detailing the negative selection screen carried out in this study.
Transposon mutagenesis was accomplished by using the previously described pSAM vector and conjugation [15]. This plasmid, which contains a mariner transposon flanked by MmeI modified inverted repeats and the himar1C9 transposase, was retrofitted with an E. coli compatible antibiotic resistance cassette and promoter elements to generate pSAM-Ec (Figure S1). Multiple, independent mating events were performed to assemble three transposon mutant libraries to be used in three separate replicate screens (Figure 1A). It was determined that each pool contained a minimum of 50,000 distinct insertion variants by enumerating selectable, transposon positive colonies immediately following conjugation when mutant siblings are minimal. Successful transposition of single inserts into individual bacterial chromosomes was verified by detection of the transposon by Southern blot analysis (Figure S2). Because the cost of a pre-screen sequence analysis was prohibitive, we utilized a more economical means to confirm the diversity and saturation of mutant pools. The frequency of variants defective for lactose utilization, curli production and glycogen storage was determined using standard colorimetric plating assays (Figure S2), which demonstrated that our transposon mutant libraries were sufficiently saturated and ready for in vivo screening. Post-hoc sequencing later confirmed that our libraries contained 60,000 to 80,000 mutant variants each, corresponding to an insertion event approximately every 75 bp within the F11 chromosome (Table S1).
For each independent biological experiment, a single mutant library was cultivated from a frozen stock in M9 minimal media overnight at 37°C (Figure 1B). During this initial outgrowth period, variants with insertions that disabled genes critical for normal replication within broth culture were likely reduced in abundance or completely eliminated. Prior to injection of 48 hpf zebrafish embryos, a 1 ml aliquot of each overnight culture was pelleted, washed and suspended in phosphate buffered saline (PBS) at the desired density. A second aliquot was pelleted and stored as a reference ‘input’ population for later sequencing and comparative analyses with ‘output’ populations. A dose of approximately 3,000 colony forming units (CFU) from each mutant library was injected into either the pericardial cavity (PC) or into the blood via the circulation valley (CV) of zebrafish embryos (200 embryos per niche per experiment were used to ensure adequate sampling of mutant variants) (Figure S3). Injection of the PC provides a model of localized tissue infection, whereas injection into the CV results in rapid dissemination of ExPEC via the bloodstream, modeling bacteremia or sepsis-like infections [10], [11]. We previously determined that a variety of ExPEC isolates have the capacity to grow unchecked within the PC, while only a subset are able to persist and multiply within the bloodstream. Presumably, the microenvironments encountered by ExPEC following injection into the zebrafish circulatory system are more challenging than the PC with respect to nutrient availability, host defenses or other as-yet-undefined factors [10], [11]. F11 is able to survive and replicate within both the PC and blood. Once injected, selection proceeded for 18 to 20 h—a timeframe in which the mutant F11 population grows to between 104 and 106 CFU/embryo and elicits death in ∼40–60% of the animals (Figure S4). Total genomic DNA of the surviving F11 variants was recovered from batch-homogenized fish (alive and dead) to produce output pools. In parallel, equal amounts of genomic material from input and corresponding output pools were digested using the MmeI restriction enzyme followed by enrichment of transposon-containing fragments with flanking genomic sequences via gel electrophoresis. The fragments were then excised and prepared for indexed sequencing on a single lane of an Illumina HiSeq 2000. General sequence-based features of each mutant pool, including number of detected mutant variants, sequencing depth, saturation and insertional bias, are summarized in Table S1.
To focus the list of Tn-seq-derived candidate genes for further study, we developed a qualitative ranking system using a customized algorithm referred to as Trait Enrichment Analysis (TEA). The underlying premise of TEA was inspired by molecular Koch's postulates, which are a widely acknowledged set of guidelines used by investigators to critically assess the contributions made by certain genetic elements to the virulent nature of pathogenic organisms [39]. In particular, TEA draws on the first axiom of these postulates, which states that, “The phenotype or property under investigation should be associated with pathogenic members of a genus or pathogenic strains of a species.” [40]. Thus, the primary goal of TEA is to identify correlations between bacterial genes and bacterial traits (i.e. genotype-phenotype relationships).
TEA was carried out as outlined in Figure 5. Briefly, protein sequences were acquired from an ecologically diverse, manually curated collection of 165 bacteria representing six phyla and assembled to form the TEA metaproteome database (TEA-MD, Table S3). Each bacterial strain was annotated with ‘trait categories’ defining habitat, niche of isolation, taxonomic lineage and phenotype (i.e. pathogen or non-pathogen). Homologues were retrieved for each of the 5,146 protein-coding genes contained within the F11 chromosome using the Basic Local Alignment Search Tool: BLASTp (Methods) [41]. Because proteins in the TEA-MD are linked to a set of traits defined by their source bacterium, homologue collections associated with F11 proteins can be appraised for the presence of a universal characteristic in a similar fashion to traditional gene ontology and KEGG enrichment analyses. As a result, hypotheses can then be generated regarding a protein's evolutionary origins (i.e. horizontal vs. vertical inheritance or homoplasy) and possible influence on bacterial physiology or behavior based on its combination of trait associations. A complete description of the TEA-MD, including the number of organisms and proteins comprising each trait category and the relative proportions of phyla compared to sequenced isolates in GenBank, is provided as supporting information (Figures S7 and S8).
To extend our experimental and sequence-based observations with the candidate genes EcF11_3256/bipA, EcF11_3082, EcF11_2628 and EcF11_3933, we employed a more traditional host organism for investigating ExPEC pathogenicity—the laboratory mouse [49], [50]. Evaluating Tn-seq and TEA-derived in piscis candidate genes using a murine model provided an opportunity to further establish the biological relevance of our multilayered approach. Murine models of bacteremia are an established means to assess the general capacity of ExPEC isolates to replicate and cause disease within a host [50], [51], [52]. One-to-one mixtures of wild type F11 and each mutant strain (108 total CFU) were subcutaneously injected into adult female Swiss Webster mice. Approximately 12 to 15 h later, bacterial titers within the spleen and liver were determined (Figure 7). Mutants lacking EcF11_3256/bipA (Figure 7A), EcF11_3082 (Figure 7B) and EcF11_3933 (Figure 7D) were significantly outcompeted by the wild type F11 parent strain. However, F11Δ3256 exhibited a more modest defect in the spleen (p = 0.0828). In contrast, there was no observable competitive difference between wild type and F11Δ2628 in either organ (Figure 7C).
Of note, we found that F11Δ3256 displayed a cold sensitivity phenotype when grown at 20°C in broth culture, which is consistent with reports on BipA function in other bacteria (Figure S9A) [53]. Therefore, we were initially concerned that the growth rate of this mutant at 28.5°C—the temperature used to raise zebrafish embryos—was compromising our ability to distinguish its temperature and host-dependent phenotypes. However, F11Δ3256, like the other mutants, grew similar to the wild type strain at both at 28.5°C and 37°C (Figure S9). We conclude that the attenuated fitness phenotypes observed with F11Δ3256 in competitive assays in both zebrafish and mice are independent of temperature effects.
We next tested the fitness contributions made by each of the four candidate genes above within the murine urinary tract, which is another commonly used model system for understanding ExPEC pathogenicity. Three days after inoculation of one-to-one mixtures of wild type and mutant derivatives (107 total CFU) into the urinary tract of adult female CBA/J mice, kidneys and bladders were harvested and bacterial titers determined. Of the four genes assayed, only EcF11_3933 was critical for the competitive fitness of F11 within the bladder and kidneys (Figure 7E).
Our ability to identify candidate genes necessary for ExPEC fitness within mammalian host niches through initial prescreening in embryonic zebrafish highlight the utility of this teleost host to model overlapping aspects of ExPEC biology. By comparing the specific host and niche-specific phenotypes of various ExPEC isolates and mutant strains in futures investigations, we will be able to gain a deeper understanding of the host-pathogen interactions engaged by ExPEC.
The genetic and phenotypic plasticity of ExPEC isolates, which is in part due to horizontal transfer of pathoadpative accessory genes, presumably enables these pathogens to traverse a wide range of environmental and host-associated niches [7], [8], [10], [11]. It will be important to decipher how ExPEC employ their genetic repertoires so that more specific and efficacious therapeutic strategies can be developed. Moreover, knowledge of key events that occur during the evolutionary assembly of pathogen genomes will likely be gleaned as genotype-phenotype relationships are mapped. To this end, we functionally dissected the contents of an ExPEC reference genome using the recently described transposon-based screening technique Tn-seq (Figure 1) [14], [15]. With mutant pools derived from the ExPEC reference isolate F11 and multi-niche zebrafish infection models, we found that ∼18% of the F11 genome significantly affected bacterial fitness under the pathogenic conditions tested. These fitness determinants can be grouped based on the niches where they were required, which in some instances correlated with particular functions like iron transport and amino acid metabolism (Figure 3). Gene set and gene-based contributions to fitness can be explored through an online and curated data viewer (http://pathogenomics.path.utah.edu/F11_TnSeq/). Going forward, it is our intent that this work will expedite future comparative and functional pathogenomic analyses focused on elucidating the genetic basis of pathogen behavior and evolution.
Seventy-six genes (1.4% of genome content) were found to significantly influence the fitness of F11 in both the PC and blood of zebrafish. We suggest that these multi-niche genes largely provide core functions that are critical, but not necessarily specific, to pathogen physiology. It was not surprising that several of these genes were recognizable and annotated as being involved in the regulation of gene expression and membrane architecture—two functions that are important to pathogens and non-pathogens alike (Table 1). In addition to the translational regulator EcF11_3256/BipA, which we validated in follow up experiments using a targeted deletion mutant and multiple in vivo host model systems (Figure 4 and 7A), several other informational genes, including hdeD, lrhA, rfaH and nhaR, were also identified by our Tn-seq screens as being important to the fitness of F11 within multiple niches. These genes have been shown to affect acid stress resistance, toxin expression, adhesion to host tissues and virulence gene regulation in a diverse array of pathogens, including enterohemorrhagic E. coli (O157:H7), Xenorhabdus nematophila, Proteus mirabilis and a sub-group of ExPEC isolates like F11 known as uropathogenic E. coli (UPEC) (Table 1) [31], [32], [54], [55], [56]. Additionally, genes responsible for the production of capsule—which is a well-established virulence feature that serves as a protective polysaccharide barrier surrounding the bacterial cell wall—were also important in both the PC and blood (Table 1) [57]. Further identification and, importantly, functional characterization of multi-niche genes will lead to a deeper understanding of the general fitness and virulence gene networks employed by ExPEC and other virulent bacteria within varying host environments.
Aside from multi-niche genes, we observed a striking numerical imbalance of genes that are exclusively required for either blood-borne or tissue-localized PC fitness (772 and 122 genes, respectively). The large number of genes and variety of enriched KEGG functional categories within the blood-specific gene set corroborated previous experimental observations that indicated the blood of embryonic zebrafish is a hostile environment and highly restrictive to bacterial growth [10], [11]. Specifically, genes involved in nutrient acquisition and utilization made up a significant portion of the blood-specific genes (Figure 3). These results provide evidence that during systemic infections nutrients are more limiting than during localized infection of the PC, requiring ExPEC to employ a greater portion of their genetic repertoire to maintain fitness in the blood. However, additional factors could also account for the discrepancy between the PC and blood-specific gene sets, including instances of trans-complementation. Once injected into the PC, bacteria remain confined until the embryo succumbs to the infection [10]. Because bacteria stay in close proximity to one another within the PC, the chances for trans-complementation and the sharing of common goods are high and may result in relaxed selection. A recent study demonstrated that extensive trans-complementation can occur within mixed mutant populations of Yersinia pestis in a mouse model of pneumonic plague [58]. However, the authors noted that this appears to be a unique characteristic of Y. pestis, as other pathogens with tropism for lung tissue do not exhibit such a high degree of wild type-to-mutant rescue. In our assays, trans-complementation effects within the PC were apparently limited as we were able to identify genes required for ExPEC survival within this niche.
Ultimately, this initial genome-wide analysis of the components necessary for ExPEC fitness within pathogenic environments garnered two key insights: (i) the number of genes that influence fitness in multiple niches is relatively small and (ii) there are distinct subsets of genes employed by ExPEC within different host environments. This suggests that do-it-all genes do not necessarily dictate the generalist and plastic nature of ExPEC. Instead, ExPEC may evolutionarily maintain sets of context-specific genes that can be employed, possibly in different combinations, to respond in a highly specific-manner to a variety of conditions. This characteristic likely contributes to the seemingly redundant composition and versatility of contemporary ExPEC genomes. Further functional dissection of ExPEC genomes by Tn-seq-based approaches will help clarify which genes underlie the salient traits of this lineage and inform the design of broad-spectrum therapies that target only ExPEC while leaving beneficial or benign E. coli strains within the normal microbiota unperturbed.
A longstanding challenge associated with unbiased, high-throughput functional genomics techniques like Tn-seq has been the need to develop methods that provide an effective transition to focused follow-up studies. Much work has been done to quantitatively filter datasets using sophisticated determinations of statistical significance and false discovery control procedures. Analysis of the molecular pathways represented within functional genomics gene sets has also been addressed extensively. However, less attention has been given to systematic methods that exploit the qualitative features of individual genes and how such considerations can streamline candidate gene vetting. To address this issue, we developed a Trait Enrichment Analysis (TEA) tool to organize Tn-seq-derived candidate genes in a biologically meaningful way (Figure 5). With a manually curated set of 165 microbial proteomes, TEA is able to assign meta-features to F11 genes based on their association with specific bacterial traits (i.e. habitat, niche of isolation, phylum and phenotype). In this way, TEA provides a tractable method to categorize Tn-seq candidate genes into a more focused subset—something not easily accomplished using larger microbial sequence databases like GenBank and BioCyc or functional annotation platforms such as DAVID. TEA uncovered several genes with previously unappreciated roles in ExPEC pathogenicity, thus creating new opportunities to further understand the genetic underpinnings of this lineage.
For the purposes of our investigation, we used TEA to identify F11 genes with homologues that are more often associated with the genomes of known pathogens. We describe such genes as having ‘pathogenic identity’ and reason that, because of their affiliation with pathogens, they are more likely to contribute to pathogenic behavior—a principal tenet of molecular Koch's postulates [39], [40]. In considering the overlap between Tn-seq-derived in piscis genes and the TEA-derived gene set made up of F11 genes with pathogenic identity, we were able to identify 246 genes that potentially influence the in vivo fitness capacity of F11 specifically. We selected three of these pathogen-associated candidate genes for closer inspection.
The candidate gene EcF11_3082 was identified by Tn-seq and confirmed during one-to-one co-challenge experiments to be important for the fitness of F11 within the blood and PC of zebrafish and spleen and liver of mice (Figure 6B and 7B). This gene appears to be relatively widespread within the TEA-MD, being found in 56 genomes, 61% of which are classified as pathogens. The range of bacteria encoding homologues of EcF11_3082 includes isolates from 12 (out of 16) different niches and 3 (out of 6) phyla represented in the TEA-MD (Figure 6A). These observations suggest that EcF11_3082 may play a broad role in bacterial fitness within a variety of pathogenic and non-pathogenic contexts. EcF11_3082 is predicted to encode an MprA/EmrR-like transcriptional repressor, which was shown to regulate the expression of a downstream multidrug efflux pump encoded by emrA and emrB. Although our Tn-seq selection screen and co-challenge experiments identified EcF11_3082 as a fitness determinant, disruption of the emrA and emrB genes present in F11 did not appreciably affect fitness within the zebrafish host, as determined by Tn-seq. This suggests that either disruption of EcF11_3082 led to aberrant and deleterious expression of emrAB that resulted in attenuation of the EcF11_3082 mutant, or the particular MprA/EmrR allele harbored by F11 influences other aspects of bacterial physiology not yet defined. Further investigation is required to determine the exact influence of EcF11_3082 on pathogenicity.
EcF11_2628, which encodes a ‘conserved hypothetical protein’ that has some homology to a secreted metalloendopeptidase toxin, was also identified by Tn-seq and confirmed as an important fitness determinant using a targeted knockout strain (Figure 6D). However, we were unable to ascertain if this gene is required during colonization of a mammalian host. The F11Δ2628 mutant variant did not exhibit any statistically significant decline in competitive fitness during systemic or urinary tract infection of mice (Figure 7C and data not shown). Further investigation is required to determine the physiological context in which this putative toxin promotes ExPEC fitness. Nonetheless, homologues of EcF11_2628 were found in only 7 proteobacterial genomes within the TEA-MD (Figure 6C and D). Of these, 5 are ExPEC strains that were isolated from human patients with either meningitis (strain S88) or UTI (strains CFT073, UTI89, 536 and UMN026). Another ExPEC strain (APEC01) harboring an EcF11_2628 homologue came from a case of avian sepsis that occurred subsequent to respiratory tract infection. The sole non-pathogen that carried an EcF11_2628 homologue was ED1a, a commensal isolate from the human gut that is classified as belonging to the B2 phylogenetic group—the E. coli lineage that encompasses the vast majority of human ExPEC strains [7]. This observation is intriguing because the EcF11_2628 allele carried by ED1a could represent either a genetic vestige from a pathogenic past or a new acquisition that is edging ED1a down an evolutionary path towards a more ExPEC-like countenance. Elucidating the function of this hypothetical gene should help explain its seemingly restricted distribution within a phylogenetically narrow range of bacteria.
The final pathogen-associated gene that we examined, EcF11_3933, encodes a variant of the DNA protection protein DprA [46], [47]. However, this particular dprA allele is distinct from those previously characterized in other bacteria. Within the TEA-MD, 83% of the EcF11_3933 homologues were found in the genomes of known pathogens. Interestingly, pathogens harboring an EcF11_3933-like gene included ExPEC isolates as well as other members of Proteobacteria and species within the phyla Bacteroidetes and Firmicutes (Figure 6E). In our Tn-seq selection screens, EcF11_3933 was initially identified because it appeared to affect fitness within the PC, but subsequent analyses revealed that EcF11_3933 could also promote pathogen fitness within the bloodstream of both zebrafish and mice (Figure 6F and 7D). Indeed, there were indications that EcF11_3933 contributed to blood-borne fitness, but it was filtered out of the final Tn-seq-derived blood gene set because it did not meet all of the predetermined cutoffs, though only by a thin margin. These results identify limits to Tn-seq as it is applied here, but also highlight the value of using multiple infection models to discover and assess the roles of fitness and virulence determinants.
Additional work using a murine UTI infection model demonstrated that EcF11_3933 is also important for competitive fitness within the murine bladder and kidneys (Figure 7E). This is in contrast to EcF11_3256, EcF11_3082 and EcF11_2628, which we found to be dispensable for colonization of the urinary tract (data not shown). Interestingly, alleles of c4222, which is a homologue of EcF11_3933 harbored by the UPEC strain CFT073, were shown by microarray to be transcriptionally induced within a panel of UPEC isolates during active infection of the human urinary tract [59]. Together, previous observations and those presented here indicate that EcF11_3933 is of broad importance to the pathogenic potential of F11 and other ExPEC isolates within a variety host-associated niches. Moreover, the distribution of EcF11_3933-like genes among a diverse array of pathogens (e.g. ExPEC, the enteric pathogen E. coli O157:H7, Neisseria gonorrhoeae, Vibrio cholera, Haemophilus influenza, Photorhabdus luminescens and Erwinia amylovora), which can colonize and cause disease within a broad range of animal, insect and plant hosts, suggests that EcF11_3933 also represents a multi-lineage virulence determinant. Further analysis of EcF11_3933 and its pathogen-associated homologues may prove useful in defining common virulence strategies employed by ExPEC as well as other more distantly related bacterial species.
Cumulatively, this investigation demonstrates that the application of Tn-seq, coupled with TEA, provides an efficient means to identify novel or previously unappreciated fitness and virulence determinants. The utility of TEA could easily be enhanced through expansion of the TEA-MD by, for example, the addition of metadata obtained from other host-microbe experimental systems. TEA serves as an accelerated annotation layer for bridging high-throughput sequencing experiments and functional characterization. In moving forward, it will be important to continue the careful curation of pertinent biological information associated with the rapidly expanding and often redundant and ill-defined collections of microbial genomics data so that meaningful connections can be mined. Ultimately, the use of Tn-seq and TEA with an expanded number of bacterial isolates and host systems may provide a detailed map of the gene families and signaling cascades employed by ExPEC and other pathogens to colonize host-associated environments.
Animals used in this study were handled in accordance with University of Utah approved IACUC protocols that followed the standard guidelines described at www.zfin.org and in the Guide for the Care and Use of Laboratory Animals, 8th Edition.
All bacterial strains and plasmids used in this study are listed in Table S5. Unless specified otherwise, bacteria were cultured statically at 37°C for 24 h in 20 ml of a defined M9 minimal medium (6 g/L Na2HPO4, 3 g/L KH2PO4, 1 g/L NH4Cl, 0.5 g/L NaCl, 1 mM MgSO4, 0.1 mM CaCl2, 0.1% glucose, 0.0025% nicotinic acid, 0.2% casein amino acids and 16.5 mg/ml thiamine in H2O) prior to injection into zebrafish embryos or mice. Antibiotics (kanamycin or ampicillin) were added to the growth medium when necessary to maintain recombinant plasmids or select for mutants.
Targeted gene knockouts of Tn-seq-derived candidate genes were generated in ExPEC strain F11 using the lambda Red-mediated linear transformation system [60], [61]. Briefly, a kanamycin resistance gene was amplified using polymerase chain reaction (PCR) from the pKD4 plasmid with 40-base pair overhangs specific to the 5′ and 3′ ends of each targeted locus. PCR products were introduced via electroporation into F11 carrying pKM208, which encodes an IPTG (isopropyl-β-D-thiogalactopyranoside)-inducible lambda red recombinase. Knockouts were confirmed by PCR. Primer sets used are listed in Table S6.
Retrofitting of the previously constructed pSAM_Bt vector to be used for the transposon mutagenesis of ExPEC isolate F11 was accomplished using source vectors and primers as listed in Tables S5 and S6, respectively. The Plac promoter from pGFP-Mut3.1 (Clonetech), along with its associated ribosome binding sequence, was amplified via PCR. Engineered 5′ BamHI and 3′ NdeI restriction sites were used to sub-clone the resulting fragment into a BamHI/NdeI (New England Biolabs) double digested pSAM_Bt vector upstream of the himar1C9 transposase gene. The kanamycin resistance gene from pKD4 was amplified and ligated using the restriction sites MfeI and XbaI, replacing the erythromycin resistance gene ermG in pSAM_Bt. The resulting transposon mutagenesis vector, pSAM-Ec, was stored and propagated in the pir+ E. coli strain EcS17.
*AB wild-type zebrafish embryos were collected from a laboratory-breeding colony that was maintained on a 14 h/10 h light/dark cycle. Embryos were grown at 28.5°C in E3 medium (5 mM NaCl, 0.17 mM KCl, 0.4 mM CaCl2, 0.16 mM MgSO4) containing 0.000016% methylene blue as an anti-fungal agent.
One ml from a 24-h bacterial culture of each transposon mutant library or isogenic bacterial strain was pelleted, washed once with 1 ml sterile PBS (Hyclone) and re-suspended in approximately 1 ml PBS to obtain appropriate bacterial densities for microinjection. Prior to injection, 48 h post-fertilization (hpf) embryos were manually dechorionated, briefly anesthetized using 0.77 mM ethyl 3-aminobenzoate methanesulfonate salt (tricaine) (Sigma-Aldrich) and embedded in 0.8% low-melt agarose (MO BIO Laboratories) without tricaine. Approximately 1 nl of bacteria was injected directly into the pericardial cavity (PC) or the blood via the circulation valley (CV) located ventral to the yolk sac using a YOU-1 micromanipulator (Narishige), a Narishige IM-200 microinjector and a JUN-AIR model 3-compressor setup. For each experiment, average colony forming units (CFU) introduced per injection were determined by adding 10 nl of the inoculum into 1 ml 0.7% NaCl that was then serially diluted and plated on Luria-Bertani (LB) agar plates. For co-challenge experiments, input doses were plated on LB agar +/− kanamycin (50 µg/mL) to determine relative numbers of the wild type and mutant strains present. After injection, embryos were carefully extracted from agar and placed in either a 10 cm petri dish containing E3 medium for transposon screening or, individually into wells of a 96-well microtiter plate (Nunc) containing E3 medium for co-challenge experiments. The E3 medium used during this incubation period lacked both tricaine and methylene blue.
To harvest DNA from embryos infected with transposon mutant libraries at the end of the selection period, all infected fish were collected (both alive and dead) in a 1.6 ml tube, being careful to remove as much E3 medium as possible by gently pulsing the tube in a microfuge to facilitate separation. Embryos were suspended in 500 µl of 0.5% Triton X-100 in PBS and homogenized using a mechanical PRO 250 homogenizer (PRO Scientific). Homogenates were then centrifuged at 18,000 g for 5 minutes to sediment bacteria away from host cellular components released during Triton X-100-mediated lysis. The resulting pellet, which appears black because of remaining zebrafish debris, served as the starting material for DNA extraction using the Wizard Genomic DNA Purification Kit (Promega) as per the manufacturer's protocol with a brief modification: the duration of the initial lysis step was extended as needed to dissolve the tough pellet.
To quantify bacterial numbers at the completion of co-challenge experiments, embryos were homogenized at the indicated time points in 500 µL PBS containing 0.5% Triton X-100 using a mechanical PRO 250 homogenizer. Homogenates were then serially diluted and plated on LB agar +/− kanamycin (50 µg/mL) to determine relative numbers of wild type and mutant bacteria. For co-challenge assays, only embryos where the total bacterial CFU recovered was equal to or greater than the limit of quantification (i.e. at least 20 CFU counted on the lowest serial dilution, which corresponds to 100 CFU per embryo) are reported. For lethality assays, fish were inspected for death at the indicated time points over a 72 h period. Death is defined here as the complete absence of heart rhythm and blood flow. Survival graphs depict total pooled results from at least 3 independent experiments in which groups of approximately 15 to 20 embryos were injected at a time.
For bacteremic co-challenge experiments, seven to eight-week old female Swiss Webster mice (Charles River) were anesthetized using isoflurane inhalation and injected subcutaneously via the nape of the neck with 200 µl of a 1∶1 wild type to mutant bacterial suspension containing a total of 108 bacteria in sterile PBS. Spleens and livers recovered 12 to 15 h later were weighed and homogenized in 500 µl sterile PBS. Homogenization was done with a Bullet Blender Storm 24 (Next Advance) using three SSB32 3.2 mm stainless steel beads for 5 minutes on power level 6. Homogenates were serially diluted and plated on LB agar +/− kanamycin (50 µg/mL) to determine the number of both wild type and mutant bacteria. Bacteremic co-challenge experiments were repeated at least twice. For these assays, only organs where the total bacterial CFU recovered was equal to or greater than the limit of quantification (i.e. at least 20 CFU counted on the lowest serial dilution, which corresponds to approximately 200 to 600 CFU per g organ) are reported.
For co-challenge experiments within the murine urinary tract infection, seven- to nine-week old female CBA/J mice (Jackson Labs) mice were anesthetized using isoflurane inhalation and inoculated via transurethral catheterization with 50 µl of a 1∶1 wild type to mutant bacterial suspension containing a total of 107 bacteria suspended in PBS. Bladders and kidneys were recovered 3 days later and each was weighed and homogenized in 1 ml containing 0.025% Triton X-100. Homogenates were serially diluted and plated on LB agar +/− kanamycin (50 µg/mL) to determine number of both wild type and mutant bacteria.
For co-challenge experiments, numbers of wild type and mutant bacteria present in the inoculum and recovered from host tissues were enumerated by differential plating on selective media as described above. To determine if bacterial counts differed between wild type and mutant-infected animals, a paired t-test was performed on log10-transformed values. Graphs and statistics were generated using GraphPad Prism 5.
|
10.1371/journal.pcbi.1002129 | A Systems Biology Approach Identifies Molecular Networks Defining Skeletal Muscle Abnormalities in Chronic Obstructive Pulmonary Disease | Chronic Obstructive Pulmonary Disease (COPD) is an inflammatory process of the lung inducing persistent airflow limitation. Extensive systemic effects, such as skeletal muscle dysfunction, often characterize these patients and severely limit life expectancy. Despite considerable research efforts, the molecular basis of muscle degeneration in COPD is still a matter of intense debate. In this study, we have applied a network biology approach to model the relationship between muscle molecular and physiological response to training and systemic inflammatory mediators. Our model shows that failure to co-ordinately activate expression of several tissue remodelling and bioenergetics pathways is a specific landmark of COPD diseased muscles. Our findings also suggest that this phenomenon may be linked to an abnormal expression of a number of histone modifiers, which we discovered correlate with oxygen utilization. These observations raised the interesting possibility that cell hypoxia may be a key factor driving skeletal muscle degeneration in COPD patients.
| Chronic Obstructive Pulmonary Disease (COPD) is a major life threatening disease of the lungs, characterized by airflow limitation and chronic inflammation. Progressive reduction of the body muscle mass is a condition linked to COPD that significantly decreases quality of life and survival. Physical exercise has been proposed as a therapeutic option but its utility is still a matter of debate. The mechanisms underlying muscle wasting are also still largely unknown. The results presented in this paper show that diseased muscles are largely unable to coordinate the expression of muscle remodelling and bioenergetics pathways and that the cause of this phenomena may be tissue hypoxia. These findings contrast with current hypotheses based on the role of chronic inflammation and show that a mechanism based on an oxygen driven, epigenetic control of these two important functions may be an important disease mechanism.
| Chronic Obstructive Pulmonary Disease (COPD) is an inflammatory process of the lung that generates progressive and largely poorly reversible airflow limitation [1], [2]. COPD represents a high burden on healthcare systems worldwide, since it is the fourth cause of death and its prevalence is expected to increase in forthcoming years [3].
The disease is primarily caused by the interplay between inhaled irritants, most frequently tobacco smoking but also environmental pollutants, and influenced by genetic susceptibility [4]. In these patients, the disease results in shortness of breath and contributes to limitation of exercise tolerance, leading to a decrease in daily physical activities [5], [6]. The latter has a significant deleterious impact on both clinical outcomes and prognosis [7], [8], [9]. Rehabilitation programs including skeletal muscle training and promotion of active lifestyles are recommended by all international clinical guidelines [4] as pivotal elements in the therapeutic strategies for COPD, but they are insufficiently deployed. One of the reasons for this is that in a significant percentage of patients, skeletal muscle dysfunction and muscle wasting are hallmark systemic effects of COPD [10]. Possibly linked with this, exercise-induced oxidative stress in COPD muscles is well documented and is likely to be an important mechanism driving tissue degeneration [11], [12], [13].
The role of systemic inflammation and myogenesis in skeletal muscle wasting are still a matter of controversy. However, recent studies have shown a reduction in the expression of myogenic genes in COPD muscles [14], [15] and a reduction in the ability to induce their expression in response to training in cachectic COPD patients [16], providing evidence for a deficiency in tissue remodelling. It has been proposed that lack of activation of myogenic pathways may be the result of the over-activation of the NF-kB pathway induced by systemic inflammatory signals generated by the lung [14]. This hypothesis is supported by cell culture and animal experiments, but so far there has been little evidence that this mechanism is clinically relevant [15]. Analysing a panel of muscle biopsies from normal and COPD individuals these authors showed that COPD muscles may be unable to activate NF-kB targets in response to physical training. Therefore, the mechanisms leading to skeletal muscle abnormalities in COPD and the relationship between muscle remodelling and oxidative damage are still a matter of intense research, and the extensive literature in this field is unable to support the development of effective predictive/preventive strategies.
The complex interplay of molecular pathways that are potentially involved in regulating muscle functionality makes a systems biology approach a desirable option. Such an approach aims to model the relationship between key molecular and physiological variables in healthy and diseased individuals to derive a testable hypothesis on the disease mechanism.
In this study, we hypothesized that skeletal muscle abnormalities in COPD may be the result of an imbalance in the physiological regulation of normal muscle homeostasis induced by systemic inflammatory mediators and chronic tissue hypoxia. In addition, we assume that the nature of such alteration might be reverse engineered by observing the statistical relationship between variables defining whole body physiology, systemic inflammation and muscle transcriptional state, with particular reference to cell bioenergetics and tissue remodelling functions. We based our analysis on a clinical study representing 12 healthy subjects and 18 age and sex-matched COPD patients, before and after undergoing an 8 weeks training program. In the latter category we included patients with preserved muscle mass (COPDN) and patients showing muscle wasting (COPDL).
We discovered that COPD muscles are characterized by a lack of correlation in expression of bioenergetic and tissue remodelling pathways, which included genes specifically involved in myogenesis. Our analysis suggests that failure to activate and coordinate these functions in response to training is associated with a general lack of activation of NF-kB targets, including many pro-inflammatory signals (e.g. IL-1β). We also discovered that expression of chromatin modification enzymes, known to control muscle differentiation and energy balance in other biological systems, is abnormal in COPD muscles and correlated with oxygen availability. This finding raises the possibility that an epigenetic mechanism triggered by tissue hypoxia may be the basis of skeletal muscle wasting in COPD.
All animal work has been conducted according to relevant national and international guidelines and approved by the University of Birmingham, Medical School ethics committee.
The 8 weeks training project (TP) was a clinical investigation in which eighteen COPD patients (68±7 yrs, 17 men, FEV1 46±12% predicted and PaO2 75±0.7 mmHg) with a wide spectrum of body mass composition and twelve age-matched healthy sedentary controls (65 yrs, 10 men, FEV1 107±14% predicted and PaO2 93±0.7 mmHg) underwent a protocol of supervised endurance exercise. The inclusion criteria were: 1) diagnosis of COPD according to GOLD criteria, 2) a stable clinical condition using standard treatment with bronchodilators and inhaled corticosteroids, 3) absence of episodes of exacerbation or oral steroid treatment in the previous 4 months, 4) absence of significant co-morbidities. All procedures were performed in the Pulmonary Function Laboratory or the Rehabilitation Unit at the Hospital Clinic-IDIPAPS. The TP included twelve patients with stable COPD and normal fat free mass index (FFMI, 21 Kg/m2) (COPDN), six COPD patients with low FFMI (16 Kg/m2) (COPL), and twelve healthy sedentary subjects (FFMI 21 Kg/m2). The study was approved by the Ethics Committee of the Hospital Clinic (Barcelona, Spain) and all patients gave written informed consent. Subjects completed all phases of the protocol as well as fully contributing to the sampling regime.
Significant physiological training effects were obtained using a standard supervised interval training program. Table S1 and Figure S1A in Text S1 summarize the characteristics of the study groups and the training-induced effects on physiological variables. Constant-work rate exercise at 70% of pre-training Watts peak (Wpeak) (CardiO2 cycle Medical Graphics Corporation, USA) was carried out before and after 8-weeks training with cycloergometer, until pre-training endurance time exhaustion. Measurements before and after training were obtained at isowork-rate and iso-time.
Serum samples obtained at rest before training (Basal-BT) and at rest after the eight weeks training program (Basal-AT) were analysed by Luminex xMAP technology, according to the manufacturer's instructions BIO-RAD [17]. A number of cytokines/growth factors were selected for analysis based on the results of a previous proteomic study carried out in COPD patients and controls [18]. In that study [18] a total of 142 cytokines/growth factors were analysed and 42 of them were differentially expressed in COPD. Our selection includes 30 analytes covering ∼50% of the cytokines/growth factors previously identified [18], which demonstrated association to clinical parameters in the same study. The complete list of proteins measured and the corresponding results of the analysis is shown in Table S2 and Figure S1B in Text S1.
Skeletal muscle transcriptomics was performed on open biopsies from the m. vastus lateralis (quadriceps). In all participants these were obtained at rest, before and after training. RNA was isolated using RNeasy extraction kits (QUIAGEN, USA) according to the manufacturer instructions. Microarray gene expression analysis employing Affymetrix ® GeneChip technology was performed using Human U133 Plus2 Gene Chips according to the manufacturer's suggested protocols. Data were subsequently subject to quality control to assess integrity of the RNA using an Agilent Bioanalyzer (Agilent). These datasets were normalised using the R library gcrma, which converts CEL files using a robust multi-array average (RMA) expression measure with the help of probe sequences. Data were quantile normalised [19].
In order to identify differential changes in physiological and protein measurements we used two-factor ANOVA with disease and training as factors. Significantly different variables were selected using a threshold of P<0.01. Similarly, genes differentially expressed between sedentary and trained subjects in the three populations (healthy, COPDN and COPDL patients) were identified by t-test followed by Benjamimi-Hochberg multiple correction [20] using a false discovery rate (FDR) threshold of q<10%.
To assess the effects of pro-inflammatory cytokines on skeletal muscle, recombinant mouse IL-1β (10 mg/ml) was injected in a single dose into the tail vein of C57BL10 mice at a loading of 100 ng/mouse (4 animals per group) and samples taken from lateral gastrocnemius (glycolytic) and soleus (oxidative) muscles 24 h later, with saline-injected animals acting as controls. Samples were stored at −80°C until use. RNA was extracted using an RNeasy RNA extraction kit (QUIAGEN). Microarray expression profiling of four independent biological replicates was performed using full genome oligonucleotide arrays (OPERON) after labelling the mRNA with the Cy-Scribe post labelling kit (Amersham) according to the manufacturer's instructions. Genes differentially expressed were identified by a t-test followed by multiple testing correction (FDR<10%). Significant genes from the mouse experiment were mapped on the human COPD network by converting mouse gene identities into their human homologues using the database Homologene (data was normalised using the same procedures described for the COPD dataset).
The gene set represented by those directly connected to growth factor and cytokine receptor components of the network and differentially expressed in response to training in healthy individuals were analyzed using the Ingenuity Pathway Analysis (IPA) application (Palo Alto, http://www.ingenuity.com), a web-based application that enables discovery, visualization, and exploration of biological interaction networks. Once the gene list was uploaded to the application, each gene identifier was mapped to its corresponding gene object in the Ingenuity Pathways Knowledge Base. These genes, called focus genes, were overlaid onto a global molecular network developed from information contained in the Ingenuity Pathways Knowledge Base. Networks of these focus genes were then algorithmically generated based on their connectivity according to the following procedure implemented in the IPA software application. The specificity of connection for each focus gene was calculated by the percentage of its connection to other focus genes. The initiation and the growth of pathways proceed from the gene with the highest specificity of connections. Each network had a maximum of 35 genes for easier interpretation and visual inspection. Pathways of highly interconnected genes were identified by statistical likelihood. Networks with a Score greater than 20 and containing more than 60% of focus genes were selected for biological interpretation. Canonical pathway and functional term enrichment analysis was performed using the IPA tools, and significance for the enrichment of the genes within a particular Canonical Pathway was determined by right- tailed Fisher's exact test with α = 0.01 using the whole database as a reference set.
In order to assess whether genes correlated to VO2max in the clinical study are indeed modulated in response to hypoxia we have analysed a publicly available microarray dataset representing the transcriptional response of C57Bl/10 mice to 2 weeks of chronic hypoxia and compared those with matching controls kept in normoxic conditions. The dataset was developed by Budak et al. [31], and it is available in the Gene Expression Omnibus database under the accession number GSE9400. The details of the overall design and the dataset can be found in GEO.
For clarity, here we report a concise summary of the experimental design and data generation. Adult male C57Bl/10 mice were divided into two groups, control (room condition) and hypoxic, over a period of 2 weeks. The hypoxic group was gradually exposed to lower levels of hypoxia in a specially designed and hermetically closed hypoxic chamber. A Pegas 4000 MF (Columbus Instruments) gas blending system was used. The oxygen level was gradually decreased from 21% to 8% over one week and animals were kept at 8% oxygen for another 7 days. After two weeks animals were euthanized using CO2 and total RNA was isolated from each quadriceps femoris (QF) muscle by Tri-Reagent (Ambion, Austin, TX). Microarray analysis was then performed using an Affymetrix Mouse Genome 430 2.0 Array according to the standard Affymetrix protocol (Affymetrix, Santa Clara, CA). Data were then quantile normalized using the R library gcrma [19]. Differentially expressed genes between normoxic and hypoxic mice were identified using SAM analysis [32]. Genes that have an FDR smaller than 10% were used to map to the neighborhood of the VO2max in the COPD network. Genes linked to the GO term “ chromatin remodeling” (represented in Figure S8 in Text S1) were identified using the web-based tool DAVID [26].
With the aim of developing an initial high level model representing the relationships between systemic and local signals we set to integrate all measurements available (physiology, blood cytokine levels and muscle gene expression measurements) for the individuals recruited in the 8 weeks training study. These samples included both pre and post-training blood and biopsy samples.
We then tested the individual network modules generated using the hubs in Table S3 in Text S1 for enrichment in GO and KEGG functional terms, which revealed a remarkably coherent association between target genes and the biological functions of network hubs. Neighbourhoods of physiological variables, such as VO2 peak, were mainly enriched in glycolysis/gluconeogenesis, mitochondria respiratory chain, oxidative phosphorylation (all positively correlated to VO2 peak), and components of the ribonucleoprotein complex and RNA processing (negatively correlated to VO2 peak, Figure S2 in Text S1 and Table S4 in Text S1). Genes associated with mRNA expression of cytokine and growth factor receptor hubs were enriched in functions related to tissue remodelling and immune response (generally positively correlated with the expression of receptors for growth factors and cytokines, Table S5 in Text S1), while genes associated with expression of glycolysis/gluconeogenesis enzymes were mainly enriched in energy metabolism-related functions (Table S6 in Text S1). Systemic inflammatory cytokines were weakly connected to the network (only three significant correlations with genes linked to physiological measurements).
The union of the individual modules and consequent visualization with a force driven layout revealed a clear separation between two main sub-networks. First, a large sub-network, primarily representing the association of modules built from cytokine (both plasma and skeletal muscle) and growth factor receptor hubs and enriched by tissue remodelling functions. Second, a sub-network representing the association of modules built with physiology and skeletal muscle glycolysis/gluconeogenesis hubs and enriched in energy metabolism genes (Figure 1). Within this sub-network we found that many of the physiological variables (PaO2, HRPEAK, HRPEAKper, WATTS, VE, VO2max, VO2maxper, VO2maxkg, BODE) were highly correlated (Figure S3 in Text S1) and consequently clustered together (Figure S4 in Text S1) to form a very compact network module. The other physiological variables (VEper, FFMI, BMI and age) were separated from this module and were characterized by a limited number of connections with gene expression variables (30 genes) (Figure S4 in Text S1).
In order to assess whether the separation between tissue remodelling and bioenergetics sub-networks observed in the network constructed by integrating all measurements in the training study may be a specific feature of COPD, it was necessary to developed two separate networks representing healthy and diseased muscles (Figure 2).
Our analysis, further supported by the integration of an additional dataset representing an independent and comparable clinical study (see Methods section for details), confirmed that the separation between cytokine/growth factor receptors and bioenergetics modules is a specific feature of COPD networks (Figure 2A, 2B) (Figure S5 in Text S1). In contrast, there was extensive overlap between the neighbourhood of receptors and intermediate metabolism in the network constructed from healthy muscle biopsies (Figure S2C, S2D and Figure S3 in Text S1).
In order to assess whether uncoupling was a general feature of pathological muscles, network models representing the transcriptional state of dystrophy (Figure S6A and S6B in Text S1) and diabetic muscles (Figure S6C and S6D in Text S1) were also developed. In both cases the cytokine/growth factor receptor and bioenergetics sub-networks were localized in close proximity. It was remarkable that even in a situation where muscle functionality is severely impaired by the lack of important structural components, a condition typical of muscle dystrophy, energy and tissue remodelling functions still remain co-ordinately regulated. This observation further strengthens the concept that uncoupling between muscle remodelling pathways and bioenergetics is a specific feature of COPD muscles.
Healthy individuals responded to training by modulating expression of a relatively large number of genes (3908, Figure S7A in Text S1) compared to COPDN (953, Figure S7B in Text S1) and to COPDL patients (6 genes). There was a significant overlap (36%, FDR<10%) at the gene level and an almost complete overlap at the pathway level (Figure S7 in Text S1) between the response to training in COPDN and healthy subjects whereas none of the six genes differentially regulated in the COPDL group were modulated in healthy and COPDN individuals (Figure S7C in Text S1).
In order to identify which part of the network was modulated in response to training, these differentially expressed genes were mapped onto the integrated network described previously (Figure 1). Genes that were up-regulated after training appear to cluster primarily in proximity to receptors for the cytokines IL-1 (IL1R), CSF1 (CSF1R), plasminogen (ENO1) and the key receptor for the pro-angiogenesis growth factor VEGF (VEGFR2) (Figure 3A). A second cluster that is also predominantly populated by genes up-regulated in response to training represents the neighbourhoods of glycolytic enzymes (Figure 3A) and to the epidermal growth factor receptor (EGFR). In contrast, genes that are either up or down-regulated in response to training populated the neighbourhood of physiological variables, representing energy (up regulated) and mRNA processing (down-regulated) functions. The transcriptional response of COPDN muscles to training followed a very similar pattern although as reported in Figure S5C in Text S1, the number of genes involved was much lower (Figure 3B). As mentioned previously, COPDL patients do not respond to training with a detectable change in the transcriptional state of the muscle (Figure 3C). In addition to the transcriptional uncoupling described above, the transition between healthy and diseased individuals is therefore characterised by a reduction in the ability to transcriptionally regulate expression of genes in both tissue remodelling and bioenergetics pathways.
To better characterize molecular networks associated with the growth factor and cytokines components of the network (see Figure 1 and Figure 2) directly connected to receptor hubs and differentially expressed in response to training in healthy individuals (predominantly located in the lower part of the network), we selected and used these genes as input for the Ingenuity Pathway Analysis (IPA) software. The analysis identified 17 networks (see method section for details of the procedure) (Table S7 in Text S1). Among the most significant findings, two interconnected networks linked the transcription of several cytokines (IL1, TNFα, IFNγ, CCL2) and activation of the NF-kB complex (Figure 4A), leading to the activation of several NF-kB targets related to connective tissue formation (Figure 4B). Figure 4C shows the relationship between upregulation of gap junction complexes (JAM, JAM2, JAM3 and TJP1) and activation of important structural components of muscle fibres (e.g. tropomyosin). The network in Figure 4D shows the important functional link between activation of several Rho GTPases and muscle development, represented by genes encoding for a component of the hexameric ATPase cellular motor protein myosin (MYL5) and (MYH10).
Although not all captured by our network model, 6 myogenic markers were found to be regulated in response to training (Table S8 in Text S1). Five of these were up-regulated. Consistent with what was previously observed (above), the training-induced expression of all the genes represented in these networks was defective in COPD individuals (Figure S4 and Table S8 in Text S1).
Because of the strong correlation between expression of IL-1 receptor and genes involved in muscle remodelling and the networks identified by the Ingenuity analysis, we reasoned IL-1 might be responsible for inducing a significant component of the transcriptional response to training observed in healthy individuals. In order to test this hypothesis we performed intra-venous injections of IL-1β in mice, and characterised the transcriptional response in both glycolytic and oxidative muscles using microarray expression profiling.
The response of mice to injection of IL1β supported our initial hypothesis. IL-1β treatment induced within 24 hours the up-regulation of 336 genes in oxidative muscle, whereas in glycolytic muscles it induced the up-regulation of 263 genes and the down-regulation of 201 genes (Table 1). Genes up-regulated in both muscle fibre types were enriched for structural components of the sarcomere (ACTA1, NEB, CRYAB, ANKRD2). Apart from this, the response of the two muscle fibre types was very different. Oxidative muscles up-regulated genes encoding for components of the extracellular matrix and organ development including angiogenesis (SOX18, VEGFA, SERPINF1, ANXA2, PTEN, ENG), whereas glycolytic muscles modulated genes involved in oxidative phosphorylation (COX8A, COX5A, NDUFB8, NDUFB5, NDUFS4, NDUFV2) and ribosomal components (Ribosomal protein chains 10A,s5 and I37A, and mitochondrion ribosomal protein chains 35 and 45).
The high percentage of differentially expressed genes encoding for tissue remodelling functions (extracellular matrix components, organ structure development, cell communication) and bioenergetics (oxidative phorphorylation) which included the oxygen sensitive regulator HIF-1, suggested that IL-1β has the potential to reproduce a component of the physiological transcriptional response to training that is defective in COPD patients.
In order to assess this hypothesis, we mapped genes differentially expressed in the IL-1 β mouse experiment on the human interaction network represented in Figure 1. 78% and 68% (265 and 317 genes) of the genes modulated in response to IL-1β treatment in oxidative muscle and glycolytic muscle, respectively, in mice mapped on to the human dataset. Of these, 34 genes were up-regulated after endurance training in healthy individuals (Figure 5). A significant proportion of these genes were, in oxidative muscle, also directly connected with the receptor for Interleukin 1 in the integrated network model. This observation further validates the model and supports a role for IL-1β in the physiological response to training.
The lack of activation of tissue remodelling functions in COPD muscles, including the component linked to IL-1β signalling, may be a consequence of the inactivation of myogenic pathways due to over-activation of NF-kB signalling induced by chronic exposure of pro-inflammatory cytokines [14]. We reasoned that if this hypothesis were correct, we should observe over expression of a number of NF-kB target genes in diseased muscles.
We had previously observed that a number of NF-kB targets, identified by the IPA software in the ‘tissue remodelling’ section of the network, were up-regulated during training in healthy individuals, but failed to respond in COPD patients (Figure 4A). Although interesting, this observation was based on a loose definition of NF-kB targets, as many connections reported in the Ingenuity database are indirect. We therefore further tested the NF-kB over-activation hypothesis by analyzing expression of 94 experimentally validated targets of NF-kB (Table S9 in Text S1). Against the working hypothesis, we could not identify any NF-kB target gene differentially regulated between normal and diseased muscles. On the contrary, we could detect 13 genes that were differentially regulated in response to training in healthy individuals (all up regulated, 10%FDR and fold>1.5) but not in COPD individuals with normal or low BMI (Table 2). These observations clearly do not support the over-activation hypothesis. On the contrary, they suggest that in COPD muscles training-induced activation of NF-kB is repressed.
In order to explore an alternative hypothesis that may explain muscle wasting in COPD patients, we performed an unbiased analysis to identify functional pathways differentially modulated between healthy and diseased muscles. We therefore used two-factor ANOVA, taking into consideration both disease and training.
Consistent with the analysis described above, we identified tissue remodelling and energy-associated pathways as significantly differentially expressed for both factors analysed (Table S10 in Text S1). However, one additional category of 20 chromatin modification enzymes were also differentially expressed (Table 3). Among these we could identify four histone deacetylase enzymes, which are known to be particularly relevant for controlling expression of muscle differentiation (HDAC9 and HDAC4, SIRT2) and bioenergetics (SIRT3) related genes. Consistent with their potential role in COPD muscle wasting, expression of these genes was sufficient to discriminate between diseased and healthy muscles with great accuracy (90% of cross-validated prediction accuracy using a K-nearest neighbour - model) (Figure 6A). By mapping them on the network model described in Figure 1 we discovered that they localized close to VO2peak, (Figure 6B) suggesting that their transcription may be up-regulated by oxygen availability and/or oxidative capacity in the skeletal muscle.
In order to elucidate whether genes linked to VO2max in our network model are part of the physiological response to hypoxic conditions we analyzed a public domain microarray study developed by Budak et al. (GEO: GSE9400) [31] which represents the response of murine skeletal muscles to 2 weeks hypoxic conditions.
We discovered that 45% of genes connected to VO2max in the network model were transcriptionally regulated in the mouse model of hypoxia and that a striking 82% of these were regulated in the direction predicted by the network analysis (Figure 7). Among these, we found the genes encoding for the chromatin modifiers HDAC4 and SIRT3. HDAC4 is up-regulated, whereas SIRT3 is down-regulated in hypoxic mice compare to normal (Figure 7), which is consistent with the sign of the correlation with VO2max observed in the clinical study (Figure 6A and 6C). Interestingly, we also discovered that many other genes encoding for chromatin modifiers were differentially expressed in response to hypoxia (Figure S8 in Text S1).
The networks we have developed represent the first model linking molecular and physiology measurements in skeletal muscle of COPD patients. It provides convincing evidence that a failure to co-ordinately activate expression of several tissue remodelling and bioenergetic pathways is a specific landmark of diseased muscles. Moreover, our model is consistent with the view that the abnormal expression of a number of histone modifiers potentially regulated by oxygen availability may be responsible for alterations in both tissue remodelling and bioenergetic functions. This hypothesis has important implications as it places cell hypoxia and oxidative capacity as the main drivers for skeletal muscle abnormalities in COPD patients.
We identified several pathways which are transcriptionally regulated in healthy individuals in response to training. These involve the up-regulation of several tissue remodelling genes/pathways (Plasminogen receptor, VEGF, pro-inflammatory signals such as IL1) as well as modulation of energy and ribosome biogenesis functions. Without exception, the exercise-induced modulation of these pathways is severely impaired in COPD individuals.
The Ingenuity pathway analysis has revealed several detailed mechanisms associated with the inflammation and growth factor receptor component of the network we have inferred. Two of these networks (Figure 4A and 4C) are part of a larger network linking the effect of pro-inflammatory signals such as IL-1, TNFα and IFNγ to the training induced up-regulation of several components of the extracellular matrix in healthy subjects. Although circulating concentration of IL-1β are largely unaffected by exercise there is evidence of increased local IL-1β levels within skeletal muscle, likely in response to micro-injury of skeletal muscle with increased activity [27]. The in vivo experiment we have performed show that indeed a component of the transcriptional response observed in healthy individuals and defective in COPD patients may be mediated by interleukin-1β. Moreover, we found that mouse glycolytic and oxidative muscles respond similarly to IL-1β in respect to the up-regulation of structural components of the muscle but diverge in the regulation of genes involved in energy metabolism and ubiquitination (up-regulated in glycolytic muscle), or extracellular matrix and tissue remodelling including angiogenesis (up-regulated in oxidative muscles).
Patients with mild to moderate COPD have a greater proportion of fatigue-susceptible anaerobic (glycolytic Type II) relative to fatigue-resistant aerobic (oxidative Type I) fibres, suggesting a slow-to-fast transition. It has been proposed that <27% Type I and >29% Type IIx fibres offers a pathological threshold for COPD [33]. This is consistent with a shift towards a more glycolytic enzyme profile, and would contribute to an increase in skeletal muscle fatigability [34]. Although individual fibre phenotypes are well conserved among mammals virtually all human skeletal muscles are of mixed composition, hence mouse muscles with discrete metabolic profiles were used to identify differential responses to IL-1β according to metabolic type.
A mechanistic link between the activity of these tissue remodeling pathways and myogenesis is well supported by the current literature. For instance, the plasminogen receptor ENO1 has been demonstrated to be an important component of skeletal myogenesis by concentrating and enhancing plasmin generation of the cell surface [35]. Interestingly, ENO1 KO mice show severe defects in muscle regeneration following injury [31]. Components of the extracellular matrix, which are induced by signaling from many of the receptors present in the network are also mechanistically linked to myopathies. For example, Col6a1–deficient and Col15a1-deficient mice have a muscle phenotype that strongly resembles human myopathies [36], [37].
Inflammatory and chemo-attractant mediators are also known to be key factors in driving muscle remodelling in the normal physiological response to training [38] and in response to trauma [35]. In the latter, IL1β also promote phagocytosis of trauma-induced cellular debris by macrophages, which themselves can continue to secrete this cytokine up to 5 days post-injury [39]. CCR2 null mice with cardiotoxin induced injury has been shown to have a delayed angiogenesis and VEFG production compared to wild type mice with muscle fibre size increase observed only after restoration of tissue VEGF [40]. Other studies have also shown the crucial role of VEGF in angiogenesis and in muscle regeneration [41], [42].
There is also strong evidence that bioenergetics and tissue remodelling pathways are linked [43]. It has been shown that genes coding for components of collagen V and collagen VI play an important regulatory role in ECM maturation, where reduced expression promotes apoptosis, mitochondrial dysfunction and muscle degeneration [43]. Other studies have shown that bioenergetics knockouts, such as H6PD null mice is responsible for inducing severe skeletal myopathy by altering sarcoplasmic reticulum redox state [44].
A number of in vitro and in vivo studies have been used in the past to support the hypothesis that a systemic inflammation-driven mechanism leads to inactivation of myogenic pathways in COPD muscles. Our analysis is the first attempt to challenge this hypothesis using genome-wide data, and in a clinically relevant setting. We could not find evidence of over-expression of NF-kB target genes in COPD muscles. On the contrary, we saw that a subset of direct (Table 2) and indirect targets (Figure 4A) of NF-kB were up-regulated in healthy individuals in response to training but not in COPD patients (p<0.05). These results are consistent with a recent observation demonstrating failure to activate NF-kB in response to acute training in a small subset of COPD patients [15]. Taken together, these results suggest that training-associated inactivation rather than over activation of NF-kB may be a feature of diseased muscles.
Additional observations also argue against a primary role of chronic inflammation in muscle wasting. For example, although we clearly have identified increased levels of cytokines in COPD patients (Figure S1B in Text S1) with respect to normal individuals, we could find little correlation between the concentration of these cytokines and the muscle transcriptional state, implying that these signals may have a smaller effect on muscle physiology than previously thought. This latest observation is consistent with previous reports showing that TNFα levels measured in COPD muscles were not significantly higher than in healthy muscles [45].
Since the NF-kB over-activation hypothesis is unlikely to explain the inhibition of muscle remodelling observed in COPD, can we propose an alternative mechanism for muscle wasting in COPD?
There are several pieces of evidence in favour of the hypothesis that an imbalance in expression of oxygen-correlated chromatin modifying enzymes, which we have shown to be a landmark of COPD muscles, could explain failure to modulate both tissue remodelling and bioenergetic functions in response to training. In Figure 6 we have shown that the expression of SIRT2, SIRT3, HDAC4 and HDAC9 is sufficient to discriminate healthy and diseased muscles. At the individual gene level, healthy muscles are characterized by a higher expression of HDAC9, SIRT3 and by a lower expression of HDAC4 (Table 4) with respect to COPD muscles.
The role of HDAC9 and HDAC4 in muscle development is well documented [46]. For example, HDAC9 is a transcriptional repressor involved in feedback control of muscle differentiation, acting in concert with MEF2 to repress activity-induced genes [47], while HDAC4 is up-regulated in pathological conditions such as muscle denervation [48], and it has been described to be a critical regulator of muscle atrophy by activation of E3 ubiquitin ligases [49]. It is possible, therefore, that a lower expression of HDAC9 and a higher expression of HDAC4 in COPD muscles may be linked to a reduced ability to activate muscle remodelling. Abnormal expression of SIRT2 may also contribute to a failure to activate an appropriate muscle remodelling response. SIRT2 is a NAD+-dependent histone deacetylase that regulates muscle gene expression and differentiation by forming a complex with MyoD [50]. When over-expressed, this retards muscle differentiation. Conversely, cells with decreased SIRT2 differentiate prematurely. Interestingly, the activity of SIRT2 is dependent on the redox state of the cell [50], which showed evidence of being abnormal in COPD muscles [11], [12], [13]. In the neighborhood of SIRT2, we have also identified TXN2 that is known to play an important role in protection against oxidative stress [51]. Similarly, GAB1 is shown to play a role in oxidative stress signaling [52] and it is identified in the neighborhood of HDAC9 in the network. Since changes in ROS production are known to influence the expression of HDACs [53], the network we have identified may represent this important control mechanism.
SIRT3 is a NAD+-dependent histone deacetylase that may account for the characteristic loss of transcriptional modulation of bioenergetic genes in response to training in COPD muscles. SIRT3 is localized in the mitochondrial matrix, where it regulates the acetylation levels of metabolic enzymes, including acetyl coenzyme A synthetase 2 [54], [55]. Mice lacking both SIRT3 alleles show hyperacetylation of several mitochondrial proteins, associated with decreased levels of fatty-acid oxidation, and display a selective inhibition of electron transport chain Complex I activity leading to reduction in basal levels of ATP in several organs [56]. These and other data implicate protein acetylation as an important regulator of mitochondrial function in vivo, and it is therefore feasible that an altered expression of SIRT3 in the muscles of COPD individuals may contribute to the observed imbalance in mitochondria functionality. It is interesting that SIRT3 in our model is positively correlated with VO2peak, suggesting that the lower levels of expression of this enzyme observed in COPD muscles may be the consequence of tissue hypoxia. Consistent with this, histone deacetylase (HDAC) inhibitors reduce HIF-1α protein expression leading to down-regulation of VEGF and other angiogenesis-related genes [57], potentially explaining the reciprocal relationship between extent of muscle capillarity and the degree of COPD [58]. The abnormal expression of a relatively small number of histone modifying enzymes could therefore account for a wide spectrum of abnormal responses observed in the muscles of COPD patients, and may also explain the limited efficacy of training as a therapeutic option.
This view is supported by our observation that indeed hypoxia induces modulation of a number of chromatin modifiers in a mouse model of chronic hypoxia (Figure S8 in Text S1) and that indeed SIRT3 and HDAC4 are among them (Figure 7).
Our work represents the most accurate system level representation of COPD muscles to date. Further work is, however, needed to elucidate the precise mechanism for muscle inactivation. If the mechanism for muscle wasting suggested by our observations on HDACs were to be validated in a clinical setting this would open up a very exciting therapeutic avenue. The use of non-toxic histone deacetylase inhibitors such as valproate has already shown promising in treatment of haematological cancer [59], [60], [61], and may help to restore mitochondrial functionality and the ability to activate muscle remodelling in COPD patients. In this context it is possible that the appropriate pharmacological regime coupled with physical rehabilitation may lead to recovered muscle functionality, and improved quality of life.
|
10.1371/journal.pgen.1007303 | The UBR-1 ubiquitin ligase regulates glutamate metabolism to generate coordinated motor pattern in Caenorhabditis elegans | UBR1 is an E3 ubiquitin ligase best known for its ability to target protein degradation by the N-end rule. The physiological functions of UBR family proteins, however, remain not fully understood. We found that the functional loss of C. elegans UBR-1 leads to a specific motor deficit: when adult animals generate reversal movements, A-class motor neurons exhibit synchronized activation, preventing body bending. This motor deficit is rescued by removing GOT-1, a transaminase that converts aspartate to glutamate. Both UBR-1 and GOT-1 are expressed and critically required in premotor interneurons of the reversal motor circuit to regulate the motor pattern. ubr-1 and got-1 mutants exhibit elevated and decreased glutamate level, respectively. These results raise an intriguing possibility that UBR proteins regulate glutamate metabolism, which is critical for neuronal development and signaling.
| Ubiquitin-mediated protein degradation is central to diverse biological processes. The selection of substrates for degradation is carried out by the E3 ubiquitin ligases, which target specific groups of proteins for ubiquitination. The human genome encodes hundreds of E3 ligases; many exhibit sequence conservation across animal species, including one such ligase called UBR1. Patients carrying mutations in UBR1 exhibit severe systemic defects, but the biology behinds UBR1’s physiological function remains elusive. Here we found that the C. elegans UBR-1 regulates glutamate level. When UBR-1 is defective, C. elegans exhibits increased glutamate; this leads to synchronization of motor neuron activity, hence defective locomotion when animals reach adulthood. UBR1-mediated glutamate metabolism may contribute to the physiological defects of UBR1 mutations.
| In eukaryotic cells, the ubiquitin-proteasome system mediates selective protein degradation [1, 2]. E3 ubiquitin ligases confer substrate specificity via selective interaction with the degradation signals in substrates [3–6].
UBR1 acts not only as an E3 ligase for the N-end rule substrates, whose metabolic stability is determined by the identity and post-translational status of their N-terminal moiety, but also for substrates that do not harbor the N-terminal degrons [6, 7]. The UBR family proteins exist from yeast to man, and have been implicated in multiple cellular processes (reviewed in 7). Yeast UBR1 is not essential, but ubr1 mutants exhibit less efficient chromatin separation, and mildly increased doubling time [8]. Yeast UBR1 also participates in protein quality control, potentiating the degradation of mis-folded proteins by ER membrane ligases [9]. The loss of C. elegans UBR-1 results in delayed degradation of a regulator for post-embryonic hypodermic cell division, but does not cause obvious hypodermic defect [10]. In mammalian cell lines, the N-end rule pathway targets pro-apoptosis fragments for degradation, affecting the efficacy of induced apoptosis [11]. The simultaneous loss of two rodent UBR homologues results in embryonic lethality with severe developmental defects in the heart and brain [12]. In human, loss-of-function mutations in one of several UBR family proteins, UBR1, cause the Johanson-Blizzard Syndrome (JBS), a genetic disorder with multi-systemic symptoms including pancreatic insufficiency, growth retardation, and cognitive impairments [13]. To date, a unifying physiological function of UBR proteins in animal models and human is lacking. In fact, whether UBR1’s role as an N-end rule E3 ligase is relevant for the JBS pathophysiology remains elusive [14].
C. elegans has a single UBR1 ortholog, UBR-1. Using this simplified animal model, we reveal that the functional loss of UBR-1 leads to a specific, late onset and prominent motor pattern change, and such a change is reversed by removing a metabolic enzyme, GOT-1, which we find to synthesize glutamate from aspartate.
Glutamate is an abundant amino acid. As a metabolite, it is essential for metabolism and development. As a neurotransmitter, glutamate-mediated signaling regulates animals’ motor and cognitive functions [15–18]. As such, glutamate level needs to be tightly regulated for metabolism, as well as neuronal signaling. Aberrant glutamate signaling has been implicated in cytotoxicity [19], and neurological disorders [20].
The genetic interaction between ubr-1 and got-1 mutants suggests that UBR-1 may affect glutamate metabolism. Consistent with this notion, we found that both genes are critically required in premotor interneurons to affect the reversal motor pattern change. Further, our metabolomics analyses reveal an inversely correlated change—increased and decreased glutamate level—in ubr-1 and got-1 mutant animals, respectively. These findings reveal a previously unknown role for the UBR family protein in glutamate metabolism. They further allude to the possibility that a common cellular defect, such as that in glutamate metabolism, may contribute to UBR’s multi-systemic functions.
Wildtype C. elegans generates movements through propagating body bends. In a genetic screen for mutants with altered motor patterns, we isolated hp684 (Fig 1A), a mutant that is capable of reversal movements, but does so with limited body flexing (Fig 1B; S1 Movie). The stiffness is prominent during prolonged reversals, and is progressive as animals develop from the last-stage larvae into adults.
We quantified the motor phenotypes of one-day-old wildtype and hp684 adults. Because the forward movement is the preferred motor state under laboratory conditions, we assayed their motor behaviors on plates where they were frequently induced for prolonged reversals (Methods). We compared their body curvature and duration of reversals during these events (Fig 1D–1F). hp684 mutants exhibited significantly decreased mean curvature during reversals (Fig 1D), significantly longer duration (Fig 1F), and reduced reversal initiation frequency (Fig 1G) under the assay conditions (Methods). The bending curvature and duration of forward movements, captured during the same recording periods, were only mildly affected or unchanged in hp684 mutant animals (S1 Movie).
We mapped and identified the causative genetic lesion in hp684 mutants (Methods): a recessive and nonsense mutation that leads to truncation of the last 194 amino acids (Q1864X) of UBR-1 (Fig 1A). The UBR family proteins exhibit conserved domain organization from yeast to humans. In addition to a highly conserved C-terminal sequences, they share the N-terminal UBR box, and internal motifs that include a region enriched for basic amino acids (BRR), and a RING finger [21]. The UBR box and its neighboring sequences interact with the substrates and E2 ubiquitin-conjugating enzyme of the N-end rule [22], whereas the RING finger is the hallmark motif utilized by a large class of E3 ligases to recruit non-N-end rule substrates [23–25].
To further verify that the motor phenotypes that we observed in hp684 mutants result from the functional loss of UBR-1, we generated multiple ubr-1 deletion alleles, hp820, hp821, and hp821hp833 (Fig 1A), by CRISPR-Cas9-mediated genome editing [26] (Methods). hp820 harbors a small in-frame deletion near the N-terminus (ΔR18-W25); hp821 a N-terminal four base pair deletion that leads to a frame-shift and a premature, N-terminal stop codon (E34X), and hp821hp833, in addition to hp821, a seven-base-pair deletion in the RING finger that leads to a frame shift and premature internal stop codon (E1315X). All alleles, like hp684, are recessive and viable, all exhibited motor defects similar to hp684, including reduced bending, increased duration, and reduced initiation frequency during reversal movements (Fig 1B and 1E–1G; S1 Fig). Also like hp684, their reversal defects were rescued by a genomic fragment that harbors only ubr-1 (Fig 1B–1G; S1 Fig).
These results confirm that the motor defects exhibited by all ubr-1 alleles result from the functional loss of UBR-1. The comparable phenotypic severity among hp684, hp820, hp821 and hp821hp833 mutants complements findings from the JBS patients, where diverse UBR1 mutations, resulting in a wide range of genetic lesions—early or late stop codons, reading frame shifts, and small internal in-frame deletions (Fig 1A)—exert similar pathologic effects [27].
Hereafter, we present quantified results from the hp684 allele unless specified, and refer to it as the ubr-1 mutants. For behavioral analyses, we present data for reversal movements, and refer to it as the motor phenotype.
To begin probing UBR-1’s function, we first examined the expression pattern of a GFP::UBR-1 reporter, which fully reversed ubr-1 mutants’ motor defects. GFP::UBR-1 exhibits strong expression in a fraction of neurons and all musculatures throughout post-embryonic development, and weak expression in hypodermal seam cells (Fig 2A). In the motor circuit, UBR-1::GFP’s expression is prominent in premotor interneurons (INs in Fig 2A) of the reversal motor circuit (Fig 2A; S2A Fig), and is absent from all ventral cord motor neurons that execute locomotion.
To determine the critical cellular origins of ubr-1 mutants’ motor defect, we examined the effect of restoring UBR-1 using exogenous and cell-type specific promoters. When UBR-1 expression was restored panneuronally (Prgef-1), the reversal motor defects (Fig 2B) in ubr-1 mutants were fully rescued, whereas restoring UBR-1 in muscles (Pmyo-3) did not (Fig 2B; S3A and S3B Fig). These results show a neuronal origin of ubr-1 mutants’ motor phenotype.
Through examining the effect of restoring UBR-1 expression in subgroups of motor circuit neurons that partially overlap with identified UBR-1::GFP-positive neurons, we confirmed a critical requirement of UBR-1 in premotor interneurons (Fig 2C; S3A and S3B Fig). Specifically, restoring UBR-1 expression by either Pglr-1 or Pnmr-1 significantly rescued ubr-1 mutants’ motor defects, including bending, duration, and frequency during reversal movements (Fig 2C; S3A and S3B Fig), whereas restoring UBR-1 in motor neurons did not (Fig 2C; S3A and S3B Fig). Both Pglr-1 and Pnmr-1 activate expression in premotor interneurons of the reversal motor circuit, AVA, AVE, AVD, and RIM [28–30]. Their activation, inactivation, and ablation affect the execution and characteristics of the reversal motor states [31–43]. Importantly, restoration of UBR-1 in other interneurons, including the premotor interneurons of the forward motor circuit (AVB), did not rescue ubr-1 mutants’ motor phenotype (Table 1).
Within this group of premotor interneurons of the reversal circuit, restoration of UBR-1 in AVE and RIM (Popt-3) exerted the strongest partial rescue (Fig 2D; S3C and S3D Fig), whereas the restoration in AVA (Prig-3), or RIM (Pgcy-13) alone, led to modest to no rescue (Fig 2D; S3C and S3D Fig). UBR-1’s role in the reversal motor circuit involves the whole network of premotor interneurons, with AVA, AVE, and maybe RIM being the most critical components.
What underlies the reduced bending in ubr-1 mutants? Premotor interneurons of the reversal motor circuit innervate the A-class motor neurons (A-MNs) (Fig 3A). Multiple A-MNs innervate body wall muscles to execute reversal movements: they are divided into the ventral (VA) and dorsal (DA) muscle-innervating subtypes through likely non-overlapping neuromuscular junctions [32, 34, 44, 45]. We examined the temporal activation pattern of a posterior cluster A-MNs, DA7, VA10, and VA11 (predicted muscle targets illustrated in Fig 3B) in freely moving adults that express a calcium sensor GCaMP6s::wCherry [46] in these A-MNs (Methods).
Consistent with the notion that A-MNs execute reversal movements, they exhibited calcium changes when animals moved backwards (boxed area in Fig 3C). While the frequency and amplitude of the calcium waveforms for A-MNs varied for each reversal event [47], their calcium profiles exhibited phase relations that are consisted with the expected temporal activation of muscle groups that they are predicted to innervate. Specifically, VA10 and VA11, two A-MNs that innervate adjacent ventral muscles exhibited asynchrony in activation (red and blue traces in Fig 3C; left panel in Fig 3C’), with variable lags (Fig 3G; Methods), as expected from the sequential contraction of adjacent muscles during reversals at different velocities. DA7 likely innervates dorsal muscles that appose VA10 and V11’s ventral targets (Fig 3B). Consistent with the alternating dorsal and ventral muscle contraction during bending, VA10 and DA7 exhibited asynchrony of activity patterns (green and red traces in Fig 3C; middle panel in Fig 3C’) with variable lags (Fig 3H). DA7 and VA11 were activated in relative synchrony (green and blue traces in Fig 3C; right panel in Fig 3C’), exhibiting shorter lags (Fig 3I) when compared to the other A-MN pairs. This indicates a more direct dorsal-ventral opposition between VA10 and DA7’s muscle targets.
We observed a striking difference in A-MN’s activation pattern in ubr-1 mutants. While they also exhibited calcium changes during reversals, all three A-MNs’ activation exhibited synchrony (blue, red and green traces in Fig 3D; Fig 3D’). This led to a drastic reduction in the mean lag times between VA10 and VA11 (Fig 3G), and between VA10 and DA7 (Fig 3H), whereas the short lags between DA7 and VA11 remained statistically unchanged (Fig 3I) between wildtype and ubr-1 mutants.
Hence, reduced bending in ubr-1 mutants was caused by increased synchronization, not lack of A-MNs’ activities. Importantly, when UBR-1 was restored in the premotor interneurons of the reversal circuit, A-MN’s phasic relationships were restored in ubr-1 mutants (Fig 3E and 3E’, Fig 3G–3I). Therefore UBR-1 plays a critical role in premotor interneurons to ensure sequential motor neuron activation, which underlies bending during reversal movements.
If UBR-1, an E3 ligase, affects the animal’s motor pattern through negative regulation of a biological pathway, the pattern change should be rescued by a simultaneous decrease of the activity of the pathway. Accordingly, we screened for genetic suppressors of ubr-1 mutants’ motor phenotype.
We isolated hp731, which restored bending in ubr-1 mutants during reversal movements (Fig 4C; S2 Movie). Notably, both hp731 and ubr-1; hp731 mutant animals exhibited slightly deeper bending than wildtype animals (Fig 4C; S2 Movie). hp731 also significantly rescued ubr-1 mutant’s change in reversal duration and frequency (S4A and S4B Fig). Consistent with synchronized A-MN activation underling ubr-1’s lack of bending, VA10, DA7 and VA11’s phasic relationships were restored in ubr-1; hp731 mutants (Fig 3F–3I).
We identified the causative mutation in hp731 mutant animals. hp731 harbors a causative, lf, and missense C184Y mutation (Methods) at the pyridoxal phosphate-binding domain in GOT-1.2 (Fig 4A), one of the four predicted C. elegans glutamate-oxaloacetate transaminases (GOTs). GOT enzymes catalyze the transfer of an amino group between aspartate and α-ketoglutarate (α-KG) and convert them to oxaloacetate (OAA) and glutamate [48, 49] (Fig 5A).
The genetic interaction between ubr-1 and got-1.2 is remarkably specific: lf mutations for the other three GOT homologues did not restore bending of ubr-1 mutants (Table 2). lf mutations in other metabolic enzymes or transporters that may be involved in glutamate and aspartate metabolic pathways (Table 2), including the alanine aminotransferase, glutamine synthetase, and glutaminase, did not rescue ubr-1 mutant’s motor defects either. Henceforth, we refer to got-1.2 as got-1.
To determine the endogenous expression pattern of GOT-1, we generated a functional GOT-1 reporter by inserting GFP at the endogenous got-1 locus. GOT-1::GFP exhibited cytoplasmic expression in all somatic tissues, including broad expression in the nervous system (Fig 4B; S2B Fig). To determine the critical cells that mediate the genetic interactions between ubr-1 and got-1, we restored cell-type specific GOT-1 expression in ubr-1; got-1 mutants, and assessed their effect on the animal’s motor pattern. Restoring GOT-1 in all neurons (Prgef-1), but not in all muscles (Pmyo-3), fully reverted the motor pattern of ubr-1; got-1 to the reduced bending as in ubr-1 mutants (Table 1); therefore, both UBR-1 and GOT-1 function through neurons to regulate motor patterns.
Because GOT-1 is more broadly expressed in the nervous system than the GFP::UBR-1 reporter (Fig 4B), we examined whether GOT-1 functions through UBR-1-expressing neurons to regulate bending. Similar to our observation for neuronal sub-type UBR-1 rescue (Fig 2), restoring GOT-1 in premotor interneurons of the reversal circuit (Table 1) was required for reversion of ubr-1; got-1’s bending pattern to that of ubr-1. Similarly, restoring GOT-1 in the same subset of these premotor interneurons, including AVE and RIM, exerted the most significant, partial reversion of ubr-1; got-1’s motor defects (Fig 4D; S4 Fig).
Therefore, not only does a predicted metabolic enzyme GOT-1 exhibit genetic interaction with UBR-1, but also both proteins exhibit similar prominent requirement in premotor interneurons to regulate the reversal motor pattern. These results raise the possibility that metabolic dys-regulation may underlie ubr-1’s motor defects.
The catalytic activity of GOT-family transaminases enables reversible conversions between aspartate and glutamate (left panel in Fig 5A). Their in vivo activity and physiological function, however, have not been examined in animal models.
To determine the metabolic changes associated with the loss of GOT-1, we quantified the amino acid levels of synchronized wildtype and got-1 adults by high performance liquid chromatography (HPLC). As previously reported [50], glutamate is an abundant amino acid, whereas aspartate is maintained at a low abundance in C. elegans (upper panel in Fig 5B). In got-1 mutants, glutamate level exhibited a decrease of 27.6 ± 7.0% (mean ±SEM, n = 4, P = 0.0477 against wildtype animals), whereas the aspartate level exhibited a massive increase of 27.3 ± 4.76 folds (n = 4, P = 0.0076 against wildtype animals) (lower panel in Fig 5B; Fig 6A and 6B). These results implicate that in vivo, GOT-1 preferentially synthesizes glutamate from aspartate (right panel in Fig 5A). Among the C. elegans GOTs, GOT-1 appears to be the key glutamate-synthesizing enzyme, because removing its homologue, GOT-2, did not lead to glutamate reduction (Fig 6A).
We noted that compared to the massive aspartate accumulation, the reduction of glutamate was mild in got-1 mutants. This may result from compensatory activation of glutamate synthesis using other amino acids. As reported [50], alanine is the most abundant amino acid in C. elegans (upper panel in Fig 5B). In got-1 mutants, alanine level exhibited a decrease of 48.5 ± 12.4% (n = 4, P = 0.0050 against wildtype animals) (lower panel in Fig 5B; Fig 6C), supporting the notion that alanine becomes the compensatory source for glutamate when conversion from aspartate is blocked.
Such a drastic shift in equilibrium of the three key amino acids—aspartate, glutamate, and alanine—in got-1 mutants must exert indirect metabolic consequences. To determine whether GOT-1’s loss affects the global metabolic state, we performed liquid chromatography-mass spectrometry (LC-MS) analyses on whole worm lysates. Indeed, got-1 mutants exhibited increased AMP/ATP and NADP/NADPH, and decreased GSH/GSSG glutathione ratios (Fig 6E and 6F), two hallmark features for increased cellular toxicity and metabolic stress [51].
We conclude that in C. elegans, GOT-1 synthesizes glutamate, and maintains glutamate level using aspartate. The loss of GOT-1 leads to glutamate reduction, aspartate accumulation, and potential compensatory glutamate synthesis using other amino acids.
Removing the glutamate-synthesizing GOT-1 restored ubr-1’s bending pattern, suggesting that ubr-1’s motor defects may be associated with glutamate homeostasis.
We assessed the amino acid level in ubr-1 mutants by HPLC. In ubr-1 mutants, the glutamate level was increased by 22.2 ± 9.3% (n = 5, P = 0.0479 against wildtype animals). In ubr-1; got-1 mutants, similar to got-1 mutants, the glutamate level was reduced by 21.6 ± 2.4% (n = 4, P = 0.0230 against wildtype animals; P = 0.0153 against ubr-1 mutants), whereas the aspartate level was increased by 30.2 ± 5.3 folds (n = 4, P = 0.0033 against wildtype animals, P = 0.0031 against ubr-1 mutants) (Fig 6A). Interestingly, there was a mild increase of aspartate in ubr-1 mutants (46.3 ± 12.4%, n = 4, P = 0.0302 against wildtype animals) (Fig 6B), which may reflect a compensatory metabolic response to reduce glutamate accumulation. Reminiscent of the genetic interactions exhibited at the behavioral level, removing GOT-1’s homologue, GOT-2, which did not restore ubr-1’s bending, did not reduce ubr-1’s glutamate either (Fig 6A).
To determine if homeostasis of other metabolic substrates for GOT-1, α-KG and OAA, also contributes to ubr-1’s motor defects, we surveyed their levels in ubr-1, got-1 and ubr-1; got-1 adults by LC-MS/MS. Unlike the case for glutamate, which exhibited inversely correlated changes between ubr-1 and got-1 mutants, and between ubr-1 and ubr-1; got-1 mutants, the α-KG and OAA levels did not exhibit consistent, or correlated changes (Fig 6C). These results support that restored bending in ubr-1; got-1 animals primarily results from reduced glutamate, not from indirect—consequential or compensatory—metabolic effects that the reduced glutamate may have incurred.
Glutamate is a neurotransmitter, as well as an abundant amino acid partaking in other metabolic pathways including amino acid synthesis, energy production and urea cycle. An elevation of glutamate level may result in not only increased neuronal signaling in glutamatergic neurons, but also cellular stress in all UBR-1-expressing cells.
Because most of the critically required premotor interneurons for UBR-1’s role in motor pattern are cholinergic [52], the developmental effect of increased glutamate in these neurons may play a prominent role. To explore this possibility, we examined these interneurons using a transcriptional reporter (Pnmr-1-RFP) [18] that allows visualization of their somata (Fig 7; S6 Fig), and a translational reporter (GLR-1::GFP) [18] that labels synapses of the AVA and AVE premotor interneurons[18] (Fig 7).
In adult ubr-1 mutants, both reporters exhibited prominent and moderate decrease in fluorescence intensity in premotor interneurons AVA and AVE, respectively (Fig 7A–7D for Pnmr-1; E-F for GLR-1::GFP). Such a decrease was not observed for other interneurons including RIM (Fig 7A–7D). We further noted that unlike the smooth and round somata in wildtype animals (S6 Fig, wildtype), in ubr-1 mutants, the AVA and AVE somata exhibited rough edges and short branches in L4 stage larvae and adults (S6A Fig, L4 and Adult panels), whereas other premotor interneurons such as RIM somata appeared normal as in wildtype animals (S6B Fig). The late onset of reversal bending defects coincides with the morphological change in AVA and AVE becoming prominent by the end of the larval development (S6A Fig). Their fluorescent intensity and morphology defects were significantly rescued by restoring UBR-1 expression in premotor interneurons (Fig 7A–7D; S6C Fig), and in adult ubr-1; got-1 mutants (Fig 7A–7D; S6C Fig).
These results support the notion that an increased glutamate level may lead to developmental defects in ubr-1 mutants, and premotor interneurons, in particular AVA and AVE, may be more susceptible to such a perturbation than other neurons or cells.
Our analyses attributed the motor defect of ubr-1 mutants to altered glutamate metabolism, most critically from premotor interneurons. To address how UBR-1 may negatively regulates glutamate level through GOT-1, we first assayed the total GOT activity of the whole worm lysates. We observed a moderate increase of GOT activity in ubr-1 mutants (Fig 8A). Consistent with the presence of multiple GOT homologues, the total GOT activity was drastically reduced, but not abolished in got-1 mutants. The increased GOT activity in ubr-1 mutants was attenuated in ubr-1; got-1 double mutants. The attenuation effect was also specific to GOT-1: the loss of GOT-2 did not reduce GOT activity in ubr-1 mutants (Fig 8A).
An increased GOT activity is consistent with an elevated glutamate level in ubr-1 mutants. However, we did not observe changes in GOT-1 protein level in ubr-1 mutants. The level of GOT-1::GFP, expressed either from its endogenous locus or from an exogenous panneuronal promoter (Prgef-1), was similar between wildtype and ubr-1 mutants (Fig 8B). These results suggest that GOT-1 is not a direct substrate of UBR-1. UBR-1-mediated regulation of glutamate metabolism may involve targeting other pathway components.
Dysregulated E3 activities have been implicated in neurodevelopmental and neurological disorders [53]. Using the C. elegans model and its motor output as the functional readout, we reveal a previously unknown role of UBR-1 in glutamate metabolism. The absence of UBR-1 elevates glutamate level, which promotes simultaneous activation of A-class motor neurons, leading to reduced bending during reversal movements. This defect is compensated by removing a glutamate-synthesizing enzyme and reducing glutamate level in premotor interneurons of the reversal motor circuit. Aberrant glutamate metabolism may underlie ubr-1 mutant’s physiological defects.
Coinciding with an elevated GOT activity, ubr-1 mutants exhibit increased glutamate level. Removing a key glutamate-synthesizing enzyme GOT-1 reduces the glutamate level in ubr-1 mutants. These metabolic changes parallel those in motor behaviors: removing the activity of GOT-1 enzyme restores bending in ubr-1 mutants. These results imply that elevated glutamate level is associated with ubr-1’s defective bending.
Genetic mutations in key metabolic enzymes can exert effects beyond their primary substrates and immediate metabolic pathways. For example, upon the loss of GOT-1, the reduction of glutamate synthesis from aspartate likely induces compensatory utility of other amino acids, thus the reduction of alanine. When converting aspartate to glutamate, GOT-1 may indirectly affect the equilibrium of other substrates, OAA and α-KG, metabolites that take part in multiple metabolic functions, including energy production, amino acid homeostasis, nucleotide synthesis, and lipid synthesis.
Our metabolomics analyses of got-1 mutants revealed no consistent or correlated changes in the level of OAA, α-KG (Fig 6C), and all other TCA cycle metabolites from wildtype controls (S5 Fig). Because the same samples exhibited consistent and correlated changes between the aspartate, glutamate and alanine levels, the variability in other metabolite levels likely reflects flexibility of compensatory or adaptive changes in response to the primary metabolic dysfunction upon the loss of GOT-1. Indeed, the only consistent and anti-correlated metabolic change that we observed between ubr-1 and got-1 (and ubr-1; got-1) mutants was the glutamate level. Cumulatively, these results strongly support that the primary metabolic change—the level of glutamate—causes the motor pattern change in ubr-1 and ubr-1; got-1 mutants.
Our analyses measured glutamate level in whole animals, not the specific neuron groups that are critically required for both UBR-1 and GOT-1’s effect on reversals. Examples of genes with broad expression, but cell-type specific functions have been reported [32, 54, 55]. These and other results–such as the presence and requirement of both UBR-1 and GOT-1 in premotor interneurons in mediating their effects on the reversal motor pattern–suggest that the glutamate level change in premotor interneurons significantly contribute to the motor pattern change in ubr-1 and ubr-1; got-1 mutants. We reckon that other UBR-1-expressing neurons likely also contribute to such an alteration. Furthermore, the glutamate change in other neurons and cells may cause other physiological effects that were not examined in this study, where we only followed the most obvious motor defect.
Glutamate has multiple functions. Being a key neurotransmitter, glutamate also serves as the sole precursor of another neurotransmitter GABA. Glutamate is also one of the abundant amino acids partaking in other metabolic pathways including the amino acid synthesis, energy production and urea cycle. An elevation of glutamate not only affects signaling in glutamatergic and likely GABAergic neurons, but also increases metabolic stress in all cells.
Because both UBR-1 and GOT-1 exert strong effects on the reversal motor pattern through premotor interneurons, most of which cholinergic [52], the developmental effect of increased glutamate may play a prominent role in these interneurons. Indeed, we observed morphological changes of premotor interneurons that coincided with the onset of motor pattern change in ubr-1 mutants. Interestingly, within the reversal motor circuit, premotor interneurons that exhibited the prominent changes, AVA and AVE, provide direct synaptic inputs to the A-class motor neurons. These observations suggest that compared to other cells and neurons, the development and function of premotor interneurons, in particular AVA and AVE, may be more susceptible to glutamate increase, likely through both increased glutamatergic inputs and cellular stress.
Despite its broad expression, GOT-1 can only influence ubr-1’s motor defects through neurons. These results indicate that maintaining glutamate level in neurons, especially the premotor interneurons, is critical for the proper motor output of ubr-1 mutants.
In the mammalian brains, glutamate has to be locally synthesized because it cannot cross the blood-brain barrier [56]. The activity of transaminases, mainly by GOT and ALT, establishes and maintains homeostasis of three abundant amino acids, glutamate, alanine, and aspartate [57, 58]. A recent study suggests that GOT also contributes to glutamate synthesis at synapses [59].
The main sources of glutamate synthesis and homeostasis in C. elegans neurons have remained elusive [60]. Our study establishes GOT-1 as a key enzyme for glutamate synthesis. Further, it provides the first evidence of a critical role of GOT-1 in glutamate metabolism in the nervous system.
How UBR-1 negatively regulates glutamate level remains to be elucidated. UBR family proteins are often addressed in the context of N-end rule E3 ligases [6, 7, 10]. However, UBR proteins have non-N-end rule substrates, and N-end rule substrates do not account for all described functions of UBR proteins [21, 61–64], and their physiological relevance remains to be clarified.
For example, in mammalian cells, UBR1’s N-end rule substrates include pro-apoptotic fragments [11]. In C. elegans, the CED-3 caspase promotes apoptosis. During postembryonic development, it interacts with UBR-1 to expose the N-degron of LIN-28 [10], a regulator of seam cell patterning [65] to promote its degradation. However, neither UBR-1 nor LIN-28’s N-degron were essential for LIN-28’s degradation [10]. Removing either apoptotic regulators CED-3 or CED-4, which generates pro-apoptotic fragments, or LIN-28, neither mimic nor rescue ubr-1 mutants’ bending defects (Table 2; S3 Movie). Substrates for neuronal function of UBR-1 remain to be identified.
Regardless of the nature of UBR-1 targets, our results show that UBR-1 is unlikely to directly regulate GOT-1’s protein stability: we did not observe a change in the level of GOT-1::GFP in ubr-1 mutants. We speculate that UBR-1-mediated regulation of glutamate level may involve targeting other components that subsequently affect glutamate metabolism. For example, the activity of many rate-limiting metabolic enzymes in glycolysis, fatty acid synthesis, cholesterol synthesis, and gluconeogenesis are regulated by phosphorylation and de-phosphorylation [66, 67]. UBR-1 may mediate the ubiquitination of kinases or phosphatases that modify GOT-1’s activity. A large-scale ubr-1 suppressor screen may be required to yield insights on UBR-1’s direct substrates.
The UBR family proteins have been extensively examined in yeast, cultured cells, and mouse models, but only recently in C. elegans. In yeast, UBR1 affects cell cycle progression, but is non-essential [6]. Mouse models revealed the functional redundancy of multiple UBR homologues, where combinatorial knockout of UBR1 and UBR2 results in embryonic lethality [12]. C. elegans UBR1 affects the stability of LIN-28, a regulator of postembryonic hypoderm seam cell division [10]; both UBR1 and LIN-28 are non-essential for viability. The simplicity and viability of the C. elegans ubr-1 model has allowed us to reveal, and genetically dissect a previously unknown physiological role of UBR-1 in glutamate homeostasis. We note that aberrant glutamate metabolism may cause systemic cellular and other unexamined neuronal defects in ubr-1 mutants. However, the simplicity and sensitivity of the C. elegans premotor interneuron circuit, and the prominent motor pattern change provided a quantifiable functional readout to afford genetic pathway dissection.
Our study provides the first demonstration of a UBR protein’s physiological role in glutamate regulation. We noted with interest that in a case study of a JBS patient with severe cognitive impairments, GOT activity levels, used as a biomarker of inflammation, were increased [68]. In addition to its role in development, there is a growing body of evidence for glutamate signaling in non-neuronal tissues [48, 69, 70]. It will be of interest to examine whether glutamate level and signaling in other JBS animal models and JBS patients are aberrant, and whether they contribute to JBS pathophysiology.
Animals of various genotypes were crossed into a reporter hpIs460 (Punc-4-GCaMP6::wCherry), using a calcium reporter GCaMP6 fused in-frame at its C-terminus with wCherry for both tracking and ratiometric correction for motion artifacts in moving animals during recordings [46]. A-MN imaging in moving young adults was carried out as described previously [32]. Briefly, C. elegans were placed on a 2.5% agar pad, immersed in a few drops of M9 buffer, and covered by a coverslip to allow slow crawling. Each recording lasted for 3 minutes. Images were captured using a 40x objective on a Zeiss Axioskop 2 Plus equipped with an ASI MS-40000 motorized stage, a dual-view beam splitter (Photometrics) and a CCD camera (Hamamatsu Orca-R2). The fluorescence excitation light source from X-CITE (EXFO Photonic Solution Inc.) was reduced to prevent saturation of imaging field. The fluorescent images were split by Dual-View with a GFP/RFP filter set onto the CCD camera operated by Micromanager. The 4x-binned images were obtained at 50-msec exposure time (10 frames per second). Regions of interest (ROIs), containing the soma of DA7, VA10 and VA11, were defined and simultaneously tracked using an in-house developed Image J plug-in [32]. The ratio between GFP and RFP fluorescence intensities from individual ROI was plotted over time to produce index for calcium profile for each neuron. Displacement for the DA7 soma, which exhibited the strongest RFP signal, was plotted over time to generate the example instantaneous velocity profiles. For example velocity profiles, we also manually annotated videos to verify the reversal periods. Reversal events longer than 5 seconds were used for cross-correlation analyses.
The pair-wise phase relationships between the activity patterns of A-MNs during each reversal event were assessed by cross-correlation analyses. In each example trace, cross-correlation during the entire period of reversal was calculated to determine the phase lags between the calcium signals of VA11, DA7, and VA10. The maximum of the cross-correlation function denotes the time point when the two signals are best aligned; the corresponding argument of the maximum correlation values denotes the lag between the two neurons. The absolute time lags were used to represent the synchrony of VA11, DA7 and VA10’s activation. N (≥10) refers to the number of reversal events analyzed for each genotype per experiment.
Synchronized last-stage larval worms were grown on 100 mm NGM agar plates seeded with OP50 bacteria. Worms were collected using M9 buffer and were washed thoroughly to remove bacteria. Worm pellets were snap-frozen in liquid nitrogen and pulverized using a cell crusher. Amino acids and other metabolites were extracted by addition of ice-cold extraction solvent (40% acetonitrile, 40% methanol, and 20% water) and incubated on dry ice for an hour with occasional thawing and vortexing. Samples were then moved to a thermo mixer (Eppendorf) and shook for an hour at 4°C at 1400 rpm. These samples were centrifuged at 14000 rpm for 10 min at 4°C. Supernatant was transferred to fresh tubes and lyophilized in a CentreVap concentrator (Labconco) at 4°C. Samples were stored at −80 °C until used for HPLC or LC-MS/MS analyses.
Amino acid quantitation was performed using the Waters Pico-Tag System (Waters). After hydrolysis and pre-column derivatization of the sample by PITC, samples were analyzed by reverse phase HPLC (Amino acid facility, SPARC BioCentre, Sick Kids, Toronto, Canada). LC-MS/MS metabolite analysis was performed as described previously [77]. Raw values were normalized against the total protein concentration as determined by a Bradford Protein Assay (Bio-Rad). Results were compared from at least three sets of independent experiments (N≥3), with all samples collected and analyzed in parallel in each replica.
Synchronized late L4 larval stage worms were grown on 100 mm NGM agar plates seeded with OP50 bacteria. Worms were collected using M9 buffer and were washed thoroughly to remove bacterial contamination. Samples were homogenized by sonication in 100-200ul ice-cold AST buffer. Samples were centrifuged at 13,000g for 10 minutes to remove insoluble materials. The supernatant was used for the AST assay and pellets were used to quantify the total protein in the samples (as described above).
GOT (also referred to as AST) activity was measured using a colorimetric assay with AST Activity Assay Kit (Sigma-Aldrich) according to the manufacturer’s instructions. Data were normalized by the total protein content of the whole worm lysate as determined by a Bradford Protein Assay.
Mixed stage C. elegans were grown on 100 mm NGM plates seeded with OP50 and collected using M9 buffer. Lysates were prepared as described previously [78]. For western blot analyses, total protein concentration was determined using a Bradford Protein Assay (Bio-Rad). Anti-GFP antibodies (Roche) were used to probe for GOT-1::GFP and tubulin was used for the loading control.
For bending curvature and calcium imaging analyses, statistical significance was determined using the Kruskal-Wallis test and the two-way repeated measures (RM) ANOVA and subsequent post-hoc analysis. For metabolite analyses, two-tailed Student’s t-tests were applied to determine statistical differences. p< 0.05 were considered to be statically significant. All statistics were performed using Prism software (GraphPad).
|
10.1371/journal.ppat.1004892 | The Recent Evolution of a Maternally-Inherited Endosymbiont of Ticks Led to the Emergence of the Q Fever Pathogen, Coxiella burnetii | Q fever is a highly infectious disease with a worldwide distribution. Its causative agent, the intracellular bacterium Coxiella burnetii, infects a variety of vertebrate species, including humans. Its evolutionary origin remains almost entirely unknown and uncertainty persists regarding the identity and lifestyle of its ancestors. A few tick species were recently found to harbor maternally-inherited Coxiella-like organisms engaged in symbiotic interactions, but their relationships to the Q fever pathogen remain unclear. Here, we extensively sampled ticks, identifying new and atypical Coxiella strains from 40 of 58 examined species, and used this data to infer the evolutionary processes leading to the emergence of C. burnetii. Phylogenetic analyses of multi-locus typing and whole-genome sequencing data revealed that Coxiella-like organisms represent an ancient and monophyletic group allied to ticks. Remarkably, all known C. burnetii strains originate within this group and are the descendants of a Coxiella-like progenitor hosted by ticks. Using both colony-reared and field-collected gravid females, we further establish the presence of highly efficient maternal transmission of these Coxiella-like organisms in four examined tick species, a pattern coherent with an endosymbiotic lifestyle. Our laboratory culture assays also showed that these Coxiella-like organisms were not amenable to culture in the vertebrate cell environment, suggesting different metabolic requirements compared to C. burnetii. Altogether, this corpus of data demonstrates that C. burnetii recently evolved from an inherited symbiont of ticks which succeeded in infecting vertebrate cells, likely by the acquisition of novel virulence factors.
| How virulent infectious diseases emerge from non-pathogenic organisms is a challenging question. Here, we address this evolutionary issue in the case of Q fever. Its causative agent, the intracellular bacterium Coxiella burnetii, is extremely infectious to humans and a variety of animals. However, uncertainty persists regarding its evolutionary origin, including the identity and lifestyle of its ancestors. In this article, we show that C. burnetii arose from a rare evolutionary transformation of a maternally-inherited endosymbiont of ticks into a specialized and virulent pathogen of vertebrates. While arthropod symbionts are typically transmitted maternally and thought not to be infectious to vertebrates, we establish here that one Coxiella symbiont has evolved the necessary adaptations to exploit the vertebrate cell, leading to the emergence of Q fever.
| ‘Query fever’ (Q fever) is a highly infectious zoonotic disease first identified in 1937 [1,2,3,4,5]. The causative agent, the obligate intracellular bacterium Coxiella burnetii, infects a variety of vertebrate species, including humans. Sporadic cases in humans occur annually worldwide, but occasional outbreaks are also common [1,2,3,4]. For example, in the Netherlands more than 4,000 human cases were reported between 2007 and 2010 [6]. While most human cases are self-limiting with fever and fatigue, acute forms range from mild flu-like symptoms to pneumonia or hepatitis. The disease can also become chronic (mainly endocarditis), and, though rarely fatal, remains highly debilitating even when treated with antibiotics [1,2]. Most human cases are linked to contact with infected livestock, especially goats and sheep, which suffer abortion and reproductive disorders. Infection usually occurs by the inhalation of aerosolized resistant small cell variants that are present in the excretions of infected animals. Other modes of transmission including ingestion of unpasteurized milk or dairy products; human-to-human contact is also possible but considered rare [1,2,4,5]. One of the most virulent reference strains of C. burnetii (strain RSA 493 / Nine Mile I [7]) was isolated from a guinea pig on which field-collected Rocky Mountain wood ticks Dermacentor andersoni had fed, suggesting that transmission through tick bites may also occur [8]. The small cell variants of the bacterium can survive and remain highly infectious for long periods in the environment, leading to the classification of C. burnetii as potential bioterrorism agent [9].
The evolutionary origin of Q fever is unclear since the C. burnetii ancestor and its primary lifestyle remain entirely unknown. Historically, C. burnetii was assigned to the taxonomic order Rickettsiales (Alphaproteobacteria), but it has been recently considered more closely related to the Legionellales order (Gammaproteobacteria) because of its genetic proximity to the Legionnaires' disease agent, Legionella pneumophila [10]. The Legionellales order includes many other intracellular bacteria infecting non-vertebrate species, such as, for instance, Rickettsiella species that are both widespread and biologically diverse in arthropods [11,12,13]. Within the Coxiella genus, the only known relative of C. burnetii which has been formally identified is C. cheraxi, a pathogen of crayfishes [14]. Many past descriptions of Coxiella were likely biased toward the detection of pathogenic strains since most C. burnetii isolates were collected from humans or domestic ruminants during Q fever outbreaks [1,5,15]. However, the advent of 16S rRNA gene sequencing as a universal DNA barcoding marker in bacteria has led to the description of a few novel Coxiella-like organisms in non-vertebrate species (listed in [16]), and particularly in ticks [17,18,19,20,21,22,23,24,25]. All these Coxiella-like organisms are closely related, but genetically distinct to C. burnetii, suggesting that some diversity exists within the Coxiella genus. The highly conserved nature of the 16S rRNA gene sequences has prevented researchers from establishing the exact relationship between C. burnetii and Coxiella-like organisms, and a sister clade relationship is commonly assumed [16,18,21,22,23].
The Coxiella-like organisms differ from C. burnetii in their biological traits and some may behave as subtle symbionts engaged in intricate interactions with ticks. In ticks belonging to Ornithodoros, Amblyomma and Rhipicephalus genera, Coxiella-like organisms were found to massively infect ovaries and to be maternally inherited through the egg cytoplasm [18,20,21,26]. In these tick species, the presence of the bacteria in the Malpighian tubules further suggests a possible role in nutrition by potentially provisioning their hosts with essential nutriments [20,21,26]. Indeed, the elimination of these bacteria with an antibiotic treatment was shown to negatively impact the fitness of the lone star tick A. americanum [27]. Accordingly, when the Coxiella-like bacterium found in A. americanum was recently sequenced [16], no recognizable virulence genes were found, indicating that this bacterium is likely non-pathogenic. In contrast, its genome encodes major vitamin and cofactor biosynthesis pathways, suggesting that it may be a vitamin-provisioning endosymbiont. This interaction exhibits the typical hallmarks of maternally-inherited symbionts with essential roles in arthropod biology [28,29]. Such patterns have been found in other exclusive blood-feeding species like bedbugs [30] and tsetse flies [31], two insect groups which rely on a single food source throughout their developmental cycle and harbor beneficial microbes that provide nutrients absent from their restricted diets. The Coxiella-like organisms of ticks share obvious similarities with these beneficial endosymbionts.
Here, we examine the origin of the Q fever pathogen, C. burnetii, by inferring the evolutionary processes that have shaped diversity within the entire Coxiella genus. To this aim, we first sampled an extensive range of ticks, with 58 tick species examined, and developed a sensitive detection method that reveals a wider Coxiella diversity than recognized in past studies. Second, instead of relying solely on the 16S rRNA gene, a molecular marker that is notoriously inadequate for inferring reliable fine-scale phylogenies [32], we used a novel multilocus typing method, allied to Whole Genome Sequencing (WGS) data, and conducted phylogenetic analyses on a large amount of DNA sequence data. Third, we examined two major ecological features of Coxiella-like organisms, i.e. their ability to be maternally-inherited through the tick egg cytoplasm and to grow in a vertebrate cell environment suitable to C. burnetii. Altogether, this corpus of data has led to the characterization of a large genetic and ecological diversity within the Coxiella genus, far beyond the C. burnetii type species. The Coxiella-like organisms of ticks form an ancient lineage of maternally-inherited tick endosymbionts that do not lie as a sister-clade to C. burnetii but rather form a basal lineage illustrative of the ancestral Coxiella life style.
We performed an extensive screening for the presence of Coxiella in 916 tick specimens from 58 species belonging to the two main tick families, Ixodidae (hard ticks, 36 species) and Argasidae (soft ticks, 22 species) (Fig 1 and Table 1). Except for 37 specimens (6 species) derived from laboratory colonies, all other tick specimens were sampled from natural populations in Europe, Americas, Africa, Oceania and Asia (n = 112 localities). In these populations, ticks were collected either in the host habitat or directly on hosts (Table 1). To detect C. burnetii and its relatives, we developed a detection method based on a nested polymerase chain reaction (PCR) using total tick DNA extracts to amplify a 539–542 base-pair (bp) fragment of the Coxiella rpoB gene (Table A in S1 Text).
Using this procedure, all the tick-borne bacteria we detected belong to the Legionellales order and can be unambiguously assigned either to Coxiella or to its sister genus, Rickettsiella. Whole tick DNA extracts from more than two thirds of the specimens (637 out of 916, 69.6%) and the species (40 out of 58, 70.0%) were found to be positive for Coxiella (Fig 1 and Table 1). Coxiella was found in most tested genera of hard ticks (Rhipicephalus, Ixodes, Amblyomma, Dermacentor, Haemaphysalis) and soft ticks (Ornithodoros, Argas). In almost all infected species, Coxiella was detected in >90% of the examined specimens, indicating high Coxiella prevalence in diverse tick species. For example, infection was apparently fixed in populations of most Rhipicephalus and Ornithodoros species (Table 1). In contrast, Coxiella was frequently absent in Ixodes species and displayed highly variable prevalence in the five infected species (out of 12 screened).
Other Legionellales bacteria of the genus Rickettsiella were found in 52 specimens (5.7%) from six Ornithodoros species and three Ixodes species (Fig 1 and Table 1). In two of the three Rickettsiella-infected Ixodes species, i.e. I. ricinus and I. uriae, Coxiella was also found, but in different individuals and in distinct populations (i.e., no co-infection by Coxiella and Rickettsiella occurred at individual and population levels; Table 1). Adding the Rickettsiella-positive samples, we found that 689 of the 916 examined tick specimens (75.2%) and 44 of 58 screened species (76%) harbored either Coxiella or one of its relatives.
To characterize Coxiella genetic diversity, we developed a multi-locus typing method based on five conserved bacterial genes including rpoB and four other housekeeping genes: 16S rRNA, 23S rRNA, GroEL and dnaK (Table A and Fig A in S1 Text). Multi-locus sequences were obtained from a subsample of 85 Coxiella- and 12 Rickettsiella-positive tick specimens (one to four specimens per infected species were examined). All five bacterial genes were successfully amplified from 71 Coxiella- and 12- Rickettsiella positive specimens representing 35 Coxiella- and six Rickettsiella-infected tick species. For five other Coxiella-infected species (i.e., 14 individual ticks), only three to four bacterial genes were successfully amplified. The sequences were easily readable without double peaks, indicating that there was no coinfection of Coxiella/Rickettsiella strains in any specimen.
The overall dataset included 33 to 40 alleles per bacterial gene (Table 2) and 51 new multi-locus genotypes (43 in Coxiella and eight in Rickettsiella). Within the Coxiella genus, all pairs of 16S rRNA gene sequences are at least 93% identical (Table 2) and range in threshold values typically used to delineate other Legionellales genera such as Legionella [33] and Rickettsiella [34]. Each of the infected tick species harbored a specific bacterial genotype or a set of closely related genotypes. None of the Coxiella multi-locus genotypes identified in ticks was identical to those of the 15 C. burnetii reference strains (Table B in S1 Text), although some showed moderate levels of nucleotide identity: pairwise identity between the two groups ranged from 77.8% to 97.7%. For each bacterial gene, the genetic diversity was significantly higher in the Coxiella strains of ticks than in C. burnetii as illustrated by the metrics on their respective genetic diversity (Table 2, paired t test, all P < 0.02).
We constructed a multi-gene phylogeny of the entire Coxiella genus using a dataset that included the Coxiella and Rickettsiella sequences from ticks, the 15 C. burnetii reference genomes, as well as sequences from Legionella spp. and more distant outgroups that were available in GenBank (Table B in S1 Text). The concatenated sequences included 3009 unambiguously aligned base pairs (bp). Prior recombination tests showed that Coxiella and Rickettsiella strains did not exhibit a strictly clonal structure, but rather experienced significant genetic exchanges. We thus applied a sequence-based network approach that does not force relationships to be tree-like but rather incorporates recombination into the phylogenetic reconstruction. The network results (Fig 2), as well as the results from the Maximum Likelihood (ML) tree-based analysis (Fig B in S1 Text), consistently showed that the Coxiella genus can be split into four main clades (labeled A-to-D hereafter) with each clade clustering the Coxiella genotypes found in five to 15 tick species. The phylogenetic analyses also highlight that all C. burnetii isolates cluster into a unique subclade embedded within the A clade (Fig B and C in S1 Text and Fig 2). Notably, the closest relatives of C. burnetii are the Coxiella strains from soft ticks of the Ornithodoros and Argas genera, suggesting that the common ancestor of C. burnetii originated from a Coxiella hosted by soft ticks.
The partitioning of Coxiella diversity among tick species revealed a complex structure, indicating a role for both co-divergence and horizontal transfer events in the evolution of this bacterial group. Closely related Coxiella-like organisms were frequently found in closely related tick species, a pattern suggestive of co-divergence between Coxiella and ticks (Fig B in S1 Text and Fig 2). For instance, all the Coxiella-like organisms found in the 12 examined Rhipicephalus tick species cluster together within the C clade, whereas all the Coxiella-like organisms in the Ixodes species cluster within the B clade (Fig B in S1 Text and Fig 2). Conversely, some Coxiella-like organisms found in related tick species are only distantly related and do not cluster together (e.g., the Coxiella-like organisms of Ornithodoros soft ticks are scattered among the A, B and C clades), a pattern suggestive of horizontal transfers among tick species.
Further analyses were conducted by examining public repositories of DNA sequencing data generated by the whole genome sequencing (WGS) projects of the cattle tick R. microplus and the deer tick I. scapularis. Using the 1,995,281 bp C. burnetii (str. Nine Mile I RSA 493) genome as a probe, we found clear evidence of Coxiella infections in R. microplus, but not in I. scapularis. A total of 31 contigs (514–2,349 bp, totaling 34,990 bp) from R. microplus sequencing were uniquely attributable to Coxiella. They matched 50 genes of C. burnetii with 68-to-100% nucleotide identity (Fig A and Table C in S1 Text and Fig 3A). Alignment of the 31 Coxiella contigs to other bacterial genomes, including the 15 C. burnetii reference genomes (19,304 unambiguously aligned bp), corroborates the finding of our prior five loci-based analyses: the Coxiella strain identified in R. microplus is evolutionarily related, but distinct, to C. burnetii (Fig 3B).
It should be noted that the R. microplus WGS DNA examined above was extracted from eggs of an inbred strain of ticks (Deutsch strain), first derived from a few field specimens sampled in Texas in 2001, and reared for at least seven generations in the laboratory. The presence of Coxiella DNA in the WGS of R. microplus eggs thus raised the issue of their maternal inheritance in ticks. To address this question, 24 gravid females of four Coxiella-positive tick species were collected either from seabird nests (O. maritimus, n = 8 females), from a dog (R. sanguineus, n = 1) or from laboratory colonies (R. microplus, n = 7; A. americanum, n = 8) in order to test for the presence of Coxiella-infection in the cytoplasm of their progeny (8 to 14 surface-sterilized eggs per female were individually examined; i.e., 244 eggs in total). The occurrence of maternal transmission was detected in all four tick species and in almost all eggs: O. maritimus—79 Coxiella-positive eggs out of 80, R. sanguineus—14 out of 14, R. microplus—68 out of 70, and A. americanum—80 of 80. The mean transmission rate can thus be estimated at 0.988 (95% confidence interval, 0.965–0.994), demonstrating highly efficient maternal transmission of Coxiella in ticks. Maternal inheritance is thus widespread in the Coxiella genera, being found in three different clades (A: Coxiella-like organism of O. maritimus; C: R. sanguineus and R. microplus; D: A. americanum).
We next compared the metabolic requirements of Coxiella-like organisms with those of C. burnetii by assessing their ability to replicate in both an axenic medium ACCM2 (mimicking the environment of the acidified lysosome-like vacuoles of phagocytes typically colonized by C. burnetii; [35]) and directly inside vertebrate host cells. First, ACCM2 was inoculated with Coxiella-like organisms extracted from eggs of either O. maritimus, R. microplus or A. americanum. Inoculated media were incubated for 10 days under standard conditions used to amplify C. burnetii. Although our C. burnetii positive controls readily replicated in the media, the Coxiella obtained from eggs of the three tick species did not grow. We then incubated egg homogenates from ticks of O. maritimus and R. microplus with mammalian cell cultures for seven days. Similar to results under axenic conditions, the incubation of vertebrate cell lines with Coxiella-like organisms failed to produce Coxiella-containing vacuoles, whereas the same cell lines incubated with C. burnetii under the same conditions were readily infected. The apparent inability to amplify tick-borne Coxiella through standardized protocols, well-characterized for C. burnetii, suggests that, despite their phylogenetic proximity, the Coxiella-like bacteria are adapted to radically different environments.
Since its original description, C. burnetii infections have been characterized in a wide variety of hosts. While only two species have been formally identified within the Coxiella genus, we show here that a far greater diversity of Coxiella exists in ticks. We detect the presence of Coxiella-like organisms in many more tick species than previously known [17,18,19,20,21,22,23,24,25] and describe a far wider genetic diversity among these bacteria than previously suspected. The incidence of Coxiella, as well as of its sister genus Rickettsiella, in ticks is exceptionally high, with approximately three quarters of tick species infected. Although possible tick-borne transmission of C. burnetii has been reported [1,2,8], none of the 43 new Coxiella genotypes identified here are identical to C. burnetii. We also demonstrate for genetically divergent Coxiella strains (i.e., members of the A, C and D clades) found in four tick species that infection is primarily transmitted maternally via the egg cytoplasm. These results converge to support the hypothesis that these Coxiella-like organisms are specific endosymbionts of ticks. Phylogenetic evidence further shows that one of the Coxiella-like organisms belonging to the A clade and primarily hosted by soft ticks has served as the progenitor of C. burnetii.
Three complementary lines of argument indicate a much longer evolutionary history for Coxiella-tick associations than for vertebrate-Coxiella associations. The first lies in the broad distribution of Coxiella and Rickettsiella bacteria across tick species, genera and families. The second concerns the extensive genetic diversity found in tick-borne Coxiella strains compared to C. burnetii strains, as illustrated by the clear subdivision of this genus into four highly divergent clades (A-D). Finally, the clustering of all C. burnetii strains within one of the clades of tick-borne Coxiella shows that the ancestor of C. burnetii was a tick-associated bacterium which succeeded in infecting vertebrate cells. The remarkably low genetic diversity of C. burnetii, previously noted in other studies [36,37], indicates a unique and recent emergence of this highly infectious vertebrate pathogen. Interestingly, this hypothesis was initially raised a decade ago from observations of the profound differences in genome architecture of C. burnetii relative to other pathogenic intracellular bacteria [38]. It was again emphasized from the genome sequencing of new C. burnetii strains [39]. Our data brings further support to this hypothesis by demonstrating that C. burnetii roots within the Coxiella phylogeny. Comparative genome sequences of C. burnetii [38, 39] and of the Coxiella-like organism from A. americanum [16] also suggest that Coxiella bacteria differ substantially in terms of genome size and gene content. The C. burnetii genome (A clade) has a size of ca 2Mb [38, 39], whereas the genome of the Coxiella-like organism isolated in A. americanum (D clade) is only about a 1/3 of this size (ie. 0.66 Mb) with a large percentage of missing genes [16]. This reduction in genome size may limit the transition to pathogenicity, and suggests that some Coxiella-like organisms may have evolved towards exclusive and irreversibly specialized interactions with their tick hosts. Overall, the diversity of genome sizes emphasizes that members of the different Coxiella clades may have retained a variety of evolutionary strategies to favour their spread and persistence in their hosts.
We identified Coxiella as a major emerging clade of bacterial endosymbionts allied to ticks. Coxiella-like organisms are maternally-transmitted through the egg cytoplasm at high frequency with 98–100% mother-to-offspring transmission, a pattern also reported in previous studies [20,21,26]. This transmission pattern is the rule for a variety of bacterial endosymbionts that live exclusively within arthropod cells [28,29,40]. While some, like Wolbachia, are globally common symbionts estimated to infect ca. 40% of insect species [41,42], others are globally rare, but common and important in particular arthropod groups [28]. This is precisely the case for Coxiella-like organisms; although they have not been found in other arthropod species, they are commonly associated with ticks. This leads to the obvious question of the phenotypic consequences of Coxiella-tick interactions. In some cases, Coxiella-like organisms of ticks likely act as obligate mutualistic symbionts required to support normal tick development, potentially provisioning their hosts with essential nutriments absent in vertebrate blood [16,20,21,26,27]. The ubiquity of Coxiella in some tick groups-such as in the Rhipicephalus genus in which infection is at fixation- corroborates the hypothesis of an obligate endosymbiont. This is not, however, the case for all tick species since some, such as I. ricinus and I. uriae, harbour Coxiella-like organisms at much lower frequencies. In these tick species, Coxiella is more likely to behave as a conditional mutualist-i.e., that confers advantages under certain environmental conditions- or as a reproductive parasite-i.e., that manipulates host reproduction toward the production of daughters (the transmitting sex), as commonly observed in arthropods with a variety of facultative symbionts [28,40]. It should also be noted that other endosymbionts also occur in ticks and may have evolved under complex multispecific interactions [17,24]. For instance, whereas the soft tick O. moubata was not found to be infected by a Coxiella-like organism in the present study, this tick species has been found to be infected by an endosymbiont belonging to the Francisella genus [17]. Endosymbionts other than Coxiella may thus interact with ticks, a pattern suggesting that endosymbiotic systems can be dynamic across tick lineages. These different hypotheses will now require specific testing.
Another question remains concerning the degree of vertebrate infection risk by the Coxiella-like organisms of ticks. Ticks are found worldwide and blood-feed on many different hosts; a combination of traits that may facilitate tick-to-vertebrate transfers of Coxiella. However, the bacteria observed in this study seem confined to ticks and, to our knowledge, none have ever been isolated from a vertebrate or associated with clinical symptoms. This suggests that these tick-associated bacteria currently pose a much lower infection risk to vertebrates than C. burnetii. As discussed above, the genome reduction of the Coxiella-like organism isolated in A. americanum, with the lack of nearly all the genes associated with pathogenicity [16], corroborates this view. Moreover, the inability to grow tick-borne bacteria in vertebrate cells highlights the significant barrier that must be overcome by the bacteria to successively achieve tick-to-vertebrate transmission. This type of transmission may, nonetheless, occasionally occur; an avian Coxiella-like organism was recently reported to induce fatal systematic infections in domestic birds [43,44,45]. A very similar infection pattern was found for another maternally inherited endosymbiont, Arsenophonus, a widespread bacterium in different insect groups [41,46,47,48,49]. In particular, some Arsenophonus strains were detected in the phloem of plants fed on by infected phytophagous insects and were assumed to be opportunistic plant pathogens [50]. In such cases, the plant host may act as an ecological arenas for the global exchange of endosymbionts like Arsenophonus, serving as a possible intermediate host for the horizontal transfer of bacteria among insect species [48]. In the case of Coxiella-like organisms, the extent of exchange between different tick species via the vertebrate host is yet to be established, but could be favoured by tick co-feeding (ticks feeding in close proximity on the host). The genetic similarity between Coxiella-like organisms found in unrelated tick species highlights the capacity to shift tick host species. Future research is now needed to assess the potential of different Coxiella-like organisms to infect vertebrates.
The reasons why C. burnetii is a highly virulent pathogen of vertebrates, but not Coxiella-like organisms (especially those from the A clade) remain unknown. As an intracellular pathogen with airborne transmission, C. burnetii has evolved specific mechanisms to survive in the abiotic environment, as well as to infect and exploit vertebrate cells [15]. Several evolutionary pathways may explain the acquisition of the genetic material necessary for this major lifestyle transition; this includes spontaneous genetic mutations in the genome of a tick-Coxiella ancestor, or the more likely transfer and integration of virulence genes from a co-infecting pathogen. The opportunity of gene transfer among bacteria, irrespective of their pathogenic or symbiotic properties, relies on their frequent co-occurrence within the same tick host [25,51,52]. The Coxiella-like organisms of the A clade may have dynamic genomes as observed in many arthropod symbionts: although they reside in confined intracellular environments, arthropod symbionts commonly experience variable degrees of recombination and gene transfer [53,54,55,56,57]. These gene transfers have served as immediate and powerful mechanisms of rapid adaptation in many endosymbionts, such as Wolbachia [56] and Hamiltonella [55,57]. This mechanism may explain the evolutionary transition from a Coxiella tick-symbiont of the A clade to the vertebrate pathogen C. burnetii. Other genetic connections are also possible; several C. burnetii genes that may contribute to major virulence traits, such as tissue tropism, are similar to eukaryotic genes and may have been acquired through lateral gene transfers from eukaryotes [38,39]. Detailed studies of virulence genes in C. burnetii and their homology with Coxiella-like organisms of the A clade will now be necessary to understand the remarkable emergence of the Q fever agent.
The evolutionary transition observed within the Coxiella genus is one of the rare cases reported to date of an arthropod-inherited symbiont evolving metabolic adaptations leading to the emergence of a vertebrate infectious disease. Another such transition occurred in the Rickettsia genus. The best-known members of this genus are transmitted by blood-feeding arthropods and are pathogenic in the vertebrate host. However, in recent years, many maternally-inherited Rickettsia endosymbionts found exclusively in arthropods have been discovered [58,59]. The examination of the evolutionary history of the Rickettsia genus revealed that this bacterium originated from endosymbionts of invertebrates and only secondarily became vertebrate pathogens [58,59]. Like Coxiella, some Rickettsia species of blood-feeding hosts have underwent a horizontal transmission through a vertebrate host, leading to pathogen emergence. Other bacteria, such as Arsenophonus [47,48] and Sodalis [60] may have had similar life cycle transitions, but the case of Coxiella is unique in that the arthropod host is no longer required to complete its life cycle.
In conclusion, we show that C. burnetii arose from a rare and recent event: the evolutionary transformation of a maternally inherited endosymbiont of ticks into a specialized and virulent pathogen of vertebrates. This raises a series of exciting questions related to both how Coxiella endosymbionts made the major evolutionary transition leading to the emergence of Q fever and their role in the population dynamics of ticks. Identifying the evolutionary processes that transform symbiotic bacteria into emerging pathogens will require further exploration into the biology of the entire Coxiella genus.
The examined specimens represent the two main tick families, nine genera, 58 species and 112 populations from around the world (Table 1). Field specimens were sampled on various host species belonging to major mammal and bird families or from their habitats. We also used specimens from laboratory colonies reared in captivity for at least three generations for six tick species (derived from field specimens collected in North America, South America, Africa and China). All samples were preserved in 70–90% ethanol at room temperature until use. Before storage, tick eggs collected under laboratory conditions were surface-sterilized with 2.6% sodium hypochlorite and 0.5% SDS for 1 min and washed with sterile water to avoid external bacterial contamination.
Tick DNA was individually extracted using the DNeasy Blood & Tissue Kit (QIAGEN) following manufacturer instructions. DNA template quality was systematically verified by PCR amplification of the 18S ribosomal RNA (18S rRNA) or the cytochrome oxydase 1 (C01) arthropod primers (Table A in S1 Text). Tick DNA samples were then tested for Coxiella presence using a nested PCR assay and sequencing of the rpoB gene using Coxiella-specific primers. The use of nested PCR was efficient at decreasing the probability of contamination from unwanted amplification products. Additional PCR assays on the 16S rRNA, 23S rRNA, GroEL and dnaK genes were conducted on a subsample of Coxiella-positive tick DNA to obtain additional DNA sequences for phylogenetic analyses. We used 15 recently published genomes of C. burnetii (mainly isolated from humans and ruminants) and the genome of Rickettsiella grylli from woodlice (listed in Table B in S1 Text) as references to design PCR primers. The efficiency of our typing method was ascertained through positive PCR amplification and clear sequences for the five loci in four cultured reference strains of C. burnetii (Table B in S1 Text). Gene features, primers and PCR conditions are detailed in Table A in S1 Text. All PCR products were visualized through electrophoresis in a 1.5% agarose gel. Positive PCR products were purified and sequenced in both directions (EUROFINS). The chromatograms were manually inspected and cleaned with CHROMAS LITE (http://www.technelysium.com.au/chromas_lite.html) and sequence alignments were done using CLUSTALW [61], both implemented in MEGA [62].
Coxiella sequences were also searched for in the whole genome sequence (WGS) data of R. microplus and I. scapularis (GenBank accession numbers ADMZ02000000 and ABJB000000000, respectively) using the 1,995,281 bp C. burnetii genome (str. Nine Mile I RSA 493, GenBank accession number NC002971) as a probe and the Basic Local Alignment Search Tool (BLAST) with default parameters. Table C in S1 Text reports the number and content of Coxiella contigs that were detected in the R. microplus WGS data.
The GBLOCKS program [63] with default parameters was used to remove poorly aligned positions and to obtain non-ambiguous sequence alignments. Sequences of individual genes that differed by one or more nucleotides were assigned distinct allele numbers using DNASP [64], with the option of excluding sites with alignment gaps and/or missing data. Tick-borne Coxiella strains are defined as each unique combination of alleles. The genetic diversity estimates (Ps, number of polymorphic sites; Ad, allelic diversity; π, nucleotide diversity; D, average number of nucleotide differences between sequences) were computed using DNASP. Other statistical analyses were carried out using the R statistical package. All sequence alignments were checked for putative recombinant regions using the GENECONV [65] and RDP [66] methods available in the RDP3 computer analysis package [67].
Phylogenetic analyses were based on single and concatenated sequences of the five bacterial genes used in the multi-locus typing scheme and on the 50 Coxiella genes found in the R. microplus WGS data. Sequence alignments included Coxiella and Rickettsiella sequences obtained in this study from tick DNA, as well as sequences available in GenBank from reference strains of C. burnetii, Rickettsiella grylli, Legionella pneumophila, L. longbeacheae, and two more distantly related bacteria, Escherichia coli and Salmonella enterica (Table B in S1 Text). The evolutionary models fitting the sequence data most closely were determined using the Akaike information criterion with the program MEGA. For each data set examined, the best-fit approximation was the general time reversible model with gamma distribution and invariant sites (GTR+G+I). Network-based phylogenetic analyses were done using SplitsTree, implementing the evolutionary model under the agglomerating NeighborNet algorithm [68]. Tree-based phylogenetic analyses were done using maximum-likelihood (ML) analyses. A ML heuristic search using a starting tree obtained by neighbor-joining was conducted in MEGA. Clade robustness was assessed by bootstrap analysis using 1,000 replicates.
We first assessed the ability of tick-borne Coxiella to replicate in an axenic medium as follows. Tick eggs were surface-sterilized as described above and homogenized by hand in sterile water. Eggs homogenates were used to inoculate 2ml of the axenic medium ACCM2 [35] and incubated three days in a humidified atmosphere of 5% CO2 and 2.5% O2 at 37°C. 50μl of each culture were then diluted in 2ml of fresh ACCM2 and further incubated under the same conditions for 10 days to assess bacterial growth. We then assessed the ability of tick-borne Coxiella to replicate inside vertebrate host cells as follows. Surface-sterilized tick eggs were homogenized by hand in 1 ml of 10% Fœtal Bovine Sérum (FBS) supplemented MEM medium (GIBCO). The homogenate (0.5 ml) was diluted in 25 ml of 10% SVF-MEM and centrifuged at 4000 rpm (2000g) at 4°C for 30 min. Ten ml of the supernatant was mixed with 10 ml of 10% FSB-MEM and again centrifuged at 2000g at 4°C for 30 min. Ten ml of the supernatant was harvested and filtered through a sterile 0.45 μm pore size filter (MILLIPORE). Two flasks containing confluent Sheep Fœtal Thymus cells (SFT) cells were inoculated with 5 ml of the obtained filtrate and incubated at 35°C and allowed to grow for 12 weeks. Cell culture flasks were observed daily for the presence of contamination or growth signs such as vacuoles containing Coxiella, during the first week then once a week. As a positive control, a homogenate of C. burnetii was used following the same protocol.
Nucleotide sequences of PCR-amplified fragments of tick-borne Coxiella and Rickettsiella genes have been deposited in the GenBank nucleotide database under accession codes KP994768-KP994862 (16S rRNA), KP994678-KP994767 (23S rRNA), KP985445-KP985537 (GroEL), KP985265-KP985357 (rpoB) and KP985358-KP985444 (dnaK).
|
10.1371/journal.pmed.1002831 | Mediating roles of preterm birth and restricted fetal growth in the relationship between maternal education and infant mortality: A Danish population-based cohort study | Socioeconomic disparities in infant mortality have persisted for decades in high-income countries and may have become stronger in some populations. Therefore, new understandings of the mechanisms that underlie socioeconomic differences in infant deaths are essential for creating and implementing health initiatives to reduce these deaths. We aimed to explore whether and the extent to which preterm birth (PTB) and small for gestational age (SGA) at birth mediate the association between maternal education and infant mortality.
We developed a population-based cohort study to include all 1,994,618 live singletons born in Denmark in 1981–2015. Infants were followed from birth until death, emigration, or the day before the first birthday, whichever came first. Maternal education at childbirth was categorized as low, medium, or high. An inverse probability weighting of marginal structural models was used to estimate the controlled direct effect (CDE) of maternal education on offspring infant mortality, further split into neonatal (0–27 days) and postneonatal (28–364 days) deaths, and portion eliminated (PE) by eliminating mediation by PTB and SGA. The proportion eliminated by eliminating mediation by PTB and SGA was reported if the mortality rate ratios (MRRs) of CDE and PE were in the same direction. The MRRs between maternal education and infant mortality were 1.63 (95% CI 1.48–1.80, P < 0.001) and 1.19 (95% CI 1.08–1.31, P < 0.001) for low and medium versus high education, respectively. The estimated proportions of these total associations eliminated by reducing PTB and SGA together were 55% (MRRPE = 1.27, 95% CI 1.15–1.40, P < 0.001) for low and 60% (MRRPE = 1.11, 95% CI 1.01–1.22, P = 0.037) for medium versus high education. The proportions eliminated by eliminating PTB and SGA separately were, respectively, 46% and 11% for low education (versus high education) and 48% and 13% for medium education (versus high education). PTB and SGA together contributed more to the association of maternal educational disparities with neonatal mortality (proportion eliminated: 75%–81%) than with postneonatal mortality (proportion eliminated: 21%–23%). Limitations of the study include the untestable assumption of no unmeasured confounders for the causal mediation analysis, and the limited generalizability of the findings to other countries with varying disparities in access and quality of perinatal healthcare.
PTB and SGA may play substantial roles in the relationship between low maternal education and infant mortality, especially for neonatal mortality. The mediating role of PTB appeared to be much stronger than that of SGA. Public health strategies aimed at reducing neonatal mortality in high-income countries may need to address socially related prenatal risk factors of PTB and impaired fetal growth. The substantial association of maternal education with postneonatal mortality not accounted for by PTB or SGA could reflect unaddressed educational disparities in infant care or other factors.
| Infant mortality in high-income countries has decreased in recent decades.
However, socioeconomic inequality in infant mortality remains and may have become stronger in some populations.
This study provides in-depth knowledge on the underlying pathways from maternal socioeconomic inequality to infant mortality, which is important for developing preventive strategies to reduce potentially preventable deaths.
We conducted a national population-based cohort study that included all 1.99 million live singletons born in Denmark in 1981–2015.
Using modern modeling methods for causal mediation analysis, we found that preterm birth and restricted fetal growth may mediate the association between low maternal education and infant mortality.
The mediated effects through preterm birth and restricted fetal growth were substantial for neonatal mortality but not for postneonatal mortality. The effect through preterm birth seemed to be greater than that through restricted fetal growth.
Public health strategies could aim to reduce educational differences in the death of newborns in high-income countries by addressing socially related risk factors for preterm birth and impaired fetal growth, such as the accessibility and quality of maternity and perinatal care, or maternal lifestyle factors.
The substantial direct impact of low maternal education on postneonatal death not explained by preterm birth or restricted fetal growth could be due to unaddressed educational inequality in infant care or other factors such as housing conditions, the quality of medical care, or the use of specialized medical care.
| Infant mortality in high-income countries has decreased over the last decades [1]. Nevertheless, socioeconomic disparities in infant mortality persist [2–5] and may even have become stronger in some populations [6,7]. A new understanding of the mechanisms that underlie socioeconomic differences in infant deaths is essential to guide health initiatives to reduce potentially preventable deaths.
Preterm birth (PTB), small for gestational age (SGA), and low birth weight (LBW) are not only main risk indicators for infant morbidity and mortality, particularly during the neonatal period, but are also associated with socioeconomic disadvantage [4,8–14]. The primary causes of LBW are PTB and fetal growth restriction [12]. It is possible that the association between socioeconomic disadvantage and infant mortality is mediated through PTB and SGA as a proxy of fetal growth restriction [12,15]. PTB and fetal growth restriction may reflect different mechanisms, and the risk factors for PTB and fetal growth restriction are different [16,17]. The risk of infant mortality is higher among preterm-born infants than SGA infants [12,18]. However, there is a lack of research addressing the possible mediating role of PTB and SGA in the pathway linking maternal socioeconomic disadvantage to an increased risk of infant mortality. Causal mediation analysis may help advance our understanding of when and how socioeconomic disadvantage has a large impact on infant mortality. If the mediating pathway through PTB and SGA plays a major role, policies should focus on the health of women before and during pregnancy to reduce the risk of PTB and SGA. Otherwise, interventions targeting other factors should be explored.
Specifically, using data from Danish registers, we aimed to use causal mediation analysis to examine whether and the extent to which PTB and SGA mediate the association of maternal education with infant mortality. Analyses were stratified by the time (neonatal or postneonatal period) and cause (disease or external cause) of death.
The study was approved by the Data Protection Agency and Research Ethics Committee of the Central Region in Denmark. By Danish law, no informed consent is required for a register-based study using anonymized data. The prespecified analysis plan is given in S1 Text.
The unique personal identification number in Denmark allows accurate linkage of personal data across national registers. The Danish Civil Registration System was linked to the Danish National Patient Register, the Danish Medical Birth Register, the Danish Register of Causes of Death, and the Danish Integrated Database for Labor Market Research [19]. There were 2,096,320 live singleton births registered in Denmark in 1981–2015. We excluded infants with birthweight less than 500 grams or gestational age at birth less than 22 weeks, as criteria for registration of live births and stillbirth vary internationally [20]. We also excluded infants with missing information on maternal education. Follow-up started at birth and ended at death, emigration, or the day before the first birthday, whichever came first.
The Danish Integrated Database for Labor Market Research [19] provided the information on maternal education, which was measured as the highest level of education attained at childbirth and categorized as low (primary and lower secondary education), medium (upper secondary education or academy profession degree), or high (university education at bachelor’s degree level or higher).
The outcome of interest was all-cause infant mortality (0–364 days), which was divided into neonatal mortality (0–27 days) and postneonatal mortality (28–364 days). We also investigated mortality by cause of death (death due to disease or medical condition, or external cause), as well as death due to certain conditions originating in the perinatal period [ICD-8 codes 760–779 and ICD-10 P00–P96] and congenital malformations [ICD-8 740–759 and ICD-10 Q00–Q99], according to the European Shortlist for Causes of Death [21].
Birth weight and gestational age were extracted from the Danish Medical Birth Register [22]. Gestational age was estimated by the date of last menstrual period and, for all pregnancies since 1995, was adjusted, if necessary, by ultrasonography. The mediators were dichotomized as PTB (1 if gestational age at birth < 37 weeks, 0 otherwise) and SGA (1 if birthweight below the 10th percentile for infants of the same gestational age, sex, and birth year, 0 otherwise) in the main analyses. We also used finer categorizations of PTB (<28, 28–31, 32–36, or 37+ weeks) and SGA (birthweight below the 3rd, between the 3rd and 10th, or above the 10th percentile for infants of the same gestational age, sex, and birth year).
Potential confounders (covariates) included maternal age at birth (<20, 20–24, 25–29, 30–34, ≥35 years), parity (1, 2, 3, ≥4), maternal smoking at delivery (yes, no), maternal cohabitation at delivery (single, cohabitation), maternal residence at delivery (Copenhagen, city with ≥100,000 inhabitants, small town/other), diagnosis of congenital malformation of the infant (yes, no), infant sex (male, female), and birth year of the infant (in 5-year intervals). Maternal smoking and congenital malformation and sex of the child were used to adjust for the confounding between the mediators and the outcome when estimating controlled direct effect (CDE); we preserved their potential mediating roles between the exposure and the outcome (in the statistical analysis).
The approach for causal mediation analysis was based on a counterfactual framework whereby the total effect (TE) can be decomposed into controlled direct effect (CDE) and portion eliminated (PE) [23–25]. The CDE captured the influence of maternal education on infant (neonatal and postneonatal) mortality if the link between maternal education and the mediator (PTB or SGA) was prevented or removed hypothetically. This simulated a scenario wherein the sample distributions of the mediator were no longer dependent on maternal education. PE, the difference between TE and CDE, measured the portion of the TE of maternal education that would be eliminated by eliminating the mediator. TE and CDE were estimated using inverse-probability-weighted marginal structural models (MSMs) [26]. For the MSM for the TE, we used weighted regressions of infant mortality on maternal education. The weight, which is called the inverse-probability-of-treatment weight (IPTW), was estimated for each mother in the sample as the ratio of (i) the estimated marginal probability of the mother’s actual educational attainment to (ii) the estimated probability of each mother’s actual educational attainment conditional on their aforementioned covariates (excluding maternal smoking, offspring sex, congenital malformation of offspring, PTB, and SGA). The IPTW simulates the scenario wherein these covariates, which could be confounders, are no longer associated with maternal education, thus eliminating any confounding by these covariates. To estimate the CDE, the corresponding MSM used a product of the IPTW for maternal education and an additional inverse-probability-of-mediators weight (IPMW). The IPMW was estimated for each infant in the sample as the ratio of (i) the estimated marginal probability of the infant’s actual PTB and SGA to (ii) the estimated probability of each infant’s actual PTB and SGA conditional on their aforementioned covariates. The IPMW simulates the scenario wherein the mediators (PTB and SGA) are no longer associated with maternal education, thus eliminating any mediation by the mediators. The PE was subsequently estimated from the model for TE offsetting the estimated CDE. We considered possible exposure–mediator and mediator–mediator interactions. As the mediator–mediator interactions in causal mediation analysis with multiple mediators were null, the final models included only exposure–mediator interactions. We estimated mortality rate ratios (MRRs) with their 95% confidence intervals (CIs) based on robust variance estimation. The proportion of the TE eliminated through the 2 mediators, i.e., the proportion eliminated ([MRRTE−MRRCDE]/[MRRTE− 1]), was reported if the MRRs of CDE and PE were in the same direction [23,27]. We first assessed the mediating role of PTB and SGA separately, i.e., one mediator at a time. Then we analyzed PTB and SGA together as a joint mediator, i.e., not separating their individual contributions. We also examined the mediating roles of PTB in non-SGA infants as well as the mediating role of SGA in term-born infants. The mediation analyses were performed according to birth year (1981–2015 in 5-year intervals). We used the missing-value indicator method to deal with missing values, such that missing values were treated as a separate category. A detailed description of the inverse probability weighting approach for causal mediation analysis is given in S2 Text. We also performed additional mediation analysis using a traditional approach (“with and without mediator”) [28].
We performed a sensitivity analysis to assess and adjust for violations of the uncontrolled confounding assumption [23,29]. Specifically, we considered a binary unmeasured confounding variable U indicating a common cause of PTB, SGA, and infant mortality (e.g., maternal alcohol use, maternal body mass index, prenatal care, maternal psychological stress in adolescence, or maternal school attendance [30,31]). We assumed that among infants with normal gestational age and normal birth weight for gestational age, the prevalence of U was 20% for the low maternal education group, 30% for the medium maternal education group, and 40% for the high maternal education group. We also considered a simplified assumption that the prevalence of U was the same among different maternal education groups. We evaluated the impact of unmeasured mediator–outcome confounding in 2 settings: (i) moderate confounding, where we considered if U increased the likelihood of infant mortality by a factor of 1.5, and (ii) strong confounding, where we considered if U increased the likelihood of infant mortality by a factor of 2.5 [30]. We also performed analysis with an additional adjustment for maternal country of origin. All analyses were conducted using SAS 9.4 (SAS Institute, Cary, NC, US) and Stata 13 (StataCorp, College Station, TX, US).
The study included 1,994,618 infants. Excluded infants (N = 101,702, 4.85%) had a higher infant mortality rate (1.46%) than included infants (0.43%) and tended to have a younger mother. Overall, infant mortality was 4.3 per 1,000 births (8,563 died), and 61.86% of deaths occurred in the neonatal period. Compared with infants of mothers with medium or high education, infants of mothers with low education were more likely to have increased risk of death, to have PTB, to be SGA, and to be born of mothers of younger childbearing age, with higher parity, who lived alone, and who were from small towns (Table 1).
Regarding the total association of maternal education with mortality, we observed that MRRs decreased with increasing education level (Table 2; S1 Fig). The MRRTE of association with low maternal education (versus high) was 1.63 (95% CI 1.48–1.80, P < 0.001) for infant mortality (neonatal and postneonatal), 1.57 (95% CI 1.38–1.78, P < 0.001) for neonatal mortality, and 1.75 (95% CI 1.49–2.04, P < 0.001) for postneonatal mortality. We found that PTB and SGA were associated with both maternal education and infant mortality, and the association between PTB and infant mortality was much stronger than the association between SGA and infant mortality (S1 Table). Analyses stratified by birth year found a stronger association between low maternal education and infant mortality in recent years (S2 Table).
Mediation analysis including PTB and SGA together showed that PTB and SGA played an important role in explaining the link between maternal education and all-cause infant mortality (Table 2; S1 Fig). Compared with high maternal education, the estimated proportion of the total association of education with infant death that could be reduced by “eliminating” the mediating role of PTB and SGA was 55% (MRRPE = 1.27 [95% CI 1.15–1.40, P < 0.001]) for low education and 60% (MRRPE = 1.11 [95% CI 1.01–1.22, P = 0.037]) for medium education. Regarding neonatal mortality, excess deaths were mainly due to the pathway involving PTB and SGA (low versus high education: proportion eliminated = 75%, MRRPE = 1.37 [95% CI 1.21–1.56, P < 0.001]; medium versus high education: proportion eliminated = 81%, MRRPE = 1.14 [95% CI 1.01–1.28, P = 0.032]). During the postneonatal period, other pathways, rather than these mediators, may play a major role in explaining the increased rate of deaths (low versus high education: proportion eliminated = 23%, MRRPE = 1.11 [95% CI 0.95–1.29, P = 0.172]; medium versus high education: proportion eliminated = 21%, MRRPE = 1.04 [95% CI 0.89–1.21, P = 0.614]). Regarding the mediation analyses including one mediator at a time (Table 2; S1 Fig), the mediating role of PTB (low versus high education: proportion eliminated = 46%, MRRPE = 1.22 [95% CI 1.10–1.34, P < 0.001]; medium versus high education: proportion eliminated = 48%, MRRPE = 1.08 [95% CI 0.99–1.19, P = 0.094]) in the association between maternal education and infant mortality was much stronger than that of SGA (low versus high education: proportion eliminated = 11%, MRRPE = 1.05 [95% CI 0.95–1.15, P = 0.377]; medium versus high education: proportion eliminated = 13%, MRRPE = 1.02 [95% CI 0.93–1.12, P = 0.662]). Similar patterns were also found for neonatal mortality and postneonatal mortality analyzed separately. A finer categorization of the mediators yielded a stronger mediating impact by PTB and SGA (Table 3).
Mediation analyses stratified by birth year yielded results similar to those from the main analyses, although the estimated proportion of the total association between maternal education and infant death reduced by eliminating the mediating roles of PTB and SGA seemed to decrease in recent years (S2 and S3 Tables). Analyses restricted to non-SGA infants on the mediating role of PTB found a similar pattern as the analyses considering PTB and SGA together (S4 Table). We found that SGA weakly mediated the association between maternal education and infant mortality among term-born infants (S5 Table).
Deaths due to diseases accounted for 96.8% (8,292 died) of all infant deaths, and the pattern observed for infant deaths due to diseases was similar to that of all-cause mortality (Table 4). Similar patterns were also found for deaths due to certain conditions originating in the perinatal period and congenital malformations (S6 and S7 Tables). Especially for infant deaths due to certain conditions originating in the perinatal period, a substantial mediating impact through PTB and SGA together was found (low versus high education: proportion mediated = 76%, MRRPE = 1.49 [95% CI 1.26–1.76, P < 0.001]). A much smaller portion of deaths due to external causes was found to be mediated by PTB and SGA (S8 Table).
In addition, we performed sensitivity analyses to assess the potential impact of unmeasured confounding. Under the simplified assumption that the prevalence of U was the same in each maternal education group, the CDE was unaffected. However, if the prevalence of U was assumed to differ between maternal education groups, the adjusted CDE was higher than the original CDE. Even if the unmeasured confounder was strong enough to increase the likelihood of infant death by 2.5-fold, we still observed the mediating impact of PTB on neonatal mortality (S9 Table). An additional adjustment for maternal country of origin (S10 Table) did not essentially change the results. The results from the traditional approach with and without mediators for mediation analysis differed from those of our MSM approach (S11 Table).
We found that low maternal education was associated with a higher risk of infant mortality in Denmark, and this association was partly mediated by PTB and SGA. The mediatory role of PTB and SGA was substantial in neonatal mortality, while a large direct impact of low maternal education on postneonatal mortality was observed. The mediating effects through PTB were greater than those through SGA.
A number of studies [3,5,9,10,32] have investigated the association between maternal socioeconomic disadvantage and infant mortality, but few of them have attempted to distinguish the relative roles of LBW, PTB, and SGA in the association. A British study [15], using a traditional method to explore the association between social class and infant mortality (“with and without mediator”), reported that LBW was a strong risk factor in the neonatal period but did not seem to play an important role in the postneonatal period. The traditional regression approach, including a potential mediator as a covariate, is easily implemented and understood. However, the traditional mediation analysis method is prone to yielding a flawed conclusion due to exposure–mediator interaction, mediator–outcome confounding, and mediator–outcome confounding affected by the exposure (intermediate confounding) [23,28]. Especially in the context of longitudinal design and time-varying confounders, intermediate confounders may not be rare. In a traditional mediation approach, adjustment for the mediator might lead to a spurious association between the intermediate confounder and the exposure, where the intermediate confounder becomes the confounder between the exposure and the outcome. Adjustment for intermediate confounders is required to prevent such bias, known as collider-stratification bias [23]. However, adjustment for intermediate confounders in a traditional mediation model could block part of the effect of exposure on outcome through the intermediate confounder, resulting in underestimation of the direct effect not through the mediator and overestimation of the mediating effect through the mediator [23,33]. Our findings also suggest that the traditional approach for mediation analysis is probably subject to bias (S11 Table). We used a counterfactual approach for causal mediation analysis that allows decomposing the TE into the CDE and the PE in the presence of exposure–mediator interaction [23]. Inverse-probability-weighted MSMs could address intermediate confounders and estimate the CDE, without blocking the pathway from the exposure to the outcome acting through the intermediate confounders (see S2 Text for details) [23,26]. Moreover, sensitivity analysis techniques based on a causal inference framework can be applied to evaluate the impact of unmeasured confounders [29].
Low socioeconomic status has been associated with PTB, SGA, and LBW [12–14,34,35]. Socioeconomic inequality may lead to differences in prenatal risk factors for PTB and SGA, including the accessibility/quality of maternity/perinatal care, maternal health behaviors, occupational situation, nutrition, and health outcomes [2,35–37]. These prenatal risk factors may have an influence on risks of PTB and fetal growth restriction, thereby leading to an increased risk of infant death [36]. Certain conditions originating in the perinatal period and congenital malformations are the 2 leading causes of infant death [38]. In mediation analyses of infant mortality due to certain conditions originating in the perinatal period and congenital malformations, we found similar mediating patterns as in the mediation analyses for overall infant mortality. PTB, SGA, and LBW are important indirect causes of neonatal deaths [12,36,39], and our results also showed that the mediating roles of PTB and SGA accounted for a large number of excess neonatal deaths related to low maternal education.
In accordance with previous studies, we observed that the total association of maternal education and infant mortality tended to be stronger in the postneonatal period than in the neonatal period [5,32]. Our finding that both the CDE and PE of maternal education on infant mortality differed between the neonatal and the postneonatal period may reflect separate causal pathways. The substantial impact of PTB and SGA on the association of maternal education with neonatal mortality due to diseases may reflect that the excess neonatal deaths are more influenced by prenatal and perinatal mechanisms, like PTB, placental impairment, and impaired fetal growth. Although neonatal mortality has been linked to the quality of obstetric and neonatal care [31], it is less likely that maternal disadvantage would significantly affect the care provided by the neonatal intensive care center in Denmark, where there is universal tax-paid health coverage and the utilization of these services is very high. Disorders related to short gestation and impaired fetal growth are important causes of neonatal death [8], and might partly explain why maternal education mainly acted through the mediating role of PTB and SGA.
In the postneonatal period, substantial direct impact of maternal education on offspring mortality was seen. Although Denmark has universal health coverage for all essential health services, education could reflect differences in both economic and non-economic factors. Higher levels of education are related to active acquisition of health-related knowledge and increased use of special health services. Education also indicates the ability to solve problems and the capacity to deal with stressors [40]. Infants with well-educated mothers are more likely to benefit from optimized use of health and social welfare services and to receive a higher quality of clinical care [31]. A previous study showed that children born to mothers with lower education used fewer general practitioner services and specialist services, and there were considerable differences in the use of telephone consultations with doctors according to maternal education [41]. Therefore, socioeconomic postnatal differences in the environmental and social circumstances of infants are likely to be a vital determinant of postneonatal mortality [5]. It is important to further evaluate this finding in other countries with larger gaps in the quality of and access to postnatal care.
From the public health perspective, the findings from this study improve our understanding of the underlying pathways from maternal education to infant mortality, which should be taken into consideration when designing preventative strategies. The overall educational disparities in infant mortality could possibly be reduced by tackling different intermediates along the pathways. First of all, it is critical to understand and identify socially related prenatal risk factors of PTB and restricted fetal growth. Poor maternal health has been suggested to be strongly associated with adverse birth characteristics [42]. Intervention strategies focused on improving the health of socially and economically disadvantaged women before and during pregnancy to reduce the risk of PTB and SGA may help to prevent neonatal mortality [31,42]. For postneonatal mortality, we found that pathways other than PTB and SGA may play a more critical role. Educational attainment of women is crucial for their access to labor market, income, social resources, financial status, and health behaviors [43]. Education level may have a direct impact on infant health through housing conditions, occupational status, lifestyle, diet, psychosocial stress, compliance with medical advice, the quality of medical care, and the use of specialized medical care [44]. Therefore, efforts to reduce postneonatal deaths need to minimize the link between maternal education and such factors [45]. Reducing these disparities in infant care will require a significant and coordinated effort from different sectors, such as health, housing, labor, and education. Even in a welfare society such as Denmark, pregnant women with low education need more attention and resources to address socially related risk factors to improve infant health in general, and reduce infant mortality.
The unique methodological strength of this study is that we have sufficient and good quality data for performing modern causal mediation analysis in a large study population. Causal mediation analysis can help to explain the mechanisms behind the impact of maternal socioeconomic disadvantage on infant mortality and inform the process of designing public health interventions that will prevent infant deaths. Our study estimated the joint mediating effect of PTB and SGA using the weighting-based approach for causal mediation analysis, which does not necessarily require knowing the ordering of the mediators and is able to address the possibility that the mediators affect each other. It allowed for possible mediator–exposure and mediator–mediator interactions. Furthermore, our large population-based cohort study with almost complete follow-up provided high-quality prospective data and minimized the potential influence of selection bias and recall bias.
The results of the study should be carefully interpreted given the following limitations. First, for valid inferences, causal mediation analysis requires the untestable assumption of there being no unmeasured confounders of the association between infant mortality, maternal education, and the mediators [23,28,33]. Although we adjusted for a wide range of confounders, we cannot exclude residual confounding by unmeasured maternal demographic and lifestyle factors, such as alcohol consumption, body mass index, physical inactivity, psychological stress in adolescence, or school attendance. We applied sensitivity analysis [23,29] to assess the potential impact of unmeasured mediator–outcome confounders on the main results. The results from sensitivity analysis indicated that fairly substantial confounding would be required to explain away our reported results. In addition, some of these unmeasured variables can also be considered as mediators in the pathway between maternal education and infant mortality, in which case they should not be included in the model. Second, ethnicity could be a confounder for both exposure–outcome association and exposure–mediator association. In the US and several European countries, ethnic minority groups have consistently lower birthweight than the predominant ethnic group. Ethnic disparity is likely due to variations in access to antenatal care and socio-environmental factors, such as workload, stress, and diet [46–48]. However, ethnic disparity is unlikely to be a major concern in Denmark as 91.8% of Danish women are of Danish descent [49], and analyses with additional control for country of origin of the mother (>90% of mothers originally from Denmark [50]) did not change the results. Third, average maternal education levels have increased between 1981 and 2015 in Denmark. The women with only primary and lower secondary education may have become a highly selected group at the end of the period due to increased average education level. Therefore, the stronger association between maternal education and infant mortality observed in recent years could be due to a higher prevalence of risk factors for infant mortality among women with low education. In addition, the same educational attainment does not necessarily reflect the same classifying functions in different calendar periods. The improved educational attainment over the past decades may influence the associations observed in the study. However, mediation analyses stratified by birth year interval found that the results in different intervals were similar to those from the main analyses using data during the whole study period. Finally, our study is based on register data from Denmark, which has universal health coverage for all essential healthcare services. It is important to evaluate these findings in countries with larger disparities in terms of access and the quality of perinatal healthcare.
The association between low maternal education and infant mortality mediated through PTB and SGA was large for neonatal mortality but small for postneonatal mortality. Public health preventive strategies for education-related neonatal mortality in high-income countries may need to address the socially related prenatal risk factors of PTB and impaired fetal growth. On the other hand, the substantial direct association of maternal education with deaths not accounted for by PTB and SGA during the postneonatal period could reflect unaddressed educational disparities in infant care or other factors.
|
10.1371/journal.pcbi.1000547 | Mechanical Strength of 17 134 Model Proteins and Cysteine Slipknots | A new theoretical survey of proteins' resistance to constant speed stretching is performed for a set of 17 134 proteins as described by a structure-based model. The proteins selected have no gaps in their structure determination and consist of no more than 250 amino acids. Our previous studies have dealt with 7510 proteins of no more than 150 amino acids. The proteins are ranked according to the strength of the resistance. Most of the predicted top-strength proteins have not yet been studied experimentally. Architectures and folds which are likely to yield large forces are identified. New types of potent force clamps are discovered. They involve disulphide bridges and, in particular, cysteine slipknots. An effective energy parameter of the model is estimated by comparing the theoretical data on characteristic forces to the corresponding experimental values combined with an extrapolation of the theoretical data to the experimental pulling speeds. These studies provide guidance for future experiments on single molecule manipulation and should lead to selection of proteins for applications. A new class of proteins, involving cystein slipknots, is identified as one that is expected to lead to the strongest force clamps known. This class is characterized through molecular dynamics simulations.
| The advances in nanotechnology have allowed for manipulation of single biomolecules and determination of their elastic properties. Titin was among the first proteins studied in this way. Its unravelling by stretching requires a 204 pN force. The resistance to stretching comes mostly from a localized region known as a force clamp. In titin, the force clamp is simple as it is formed by two parallel β-strands that are sheared on pulling. Studies of a set of under a hundred proteins accomplished in the last decade have revealed a variety of the force clamps that lead to forces ranging from under 20 pN to about 500 pN. This set comprises only a tiny fraction of proteins known. Thus one needs guidance as to what proteins should be considered for specific mechanical properties. Such a guidance is provided here through simulations within simplified coarse-grained models on 17 134 proteins that are stretched at constant speed. We correlate their unravelling forces with two structure classification schemes. We identify proteins with large resistance to unravelling and characterize their force clamps. Quite a few top strength proteins owe their sturdiness to a new type of the force clamp: the cystein slipknot in which the force peak is due to dragging of a piece of the backbone through a closed ring formed by two other pieces of the backbone and two connecting disulphide bonds.
| Atomic force microscopy, optical tweezers, and other tools of nanotechnology have enabled induction and monitoring of large conformational changes in biomolecules. Such studies are performed to assess structure of the biomolecules, their elastic properties, and ability to act as nanomachines in a cell. Stretching studies of proteins [1] are of a particular current interest and they have been performed for under a hundred of systems. Interpretation of some of these experiments has been helped by all-atom simulations, such as reported in refs. [2],[3]. They are limited by of order 100 ns time scales and thus require using unrealistically large constant pulling speeds. However, they often elucidate the nature of the force clamp – the region responsible for the largest force of resistance to pulling, . All of the experimental and all-atom simulational studies address merely a tiny fraction of proteins that are stored in the Protein Data Bank (PDB) [4]. Thus it appears worthwhile to consider a large set of proteins and determine their within an approximate model that allows for fast and yet reasonably accurate calculations. Structure-based models of proteins, as pioneered by Go and his collaborators [5] and used in several implementations [6]–[13], seem to be suited to this task especially well since they are defined in terms of the native structures away from which stretching is imposed.
There are many ways, all phenomenological, to construct a structure-based model of a protein. 504 of possible variants are enumerated and 62 are studied in details in ref. [14]. The variants differ by the choice of effective potentials, nature of the local backbone stiffness, energy-related parameters, and of the coarse-grained degrees of freedom. The most crucial choice relates to making a decision about which interactions between amino acids count as native contacts. Comparing to the corresponding experimental values in 36 available cases selects several optimal models [14]. Among them, there is one which is very simple and which describes a protein in terms of its atoms, as labeled by the sequential index . This model is denoted by which stands for, respectively, the Lennard-Jones native contact potentials, local backbone stiffness represented by harmonic terms that favor the native values of local chiralities, the contact map in which there are no contacts, and the amplitude of the Lennard-Jones potential, , is uniform. The contact map is determined by assigning the van der Waals spheres to the heavy atoms (enlarged by a factor to account for attraction) and by checking whether spheres belonging to different amino acids overlap in the native state [15],[16]. If they do, a contact is declared as native. Non-native contacts are considered repulsive. Application of this criterion frequently selects the contacts as native. If the contact map includes these contacts the resulting model will be denoted here as . On average, it performs worse than because the contacts usually correspond to the weak van der Waals couplings as can be demonstrated in a sample of proteins by using a software [17] which analyses atomic configurations from the chemical perspective on molecular bonds. Thus the couplings should better be removed from the contact map (in most cases).
The survey to determine in 7510 model proteins with the number of amino acids, , not exceeding 150 and 239 longer proteins (with up to 851) has been accomplished twice. First within the model [18] and soon afterwords within the model [19]. The first survey also comes with many details of the methodology whereas the second just presents the outcomes. The two surveys are compared in more details in refs. [14],[20]. The results differ, particularly when it comes to ranking of the proteins according to the value of , but they mutually provide the error bars on the findings. They both agree, however, on predicting that there are many proteins whose strength should be considerably larger than the frequently studied benchmark – the sarcomere protein titin ( of order 204 pN [21],[22]). Near the top of the list, there is the scaffoldin protein c7A (the PDB code 1aoh) which has been recently measured to have of about 480 pN [23]. Other findings include establishing correlations with the CATH hierarchical classification scheme [24],[25], such as that there are no strong proteins, and identification of several types of the force clamps. The large forces most commonly originate in parallel that are sheared [26]. However, there are also clamps with antiparallel , unstructured strands, and other kinds.
The two surveys have been based on the structure download made on July 26, 2005 when the PDB comprised 29 385 entries. Many of them correspond to nucleic acids, complexes with nucleic acids and with other proteins, carbohydrates, or come with incomplete files and hence the much smaller number of proteins that could be used in the molecular dynamics studies. Here, we present results of still another survey which is based on a download of December 18, 2008 which contains 54 807 structure files and leads to 17 134 acceptable structures with not exceeding 250 (instead of 150). These structures are then analyzed through simulations based on the model. The numerical code has been improved to allow for acceleration of calculations by a factor of 2.
The 190 structures (or 1.1% of all structure considered) with the top values of in units of are shown in Table 1 (the first 81 entries for which ) and Table S1 of the SI (proteins ranked 82 through 190), together with the values of titin (1tit) and ubiquitin (1ubq) to provide a scale. As argued in the Materials and Methods section section, the unit of force, , is now estimated to be of order 110 pN. All of the corresponding proteins are predicted to be much stronger than titin and none but two of them (1aho, 1g1k [23]) have been studied experimentally yet. In addition to the types of force clamps identified before, we have discovered two new mechanisms of sturdiness. One of them involves a cysteine slipknot (CSK) and is found to be operational in all of the 13 top strength proteins. In this motif, a slip-loop is pulled out of a cysteine knot-loop. Another involves dragging of a single fragment of the main chain across a cysteine knot-loop. The two mechanisms are similar in spirit since both involve dragging of the backbone. However, in the CSK case, two fragments of the backbone are participating.
We make a more systematic identification of the CATH-classified architectures that are linked to mechanical strength and then analyze correlations of the data to the SCOP-based grouping (version 1.73) [27]–[29]. The previous surveys did not relate to the SCOP scheme.
We identify the CATH-based architectures and SCOP-based folds that are associated with the occurrence of a strong resistance to pulling. A general observation, however, is that each such group of structures may also include examples of proteins that unravel easily. The dynamics of a protein are very sensitive to mechanical details that are largely captured by the contact map and not just by the appearance of a structure. On the other hand, if one were to look for mechanically strong proteins then the architectures and folds identified by us should provide a good starting point. We also study the dependence of on the pulling velocity and characterize the dependence on through distributions of the forces.
The current third survey has been performed within the same model as the second survey [19]. However, we reuse and extend it here because the editors of Biophysical Journal retracted the second survey [30]. All of the values of are deposited at the website www.ifpan.edu.pl/BSDB (Biomolecule Stretching Database) and can by accessed by through the PDB structure code.
The distribution of all values of for the full set of proteins is shown in Figure 1. Despite the larger limit on now allowed, the distribution is rather similar to that obtained in ref. [19] for the smaller number of proteins (and with the smaller sizes). The similarity is primarily due to the fact that the size related effects, discussed below, are countered by new types of proteins that are now incorporated into the survey. The distribution is peaked around of which constitutes about 60% of the strength associated with titin. The distribution is non-Gaussian: it has a zero-force peak and a long force tail. The zero-force peak arises in some proteins with the covalent disulphide bonds. In the model, such bonds are represented by strong harmonic bonds. Stretching of such a protein may not result in any force peak before a disulphide bond gets stretched indefinitely and hence is considered to be vanishing then. The tail, on the other hand, corresponds to the strong proteins. The top strongest 1.1% of all proteins are listed in Tables 1 (in the main text) and S1 (in the SI).
The insets of Figure 1 show similar distributions for proteins belonging to the particular CATH-based classes. There are four such classes: , , and proteins with no apparent secondary structures. It is seen that none of the 3240 proteins exceeds the peak force obtained for titin within our model. This observation is in agreement with experiments on several proteins that are listed in the Materials and Methods section. All strong proteins are seen to involve the . The peak in the probability distribution for the proteins is observed to be shifted towards the bigger values of compared to the one for the proteins. At the same time, the high force tail of the distribution for the proteins is substantially more populated than the corresponding tail for the proteins.
Figure 2 is similar to Figure 1 in spirit, but now the structures are split into particular ranges of the protein sizes: between 40 and 100 (the dotted line), between 100 and 150 (thin solid line), and between 200 and 250 (the thick solid line). The curve for the range from 150 to 200 is in-between the curves corresponding to neighboring ranges and is not shown in order not to crowd the Figure. The distributions are seen to be shifting to the right when increasing the range of the values of indicating, that the bigger the number of amino acids, the more likely a protein is to have a large value of . This observation holds for all classes of the proteins, as evidenced by the insets in Figure 2.
In most cases, the major force peak arises at the begining of stretching where the Go-like model should be applicable most adequately. One can characterize the location of during the stretching process by a dimensionless parameter which is defined in terms of the end-to-end distance, as spelled out in the caption of Table 1. This parameter is equal to 0 in the native state and to 1 in the fully extended state. In 25% of the proteins studied in this survey, was less than 0.25 and in 52% – les than 0.5. There are very few proteins with exceeding 0.8.
Table 1 does not include any (non-cysteine-based) knotted proteins. The full list of 17 134 proteins contains 42 such proteins but they come with moderate values of . However, knotted proteins with may turn out to have different properties.
A convenient way to learn about the biological properties listed in Tables 1 and S1 is through the Gene Ontology data base [31] which links such properties with the PDB structure codes. The properties are divided into three domains. The first of these is “molecular function” which describes a molecular function of a gene product. The second is “biological processes” and it covers sets of molecular events that have well defined initial and final stages. The third is “cellular component” and it specifies a place where a given gene product is most likely to act.
The results of our findings are summarised in Table 2. It can be seen, that most of the 190 strongest proteins are likely to be found in an extracellular space where conditions are much more reducing than within cells. Larger mechanical stability is advantageous under such conditions. 90 out of the strongest proteins exhibit hydrolase activity. 39 of these 90 are serine-type endopeptidases. These findings seem to be consistent with expectations regarding proteins endowed with high mechanical stability. For instance, proteases, which are well represented in Table 2 should be more stable to prevent self-cleavage.
The classification of proteins within the CATH (Class, Architecture, Topology, Homology) data base is done semi-automatically by applying numerical algorithms to structures that are resolved better than within 4 Å [24],[25]. The four classes of proteins in the CATH system are split into architectures, depending on the overall spatial arrangement of the secondary structures, the numbers of in various motifs, and the like. The next finer step in this hierarchical scheme is into topologies and it involves counting contacts between amino acids which are sequentially separated by more than a treshold. The further divisions into homologous superfamilies and then sequence family levels involve studies of the sequential identity.
We have found that only six architectures contribute to larger than . These are ribbons – 2.10 (41.8% of the proteins listed in Table 1), – 2.40 (8.9%), – 2.60 (16.3%), – 3.10 (5.4%), 3-layer (aba) sandwiches – 3.40 (5.4%), and these with no CATH classification to date (21.8%). The corresponding distributions of forces are shown in the top six panels of Figure 3 and the topologies involved are listed and named in Table 3.
Examples of architectures that are dominant contributors to a low force behavior are the orthogonal bundle (the right bottom panel of Figure 3), the up-down bundle, and the (the left bottom panel of Figure 3).
The SCOP (Structural Classification of Proteins) data base [27]–[29] is curated manually and it relies on making comparisons to other structures through a visual inspection. This classification scheme is also hierarchical and the broadest division is into seven classes and three quasi-classes. The classes are labelled through and these are as follows: mainly (), mainly (), which groups proteins in which helices and are interlaced (), with the helices and grouped into clusters that are separated spatially (), multidomain proteins (), membrane and cell-surface proteins (), and small proteins that are dominated by disulphide bridges or the heme metal ligands (). The quasi-classes are labelled through and they comprise coiled-coil proteins (), structures with low resolution (), and peptides and short fragments (). The classes are then partitioned into folds that share spatial arrangement of secondary structures and the nature of their topological interlinking. Folds are then divided into superfamilies (same fold but small sequence identity) and then families (two proteins are said to belong to the same family if their sequence identity is at least 30%). Families are then divided into proteins – a category that groups similar structures that are linked to a similar function. Proteins comprise various protein species.
Each structure assignment comes with an alphanumeric label, as shown in Tables 1, S1, and 4 which reflects the placement in the hierarchy. At the time of our download, there have been 92 972 entries in the SCOP data base that are assigned to 34 495 PDB structures. These entries are divided into 3464 families, 1777 superfamilies and 1086 unique folds. A given structure may have several entry labels but the dominant assignment is listed first. We use the primary assignment in our studies. The same rule is also applied to the CATH-based codes.
Figure 4 shows the distributions of forces for the SCOP-based classes of proteins. The results are consistent with the CATH-based classes since the class of CATH basically encompasses the and classes of SCOP. However, there are proteins which are classified only according to one of the two schemes. Thus there are 4431 proteins out of which only the total of 3368 is SCOP-classified as belonging to the and classes. At the same time, the total of the proteins in the and classes we have is 4795.
It should be noted that the peak in the distribution for is shifted to higher forces by about from the peak for . At the same time, the zero-force peak is virtually absent in . The SCOP-based classification also reveals that its class contributes across the full range of forces and, in particular, it may lead to large values of . It should be noted, as also evidenced by Table 1, that there is a substantial number of strong proteins that has no class assignment.
Figures 5 and 6 refer to the distributions of across specific folds. The first of these presents results for the folds that give rise to the largest forces. The names of such folds are specified in Figure 5. The percentage-wise assessment of the folds contributing to big forces is presented in Table 4. The top contributor is found to be the b.47 fold (SMAD/FHA domain). Figure 6 gives examples of folds that typically yield low forces.
It is interesting to note that distributions corresponding to some folds are distinctively bimodal, as in the case of the SMAD/FHA fold (b.47). This particular fold is dominated by SMAD3 MH2 domain (b.47.1.2; 352 structures) which contributes both to the high and low force peaks in the distribution. The remaining domains (b.47.1.1, b47.1.3, and b47.1.4) contribute only to the low force peak. The dynamical bimodality of the b.47.1.2 fold can be ascribed to the fact that the strong subset comes with one extra disulphide bond relative to the weak subset. This extra bond provides substantial additional mechanical stability when stretching is accomplished by the termini. We illustrate sources of this bimodality in the SI (Figure S1) for two proteins from this fold: 1bra which is strong and 1elc which is weak. In ref. [18], we have noted that various sets of proteins with identical CATH codes (e.g., 3.10.10) may give rise to bimodal distributions without any dynamical involvement of the disulphide bonds. The reason for this is that even though the contact maps for the two modes are similar, the weaker subset misses certain longer ranged contacts which pin the structure. Mechanical stability is more sensitive to structural and dynamical details than are not provided by standard structural descriptors.
This surveys identifies a host of proteins that are likely to be sturdy mechanically. Many of them involve disulphide bridges which bring about entanglements that are complicated topologically such as CSKs and CKs. The distinction between the two is that the former can depart from its native conformation and the latter cannot.
Our survey made use of a coarse grained model so it would be interesting to reinvestigate some of the proteins identified here by all-atom simulations, especially in situations when the CSK is involved. The CSK motifs may reveal different mechanical properties when studied in a more realistic model. Of course, a decisive judgment should be provided by experiment.
The very high mechanical resistance of the CSK proteins should help one to understand their biological function. The superfamily of cysteine-knot cytokines (in class small proteins and fold cystein-knot cytokines) includes families of the transforming growth-factor and the polypeptide vascular endothelial growth factors (VEGFs) [49],[50]. The various members of this superfamily, listed in Table 5, have distinct biological functions. For instance, VEGF-B proteins which regulate the blood vessel and limphatic angiogenesis bind only to one receptor of tyrosine kinase VEGFR-1. On the other hand, VEGF-A proteins bind to two receptors VEGFR-1 and VEGFR-2. All of these proteins form a dimer structure. The members of this familly are endowed with remarkably similar monomer structures but differ in their mode of dimerisation and thus in their propensity to bind ligands. Additionally, all dimers posses almost the same a cyclic arrangement of cysteine residues which are involved in both intra- and inter-chain disulphide bonds. These inter-chain disulphide bonds create the knot and slip-loops, where the intra-chain disulphide bonds give rise to a CSK motif when the slip-loop is gets dragged acrros the knot-loop upon pulling.
It has been shown experimentally [51] that such cysteine related connectivities bring the key residues involved in receptor recognition into close proximity of each other. They also provide a primary source of stability of the monomers due to the lack of other hydrogen bonds between two beta strands at the dimer interface.
The non trvial topologial connection between the monomers allow for mechanical separation of two monomers by a distance of about half of the size of the slip-loop. Our results suggest, however, that the force needed for the separation may be too high to arise in the cell.
The input to the dynamical modeling is provided by a PDB-based structures. The structure files may often contain several chains. In this case, we consider only the first chain that is present in the PDB file. Likewise, the first NMR determined structure is considered. If a protein consists of several domains, we consider only the first of them.
The modeling cannot be accomplished if a structure has regions or strings of residues which are not sufficiently resolved experimentally. Essentially all structure-disjoint proteins have been excluded for our studies. Exceptions were made for the experimentally studied scaffoldin 1aoh and for proteins in which small defects in the established structure (such as missing side groups) were confined within cystein loops and were thus irrelevant dynamically. In these situations, the missing contacts have been added by a distance based criterion [23] in which the treshold was set at 7.5 Å. Among the test used to weed out inadequate structures involved determining distances between the consecutive atoms. A structure was rejected if these distances were found to be outside of the range of 3.6–3.95 Å. The exception was made for prolines, which in its native state can accommodate the cis conformation. In that case, the distance between a proline and its subsequent amino acid usually falls in the range between 2.8 and 3.85 Å. For a small group of proteins which slipped through our structure quality checking procedure, but were found to be easily fixed (e.g. 1f5f, 1fy8, and 2f3c), we used publicly avialable software BBQ [52] to rebuild locations of the missing residues. A limited accuracy of this prediction procedure seems to be adequate for our model due to its the coarse-grained nature.
The modeling of dynamics follows our previous implementations [11],[12],[18] within model except that the contact map is as in ref. [19], i.e. with the contacts excluded. There is also a difference in description of the disulphide bonds. In refs. [14],[19] they were treated as an order-of-magnitude enhancement of the Lennard-Jones contacts in all proteins. In ref. [18] the different treatment of the disulphide bonds was applied to the proteins that were found to be strong mechanically without any enhancements. Here, on the other hand, we consider such bonds as harmonic in all proteins, in analogy to the backbone links between the consecutive . The native contacts are described by the Lennard-Jones potential , where is the distance between the in amino acids and whereas is determined pair-by-pair so that the minimum in the potential is located at the experimentally established native distance. The non-native contacts are repulsive below of 4 Å.
The implicit solvent is described by the Langevin noise and damping terms. The amplitude of the noise is controlled by the temperature, . All simulations were done at , where is the Boltzmann constant. Newton's equations of motion are solved by the fifth order predictor-corrector algorithm. The model is considered in the overdamped limit so that the characteristic time scale, , is of order 1 ns as argued in refs. [6],[53]. Stretching is implemented by attaching an elastic spring to two amino acids. The spring constant used has a value of which is close to the elasticity of experimental cantilevers. One of the springs is anchored and the other spring is moving with a constant speed, . Choices in the value of the spring constant have been found to affect the look of the force-displacements patterns and thus the location of the transition state [54],[55], but not the values of [10],[12],[18].
The dependence on is protein-dependent and it is approximately logarithmic in as evidenced by Figure 11 for several strong proteins. The logarithmic dependence has been demonstrated experimentally, for instance, for polyubiquitin [56],[57]. . The approximate validity of this relationship is demonstrated in Figure 11 for three proteins with big values of . We observe that the larger the value of , the bigger probability that the dependence on is large. When we make a fit to for 1vpf, 1c4p, and 1j8s, we get the parameter to be equal to respectively (the values of are correspondingly). However, some strong proteins may have to be as low as 0.04.
When making the survey, we have used of and stretching was accomplished by attaching the springs to the terminal amino acids (there is an astronomical number of other choices of the attachment points).
In order to estimate an effective experimental value of the energy parameter , we have correlated the theoretical values of with those obtained experimentally. The experimental data points used in ref. [14] have been augmented by entries pertaining to 1emb (117–182), 1emb (182–212) [58] (where the numbers in brackets indicate the amino acids that are pulled) and 1aoh, 1g1k, and 1amu [23]. The full list of the experimental entries is provided by Table 6. Unlike the previous plots [14] that cross correlate the experimental and theoretical values of , we now extrapolate the theoretical forces to the values that should be measured at the pulling speeds that are used experimentally. We assume that the unit of speed, , is of order 1 Å/ns and consider 10 speeds to make a fit to the logarithmic relationship. The values of parameters and for the proteins studied experimentally are listed in Table 6.
The main panel of Figure 12 demonstrates the relationship between the extrapolated theoretical and experimental values of . The best slope, indicated by the solid line, corresponds to the slope of 0.0091. The inverse of this slope yields 110 pN as an effective equivalent of the theoretical force unit of . The Pearson correlation coefficient, is 0.832, the rms percent error, , is 1.02, and the Theil coefficient (discussed in ref. [14]) is 0.281. The inset show a similar plot obtained when the extrapolation to the experimental speeds is not done. The resulting unit of the force would be equivalent to 110 pN which differs form the previous estimate of 71 pN (shown by the dotted line in the main panel) because of the inclusion of the newly measured proteins and implementation of the extrapolation procedure. The statistical measures of error here are . These measures are better compared to the case with the extrapolation because the extrapolation procedure itself brings in additional uncertainties. Nevertheless, implementing the procedure seems sounder physically. The spread between these various effective units of the force suggests an error bar of order 30 pN on the currently best value of 110 pN.
|
10.1371/journal.pgen.1003972 | Whole Genome, Whole Population Sequencing Reveals That Loss of Signaling Networks Is the Major Adaptive Strategy in a Constant Environment | Molecular signaling networks are ubiquitous across life and likely evolved to allow organisms to sense and respond to environmental change in dynamic environments. Few examples exist regarding the dispensability of signaling networks, and it remains unclear whether they are an essential feature of a highly adapted biological system. Here, we show that signaling network function carries a fitness cost in yeast evolving in a constant environment. We performed whole-genome, whole-population Illumina sequencing on replicate evolution experiments and find the major theme of adaptive evolution in a constant environment is the disruption of signaling networks responsible for regulating the response to environmental perturbations. Over half of all identified mutations occurred in three major signaling networks that regulate growth control: glucose signaling, Ras/cAMP/PKA and HOG. This results in a loss of environmental sensitivity that is reproducible across experiments. However, adaptive clones show reduced viability under starvation conditions, demonstrating an evolutionary tradeoff. These mutations are beneficial in an environment with a constant and predictable nutrient supply, likely because they result in constitutive growth, but reduce fitness in an environment where nutrient supply is not constant. Our results are a clear example of the myopic nature of evolution: a loss of environmental sensitivity in a constant environment is adaptive in the short term, but maladaptive should the environment change.
| When a population of organisms is faced with a selective pressure, such as a limiting nutrient, mutations that arise randomly may confer a fitness benefit on the individual carrying that mutation. If that individual reproduces before it is lost from the population, the frequency of that mutation may increase. Over time, many beneficial mutations will arise in a large population, but there are few high resolution experiments tracking the frequency of such mutations in an evolving population. We evolved populations of the baker's yeast in a constant environment in the presence of limiting amounts of sugar, and then used DNA sequencing to identify mutations that reached at least a 1% frequency in the population, and tracked them over time. We identified 120 mutations over three experiments, and determined that the genes and pathways that had gained beneficial mutations were largely reproducible across experiments, and that many of the mutations led to the loss of signaling pathways that usually sense a changing environment, allowing the cell to respond appropriately. When these mutant cells were faced with uncertain environments, the mutations proved to be deleterious. Environmental sensing must carry a fitness cost in a constant environment, but is essential in a changing one.
| Adaptive evolution optimizes the fitness of organisms for their environment through the accumulation of beneficial mutations by natural selection [1]. While we understand much about the mechanisms by which natural selection operates, less is known about the beneficial mutation rate [2], and the genetic basis of adaptation [3]. Of particular interest is the spectrum of mutations that are adaptive in a specific environment, defined here as “adaptive strategy”. Through the use of experimental evolution, in conjunction with technological innovations such as candidate gene sequencing [4]–[8], cDNA, [9], [10] and tiling microarrays [11], [12], and whole genome sequencing of individual clones [13]–[16] and populations [17], [18], the field has recently gained significant insight into the genetic basis of adaptation. However, while candidate gene sequencing is certainly incomplete (though still instructive) in the picture it provides, even the identification of all mutations in individual clones does not reveal a complete representation of adaptation. Sequencing a handful of selected clones from an experiment provides only a microcosm of the adaptive mutational spectrum, while sequencing many clones from an experiment begins to resample the most prevalent lineages. By contrast, sequencing terminal clones from many different experiments (e.g. [14]) provides deeper insight into the convergence or divergence of the adaptive process, but is unable to capture evolution in action, including the clonal interference that occurs in the typically large populations used in microbial experimental evolution systems. To capture the dynamics of the adaptive process, as well as the mutational spectrum that accompanies it, it is necessary to sequence very large numbers of clones, possibly from many time points during an experiment, or instead to sequence entire populations as they evolve.
In large asexual populations, selection acts positively to increase the frequency of the lineages containing beneficial mutations, while competition between coexisting adaptive lineages reduces the overall rate at which beneficial mutation increase in allele frequency, a process termed clonal interference [19], [20]. Clonal interference occurs when beneficial mutations are sufficiently common to allow multiple adaptive lineages to expand in the population concurrently [21]. By deeply sequencing populations at multiple time points it is possible to not only identify mutations, but to also track the evolutionary dynamics of adapting lineages. Three studies published thus far have performed whole genome sequencing of evolving populations [17], [18], [22], identifying SNPs at as low as 5% allele frequency in the sequenced populations; in the first two of these studies, E. coli were evolved by serial transfer, effectively in a continuously varying environment. The second of these two studies [18], sequenced deeply enough that the allele frequencies of identified mutations over time could be tracked. However, it is likely that at ever-lower allele frequencies, there will be more observable beneficial mutations, most probably with smaller fitness effects. In the third of the studies, 40 replicate yeast populations were propagated by serial transfer for 1,000 generations, and sequenced every 80 or so generations [22], allowing allele frequencies to be able to be determined down to 10% allele frequency. To better enumerate the adaptive strategy under a particular environment and to gain a better quantitative measure of the extent of clonal interference, deeper sequencing is needed however, which will likely identify additional mutations at lower allele frequencies with which to better characterize the adaptive mutational spectrum.
Different environments likely result in different adaptive strategies, and many natural environments are variable and unpredictable, with irregular fluctuations in environmental parameters. Consequently, signaling networks evolved to enable organisms to be able to sense and respond to uncertain environments [23]. Signaling networks are ubiquitous across the Tree of Life, yet the question remains, “are functional signaling networks an essential feature of a well-adapted biological system?” Intracellular symbionts have undergone extensive genome reductions, likely due to relaxed selection in a setting that has few environmental perturbations. A major functional theme in these genome reductions is the loss of genes involved in signaling and genetic regulation [24], [25]. However, this loss is likely neutral gene degradation due to genetic drift rather than adaptive evolutionary processes [26].
We sought to determine if the loss of environmental sensitivity is a viable or indeed preferred adaptive strategy. A constant environment provides an opportunity for such a system to evolve, since environmental sensing is superfluous, and perhaps even carries a fitness cost. We characterized the adaptive strategy, and the dynamics of adaptive lineages of the budding yeast S. cerevisiae evolving in a constant environment by ultra deep genome- and population-wide sequencing of three parallel evolution experiments.
To determine the genetic basis of adaptation and the dynamics of arising mutations, we developed a novel population sequencing protocol, enabling the discovery of mutant alleles as well as their frequencies (Figure S1). We sequenced samples taken every ∼70 generations from three glucose-limited, chemostat-evolved populations of haploid S288c, named E1, E2 and E3, that have been described previously [12], [27]. Libraries were sequenced to 266–1046× coverage, and we employed an overlapping read strategy to reduce the sequencing error rate (Figure S1; Table S1). We also tagged each DNA fragment with a random barcode during library construction (Figure S1), enabling us to distinguish PCR duplicates from fragments that happened to map to the same genomic location; this reduced the apparent number of PCR duplicate reads by 100-fold. Our approach enabled the detection of mutations with an allele frequency as low as 1% (Figure 1), and in total, we discovered 117 mutations across all time points in the three experiments, of which 106 were in coding regions, affecting 51 genes, 19 of which were recurrently mutated. The mutations discovered, as well as their allele frequencies at each timepoint are given in Table S2.
Our population sequencing shows that clonal interference plays a dominant role in all three experiments, as 74 of the identified mutations (63%) decrease in frequency following their maxima, and 42 of these mutations (57%) become extinct by the end of the experiment (Figure 2). These results agree with theoretical predictions [19], [28] and previous observations [29], and imply that even if a mutation rises to a level above our detection threshold, it is still likely to succumb to an expanding fitter lineage and eventually become extinct.
Evolution under conditions that promote clonal interference is also predicted to promote the accumulation of multiple beneficial mutations within an adaptive lineage before the first mutation can sweep [21]. We genotyped clones to determine the linkage of mutations above 10% frequency, and find that 91% of these mutations coexist in a clone with one or more other mutations. This value is an underestimate, since most mutations (67%) never reach 10% frequency and thus were not analyzed for linkage. While from sequencing data alone we cannot unequivocally label a mutation as beneficial versus neutral, recurrent independent mutations (see below and Figure S2) are likely to be beneficial. By this definition, all lineages that we were able to define by genotyping carry at least 1 beneficial mutation (Figure 3). Furthermore, the “winning” lineages occupying the largest proportion of the final population carry at least three beneficial mutations, and at least five mutations total (Figure 3). An exceptional case is E1, where six mutations occur in close succession (four of which are genes that we observe as recurrently mutated) and result in what appears to be a complete selective sweep (Figure 3a). These data indicate that multiple beneficial mutations – often occurring in close succession on what appears to have been a wild-type background – are necessary for a lineage to be successful. However, having multiple mutations is not sufficient for a lineage's success; for example, three lineages in E1, each with two mutations, become extinct due another lineage sweeping (Figure 3a). Furthermore, almost two thirds (49/76) of recurrent, and thus likely beneficial mutations never reach 10% frequency. The dynamics of adaptation suggest the “survival of the luckiest”, where for a new beneficial mutation to reach a high frequency, it must occur on a background that already has multiple other beneficial mutations [29]. This makes predicting the outcome of adaptive evolution difficult since the fixation probability of a beneficial mutation is no longer deterministic and proportional to the selection coefficient, but is also dependent on the genetic background on which the mutation occurs, which is distinctly a chance event. Our data show unequivocally that clonal interference between lineages carrying multiple beneficial mutations defines the dynamics of adaptation.
We sought to understand the adaptive strategy of yeast growing in a constant environment by categorizing the genes in which mutations had occurred. Grouping recurrently mutated genes by pathway, we find that 53% of these mutations across all experiments reside in genes which function in three major signaling pathways: glucose signaling and transport, cyclic adenosine monophosphate/protein kinase A (cAMP/PKA) and the high osmolarity glycerol (HOG) response pathway (Figure 4), and these pathways have statistically enriched GO terms (Table S3). To further characterize the adaptive strategy, we characterized mutations by their predicted consequences. We found that the majority (73%) of mutations are predicted to disrupt protein function, with nonsense mutations being enriched by 7.6-fold (p<2.2e-16) (Figure 5). Together, these data suggest that the general adaptive strategy in a constant environment is the loss of signal transduction pathway function (Figure 4). For the glucose signaling pathway, disruptive mutations in MTH1 and RGT1 lead to constitutive expression of the glucose transporter (HXT) genes [30], [31], which increases the amount of glucose that is able to enter the cell, facilitating growth and providing a selective advantage [27]. The cAMP/PKA pathway positively responds to glucose in wild-type cells leading to growth [32]; disruptive mutations in the three recurrently mutated repressors GPB2, IRA2 and PDE2 would cause constitutive pathway activation and growth, while loss of function in RIM15 (the second most mutated gene, with 7 mutations observed), which is repressed by the PKA pathway, is akin to having increased PKA activity through that downstream path. Rim15 function is involved in the establishment of stationary phase [33] – presumably loss of the ability to enter stationary phase must be beneficial in the constant chemostat environment. The HOG pathway mediates transcriptional response to hyper-osmotic stress and also causes a brief growth arrest [34], so the observed disruptive mutations in pathway activators would be expected to eliminate this response. All five HOG pathway mutations we genotyped occur in lineages with pre-existing MTH1 or RGT1 mutations, (Figure 3a–b), suggesting potential positive epistasis between the HOG and glucose signaling pathways.
To assess the extent of parallel adaptation we examined the overlap in genes and GO terms between experiments. E1, E2 and E3 share 50%, 61% and 21% of their mutated genes with one of the other experiments, with E1 and E2 having the most sharing. MTH1, RIM15 and GPB2 are mutated in all three experiments, with MTH1 being the most frequently observed mutated gene, having 19 independent mutations observed. We grouped enriched GO terms that share edges into GO networks to eliminate redundant GO terms and determined that E1 and E2 share all GO networks with each other (Table S4). E3 has overlap with the other two experiments, with 3 of 6 networks shared with both E1 and E2. The GO network overlap suggests that the replicate experiments followed similar functional trajectories, with the underlying mutations broadly impacting similar biological processes in all experiments, namely the disruption of environmental sensing and signal transduction.
We have shown that loss of environmental sensing through disruptive mutations in signaling pathways is adaptive in a constant environment. As signaling pathways make organisms robust to environmental changes, we hypothesized this loss would have a fitness cost in environments where nutrient availability was not constant. We thus subjected 18 clones containing mutations in one or more signaling pathway to starvation conditions. All 18 clones lost viability more rapidly than wild-type (Figure 6a). To understand which mutations were causing decreased viability, we assayed nine strains containing single mutations from E3 [27] for premature cell death, and found that mutations in or downstream the cAMP/PKA pathway (4 of 9 mutations assayed) showed significantly lower cell viability during starvation (Figure 6b). Of these, we have previously shown that mutations in 3 of these 4 genes are beneficial alone in a glucose limited chemostat [27]. Thus, our results suggest that the adaptive strategy utilized by yeast in the constant chemostat environment is maladaptive in an environment where nutrients are not constant, indicating that there is an evolutionary trade-off due to antagonistic pleiotropy (e.g. see [35]).
We have previously used fluorescent markers to track subpopulations during adaptive evolution in a constant environment [12], and observed clonal interference in each of the 8 experimental populations that we evolved, in concordance with theoretical expectations. In this work, we have greatly expanded upon this, by performing ultra deep whole genome, whole population sequencing at each of 8 timepoints across 3 of these experimental evolutions. In addition to allowing us to identify mutations at an allele frequency as low as 1%, these population sequence data also provide us with direct estimates of the frequency trajectories of the mutations. Of the 3 experiments, only one resulted in a fixation event (E1, where 4 mutations in the same lineage were fixed by the final time point at 448 generations). By contrast, most mutations that enter the population were at a lower frequency than their maximum by the end of the experiment, and indeed more than a third had gone extinct. In most cases, mutations that were subject to clonal interference were in genes that were recurrently mutated (53 out of 74 (72%)), and of those mutations that went extinct, the majority were also in genes that were recurrently mutated (32 out of 42 (76%)). Thus, clonal interference clearly plays a major role in these populations in shaping their eventual composition, with many beneficial mutations in the population going extinct during the evolution. A recent study which also used sequencing of populations undergoing experimental evolution [18] did not observe such a great extent of clonal interference, though in their experiments they only could detect mutations that reached a 5% allele frequency. In our data, of the 74 mutations we detected that were at a lower frequency by the end of the experiment than their maximum frequency (i.e. were subject to clonal interference), 38 (51%) had a maximum frequency of less than 5%. Thus, deeper sequencing is able to provide significantly more insight into the process of clonal interference. We observed 35 mutations in genes that were recurrently mutated that failed to reach a 5% frequency in the experiment, though we only identified 2 additional recurrently mutated genes by being able to get to allele frequencies lower than 5% (OSH3 and LCB3). There were 19 mutations that did not reach an allele frequency of 5% that were in genes that were not recurrently mutated – further experimentation to determine whether these mutations are adaptive, and/or even deeper sequencing would be required to confidently extend the adaptive mutational spectrum. We also observed that multiple mutations prevail, with all of the lineages that we detect as present in our populations at the end of the experiment carrying more than one mutation, with at least two predicted to be beneficial.
It is an open question as to how many lineages with beneficial mutations actually existed within the population – there are few empirical estimates of the beneficial mutation rate, and those that do exist are based on a relatively modest number of observed mutations. One estimate, based on mutations that fixed in Pseudomonas fluoresecens, is 3.8e-8 per cell division [36]. If that were similar to the beneficial mutation rate in yeast, then with a population of 1e9 growing for 448 generations, we might expect as many as 17,000 beneficial mutations to occur within any one of our experiments. Most of these would not be expected to establish – if we assumed that ∼10% establish (roughly similar to an average 10% fitness effect), then 1,700 lineages with beneficial mutations would have established in a given experiment. By contrast, Shaw et al [37], analyzing mutation accumulation lines in A. thaliana, found that approximately half of all mutations observed were beneficial. In yeast, also using mutation accumulation lines, Hall and colleagues have estimated that between 5% and 13% of mutations are beneficial [38]–[40]. With a per base pair mutation rate on the order of ∼1e-10 [41] and a genome size of 12e6, the number of cells estimated per generation to have a mutation is around 1 in 1,000. If 10% of mutations are beneficial, then 1 in 10,000 per generation may receive a beneficial mutation. Thus, in our experiments, we might expect as many as 50 million beneficial mutations to occur over the 448 generation time course, with ∼5 million establishing. While these are estimates based on relatively small number of mutations in mutation accumulation lines, even if they are over estimated by 2 orders of magnitude, it is clear that sequencing of even hundreds of randomly selected individual clones (which will likely represent a few, prevalent lineages), or even deep population sequencing will not be able to fully characterize the spectrum of beneficial mutations, nor determine an accurate estimate of their fitness effects. While to our knowledge this study is the deepest sequencing yet performed on experimentally evolving populations, it may only represent the tip of the potential adaptive iceberg (though this is likely the most important part, as these mutations likely drive the evolutionary process), while our previous work [12] was only the tip of the tip. New, higher throughput approaches, and rational ways of identifying and selecting independent lineages are clearly needed to fully understand this most fundamental of biological processes.
We observed the parallel evolution of mutations that disrupt one or more of three major signaling pathways responsible for sensing environmental stimuli and responding by governing growth rate. We propose a model for the adaptive strategy in constant, nutrient-limited environments (or at least in a glucose limited environment) (Figure 7), wherein constitutive commitment to cell division is beneficial, and thus mutations that result in unrestrained cell division are adaptive as long as the growth rate does not exceed the influx rate of nutrients into the system. By and large, these mutations are loss of function mutations. We consider the mutations in these pathways to be decoupling the sense and response to environmental stimuli, leading to an adaptive loss of environmental sensing in a constant environment. In contrast, these mutations are maladaptive in environments where nutrient abundance is not constant, such as when going through a boom and bust cycle from high glucose into starvation conditions. This may be due to depletion of the cell's reserve nutrient supply or the inability to enter a quiescent state, leading to cellular death. This adaptive loss of environmental sensitivity is a powerful example of how evolution is myopic: by evolving strategies to cope with a constant and predictable environment, genes and pathways are disrupted that would be necessary for survival when cells are confronted with an uncertain environment. It is noteworthy that the clones characterized in Wenger et al [35], also evolved in an aerobic glucose limited environment, were also more fit under a diverse set of other carbon limited environments, suggesting that their adaptive strategy also translated to other constant environments. Whether this strategy is widely applicable under an array of constant environments with different nutrient limitations remains to be determined through additional experimentation. However, recent analysis of experiments in bacteria have verified the idea that loss of function mutations can be a general strategy for adaptive evolution [42].
From a broad viewpoint, the adaptive strategy of loss of environmental sensitivity that we observed is similar to the strategy tumor cells use to proliferate. Cancer is an evolutionary process of clonal selection [43]–[45], and it is beneficial for the cells to replicate as fast as possible through the accumulation of mutations in oncogenes and tumor suppressor genes, many of which are in the homologous Ras/cAMP/PKA pathway that is recurrently mutated in our experiments [46], [47]. While the external environment humans face is dynamic and unpredictable, the human body has evolved to maintain homeostasis, exemplified by the near constant concentration of blood glucose [48]. Such mutations also come with trade-offs – when faced with an uncertain environment, many of the mutations show antagonistic pleiotropy (AP). In our data, three single mutants that we tested had a reduced fitness in the starvation environment, for which we have previously demonstrated fitness gains in the chemostat environment where they were selected – clearly cases of AP. For the multiple mutants that we tested, the loss of fitness could be due to AP, or alternatively result from a prior hitchhiking event in the evolved environment of a mutation that is deleterious in the starved environment. This mutation accumulation hypothesis (MA) is considered as an alternative to antagonistic pleiotropy when an evolved lineage shows fitness trade-offs. The fact the many, if not all of the mutations in our multiple mutants are in genes or pathways that are recurrently mutated in our chemostat evolutions makes MA seem a less likely explanation than AP. Indeed, previous experiments using E. coli evolving by serial transfer [49] showed that the rate of loss of unused functions in parallel evolving populations was consistent with AP, rather than MA, suggesting that AP may be widespread, and that when evolving in a consistent (though not necessarily constant) environment that due to the fitness cost of unneeded pathways, that there is a use it or lose if effect [50]. It has been shown using the yeast deletion collection that AP is indeed widespread, with approximately 20% of the collection of non-essential gene deletions being more fit under one of the tested conditions [51]. It is also of note that in evolving E. coli strains, mutations that result in a loss function of the sigma factor encoded by rpoS (which is involved in the general stress response) are frequently selected [52]. These mutations frequently exhibit AP, being detrimental under conditions where there is stress, the response to which needs to be balanced with growth (see [53] for review). Most of the AP examples thus provided are loss of function mutations (either from systematic gene deletion projects, or from sequencing beneficial mutations arising during experimental evolution), but a systematic catalog of AP effects of large numbers of beneficial mutations has not yet been generated. It will also be interesting to determine how clones with beneficial mutations that exhibit AP can perform adaptive escape when allowed to evolve afresh in an environment in which their previously beneficial mutations are now deleterious.
All population samples in this study have been previously described [12]. Briefly, three strains of haploid S288c that are isogenic, except that each constitutively expresses a different fluorescent protein (GFP, YFP or DsRed), were seeded in equal quantities in a 20 mL chemostat device. Each population was evolved for 448 generations at steady state under glucose limitation (0.08%) at a dilution rate of 0.2 h−1. During this evolution, the proportions of the three colored lineages were tracked using flow cytometry, and population samples were archived under deep freeze in glycerol at −80°C at regular intervals. Wild-type ancestral strain GSY1136 was also used as a reference for sequencing.
Illumina sequencing libraries were made directly from glycerol stocks of the original population samples, as well as the wild-type ancestral strain (GSY1136). Stocks were melted, and genomic DNA was extracted from 500 µl of each stock using Zymo Yeast Genomic DNA columns. The Nextera library prep kit (Epicentre) was used to construct the libraries, starting with 25–50 ng of genomic DNA. The tagmentation reaction was performed in LMW Reaction Buffer at 55°C for 10 minutes. The resulting tagged DNA was subjected to PCR using the Nextera PCR enzyme (Epicentre) under the following conditions: 72°C for 3 min, 95°C for 30 sec; 9 cycles of 95°C for 10 sec, 62°C for 30 sec, 72°C for 10 sec; final extension at 72°C for 1 min. A shortened extension time was used to bias the amplification of short fragments in order to maximize the proportion of bases being sequenced twice with overlapping paired-end Illumina reads. A modified Adapter 2 with a random hexamer barcode of sequence 5′-CAAGCAGAAGACGGCATACGAGATNNNNNNCGGTCTGCCTTGCCAGCCCGCTCAG-3′ (PAGE-purified, IDT Technologies) was used during the PCR for the population samples, while the standard Nextera Adapter 2 was used for GSY1136. No size selection was performed on the libraries, although they were concentrated through a Qiagen MinElute column. The same GSY1136 library was spiked into all 24 population libraries at a molar rate of 5%. The resulting libraries were sequenced on one lane apiece of 2×101 bp plus a 6 bp index read on the Illumina Hi-Seq 2000. In addition, two independent libraries from the same genomic DNA of GSY1136 were sequenced on one Hi-Seq lane apiece.
An overview of the sequencing analysis pipeline used to identify variants is given in Figure S3. The wild-type GSY1136 library that was spiked into each population sample was extracted with the exact tag ATCTCG using a modified version of the Fastx Barcode Splitter (http://hannonlab.cshl.edu/fastx_toolkit/index.html). Nextera adapters were trimmed off the 3′ read ends with Cutadapt v0.9.4 [54] supplied with the Nextera adapter sequence and default parameters except -m 15. The resulting FASTQ files were culled of any reads that occurred in only one read of the pair. Paired-end reads were mapped to a custom S288c reference genome with BWA (bwa-short) v0.5.9-r16 [55] using default parameters plus -I -q 10, and a sorted BAM file was created with Picard v1.45 FixMateInformation (http://picard.sourceforge.net).
The custom genome was constructed as follows: single end Illumina reads of a different ancestral wild-type strain (GSY1135) from a previous study [27] were mapped to the S288c reference sequences from the Saccharomyces Genome Database (SGD; http://www.yeastgenome.org/; downloaded 2/24/2011). SNPs were called with the GATK v1.0.5777 UnifiedGenotyper [56], [57], and a FASTA reference sequence was constructed that incorporated these SNP calls using the GATK FastaAlternateReferenceMaker.
The population data were culled of PCR duplicates using a modified version of Picard MarkDuplicates. In this program, the random hexamer barcodes were used in addition to the mapping coordinates to decide if a pair of reads was a PCR duplicate. Specifically, if more than one read pair had the same mapping coordinates in addition to the same hexamer barcode, only the pair with the highest mapping quality was retained for further analysis.
The in-lane spike-in of wild-type library was used to recalibrate the base qualities of the population data from the same lane. To achieve this, GATK CountCovariates and TableRecalibration were called on each lane of the wild-type data separately, using a variant mask for the CountCovariates step created by Samtools v0.1.16 [58] mpileup. Recalibration was visualized as successful as visualized by AnalyzeCovariates. The covariate file from the wild-type recalibration was then used as input for TableRecalibration on the population data from the same lane. Proper recalibration was assessed once again by AnalyzeCovariates.
A custom Java program was written to identify the bases in each library fragment that were sequenced twice by overlapping read pairs. This analysis was applied to both population and wild-type data, and the overlap information was stored in the custom “ZO” tag of the BAM file. A Python script implementing PySAM v0.5 (http://code.google.com/p/pysam/) was used to calculate the allele counts for each position in the reference genome, and the following filters were applied: uniquely mapping reads only, base quality score greater than 19 required, and only bases sequenced twice that were concordant in base identity between the two reads were retained.
Population SNP calls were made by comparing the allele counts in each population sample for each genomic position to the counts of the same allele and position from the wild-type data. This comparative approach filtered out any position that had false positive SNP calls due to positional effects, such as mapping or systematic sequencing errors. First, a merged wild-type file was created by combining all the spike-in control wild-type data with the two independently sequenced wild-type files. Second, only non-reference alleles that had both an allele count of at least 2 and a larger frequency in the population sample than the wild-type were retained. Third, a one-tailed Fisher's Exact Test was used to calculate if the number of non-reference alleles out of all alleles at a site was significantly greater in the population data than in the wild-type data for the same allele. These p-values were FDR corrected using the Benjamini and Hochberg method [59], and only sites with a q-value less than 0.01 were retained.
The following heuristic post-hoc filters were applied to the set of SNPs: 1) SNPs with a maximum frequency that was greater than the largest color proportion, plus 0.1, for the appropriate time point were removed (color frequency data from [12]). This removes any SNP that rose to a higher frequency than the highest color, which is not possible, unless identical SNPs arose in different colored populations. 2) Any SNP that was significant in the first time point was removed. This is because even if a new mutation present at the start of the experiment conferred a relative fitness of 2, that mutation would not be detectable in our assay in the first sampled generation. 3) Any site that was not deemed callable was removed. Callability was determined empirically with the GATK CallableLociWalker (-frlmq 0.01 -minMappingQuality 2) on the relevant population data, as well as the merged wild-type data. 4) Sites that had greater than 5% non-reference alleles in the merged wild-type data were removed. These sites were largely systematic errors. 5) SNPs where the read position of the variant allele did not vary were removed. This was defined as a read position standard deviation lower than one. 6) SNPs that had a mapping quality bias between reads containing the reference and variant alleles were removed, as calculated by a Bonferroni-corrected Mann-Whitney U test on mapping qualities.
Mutation allele frequencies were validated against a set of known mutation frequencies for experiment C1 (Figure S4) with data from [12], [27], as well as the fluorescent protein reporter frequencies for all experiments (Figure S5). All putative SNPs with a maximum allele frequency greater than 10% were confirmed by Sanger sequencing, except for the chr16:581589 mutation in experiment E2, which we were unable to amplify by PCR. While no effort was made to comprehensively catalog indels, Sanger sequencing of putative SNPs revealed six indels, which in every case were due to mapping errors of true indels near the ends of reads.
Co-occurrence of SNPs was determined by Sanger sequencing clones picked from the relevant time points for mutations greater than 10% allele frequency. The effect of each SNP (non-coding, synonymous coding, non-synonymous coding, etc.) was established with SNPeff v2.0.3 (http://snpeff.sourceforge.net/). The permissiveness of all missense mutations was calculated using SIFT [60] with default parameters.
To create the lineage dynamics plots, allele frequency data were plotted assuming linear expansion or contraction between primary data points. Since the allele frequency data were of lower resolution than the flow cytometry data (8 vs 47 time points), sometimes the inferred linear extrapolation between frequency data points resulted in an allele frequency greater than the color frequency. In these cases, the extrapolated allele frequencies were reduced to fit within the bounds of the color frequencies. Note, this fitting was performed for extrapolated points only; primary allele frequency data remained untouched.
All mutations discovered across the three experiments were divided into the following coding mutation effect categories: stop gained, start lost, stop lost, non-synonymous and synonymous. The sum of mutations within these categories was compared to the expectation using a chi-squared test. The expectation was calculated empirically by assuming random mutation throughout the genome; i.e. all possible mutations in the genome were made in silico, and the effect of the mutation was assigned to one of the categories above. The expected proportion of each category was calculated as the total for each category out of all possible mutations, and this proportion was multiplied by the total number of mutations discovered to get the expected number of mutations for each category. The same analysis was performed for coding versus non-coding mutations.
To find an enrichment of disruptive versus tolerated mutations, the totals of the stop gained, start lost, stop lost and disruptive non-synonymous categories were summed into the “disruptive” meta-category, and the synonymous, tolerated non-synonymous and non-coding mutations were summed into the “tolerated” meta-category. The SIFT predictions were used to classify non-synonymous mutations as either disruptive or tolerated. Expectations for disruptive or tolerated non-synonymous mutations were calculated empirically by summing the SIFT effect of all possible mutations for a particular protein.
Cell viability was quantified under starvation conditions using propidium iodide (PI) and flow cytometry similar to [61], in biological triplicate. Overnight cultures in 1.2 mL YPD were grown unshaken in deep-well 96 well plates at 30°C. Cultures were spun down, aspirated, and resuspended in 1.2 mL sterile water, and then diluted 1∶3 into a minimal medium described previously [12] supplemented with 2% glucose. The cultures were left undisturbed at 30°C between time points. Cell viability was measured at regular intervals post-inoculation by mixing the cultures and diluting 50 µL of culture into 250 µL water containing 250 µg PI, following by analysis by flow cytometry. The proportion of viable cells was calculated as PI-negative cells out of total cells analyzed. Significantly different viability was calculated with a two-tailed t-test between each mutant strain and wild-type at each time point. Cell viability based on PI staining was validated by colony forming unit analysis on a subset of the strains analyzed.
Gene Ontology (GO) biological process enrichments of coding mutations for each experiment were calculated using GO::TermFinder [62] at SGD with default options except “Feature Type” set to “ORF” and dubious ORFs disqualified from the analysis. For the reproducibility analysis, GO terms sharing edges were grouped into networks and GO networks were considered shared between experiments if they had at least one shared GO term.
All Illumina sequencing data are available from the NCBI Sequence Read Archive with accession number SRA054922.
|
10.1371/journal.pbio.2003143 | Task relevance modulates the behavioural and neural effects of sensory predictions | The brain is thought to generate internal predictions to optimize behaviour. However, it is unclear whether predictions signalling is an automatic brain function or depends on task demands. Here, we manipulated the spatial/temporal predictability of visual targets, and the relevance of spatial/temporal information provided by auditory cues. We used magnetoencephalography (MEG) to measure participants’ brain activity during task performance. Task relevance modulated the influence of predictions on behaviour: spatial/temporal predictability improved spatial/temporal discrimination accuracy, but not vice versa. To explain these effects, we used behavioural responses to estimate subjective predictions under an ideal-observer model. Model-based time-series of predictions and prediction errors (PEs) were associated with dissociable neural responses: predictions correlated with cue-induced beta-band activity in auditory regions and alpha-band activity in visual regions, while stimulus-bound PEs correlated with gamma-band activity in posterior regions. Crucially, task relevance modulated these spectral correlates, suggesting that current goals influence PE and prediction signalling.
| As natural environments change, animals need to continuously learn and update predictions about their current context to optimize behaviour. According to predictive coding, a general principle of brain function is the propagation of both neural predictions from hierarchically higher to lower brain regions and of the ensuing prediction-errors back up the cortical hierarchy. We show that the neural activity that signals internal predictions and prediction-errors depends on the current task or goals. We applied magnetoencephalography and computational modelling of behavioural data to a study in which human participants could generate spatial and temporal predictions about upcoming stimuli, while performing spatial or temporal tasks. We found that current context (task relevance) modulated the influence of predictions on behavioural and neural responses. At the level of behavioural responses, only the task-relevant predictions led to improvement in task performance. At the level of neural responses, we found that predictions and prediction-errors correlated with activity in different brain regions and in dissociable frequency bands—reflecting synchronized neural activity. Crucially, these specific neural signatures of prediction and prediction-error signalling were strongly modulated by their contextual relevance. Thus, our results show that current goals influence prediction and prediction-error signalling in the brain.
| The notion that the brain generates internal predictions to optimize behaviour is now well established [1–3]. Within the predictive-coding framework, predictions ground perceptual inference and are thought to be conveyed by descending connections in cortical hierarchies [4,5], which may be mediated by synchronized activity in the alpha- and beta-bands [6]. Conversely, incoming sensory or neural inputs—that are unexplained by predictions—translate into sensory prediction error (PE) signals. These “newsworthy” signals induce neural responses [4], which are thought to be propagated up sensory hierarchies in higher frequency bands such as gamma [5,6]. Accordingly, the modulation of alpha- and beta-band activity due to anticipatory predictions has been demonstrated in several modalities (visual: [7,8], auditory: [9,10], somatosensory: [11], motor: [12], see also [13]). Similarly, gamma-band PE signalling has been shown in visual [14] and auditory cortices [9,15].
Predictions can be generated about multiple attributes of stimuli, including their constituent features and their location and time of onset. Indeed, spatial and temporal predictions have been shown to act synergistically to improve visual discrimination in cued orienting tasks [3,16,17,18,19]. However, in natural cases, predictions are typically not cued but evolve dynamically (i.e., predicting the implications of hearing a car’s horn depends on the current context [e.g. location, traffic conditions, driving culture]). While some previous studies have shown that stimulus predictions can be generated and employed even when they are not behaviourally relevant [20–22], other findings suggest that the difference in neural activity triggered by predicted versus unpredicted targets is amplified by attention [23], and predictions about upcoming targets are learned and exploited more efficiently than predictions about nontargets [24]. However, it is not known whether predictions of multiple stimulus attributes are learned independently, or if the task relevance of specific predictions modulates their encoding and updating. Thus, while predictability and task relevance could constitute 2 independent sources of top-down control [25], relevance could also affect the deployment of predictions, precluding redundant or wasteful processing of task-irrelevant sensory information. In other words, predictability and task relevance may interact in selecting the most informative and relevant PEs for belief updating.
Here, to test whether the effect of predictability depends upon task relevance, we designed a task in which participants could use fluctuating spatial and temporal predictions to report either the location (left/right hemifield) or the latency (early/late relative to the cue) of visual targets. Predictions guiding task responses could be formed at different hierarchical levels of processing; at the lower level, participants could use a cue predicting the location/latency of the target in a given trial. At the higher level, they could learn the validity of the cue over several trials. We inferred the participants’ trial-by-trial predictions and PEs using an ideal-observer model based upon a hierarchical Bayesian inference [26–31]. The resulting predictions and PEs, as well as their interactions with task relevance, were used to explain time-frequency (TF) responses (measured with magnetoencephalography [MEG]) to test whether the neural correlates of predictions and PEs are modulated by task relevance.
Participants performed 2 tasks—location and latency discrimination of visual targets—in alternating blocks (Fig 1A). Each trial contained an auditory cue (a tone pair) and a visual target (a near-threshold square embedded in noise and presented peripherally). The auditory cue had the following 2 features: pitch (high versus low) and composition (ascending or descending pair). Similarly, the visual target had the following 2 features: location (right versus left hemifield) and latency (approximately 730 or approximately 1,270 ms after the cue). The following 2 cue-target contingencies were introduced in the task: cue pitch could predict target location and cue composition could predict target latency, with a varying degree of validity (Fig 1B). Participants were not informed of the cue-target contingencies or cue-validity manipulations. Thus, in certain (predictive) trials during each task block, the cue could be used to implicitly predict target location and/or latency, whereas in other trials, the cue was uninformative with respect to 1 or both target features. However, the task relevance manipulation was introduced explicitly, i.e., at the beginning of each block participants were informed whether they should discriminate the location or the latency of the target. Because cue validity varied along spatial and/or temporal dimensions, this design enabled us to orthogonalize predictability and task relevance, i.e., a stimulus could be predictable or not in the relevant or irrelevant context (determined by the current task).
In both tasks, cue validity led to improvements in discrimination accuracy depending on task relevance (Fig 1C). Thus, the interaction between predictability (a parametric factor encoding 90, 70, or 50% cue validity) and relevance (e.g., relevant: spatial predictability in a spatial task; irrelevant: spatial predictability in a temporal task) was significant for both tasks (spatial: F1,50 = 5.23, partial η2 = 0.09, p = 0.02; temporal: F1,50 = 5.10, partial η2 = 0.09, p = 0.02). In both tasks, the main parametric effect of predictability was not significant (p > 0.05, F < 2). However, there was a significant main effect of relevance (spatial: F1,50 = 34.61, partial η2 = 0.46, p < 0.001; temporal: F1,50 = 7.03, partial η2 = 0.15, p = 0.01), reflecting better overall performance in the spatial task (spatial: mean 87.2%, SEM 2.2%; temporal: mean 68.7%, SEM 2.2%) in the analyzed trials. No effect of the foreperiod on accuracy was observed in either task (paired t tests: short versus long intervals; both p’s > 0.2).
To explain the interaction between predictability and relevance on accuracy, we modelled individual participants’ responses using a Hierarchical Gaussian Filter (HGF) [26]. This Bayesian observer model allowed us to infer, on a trial-by-trial basis, the participants’ beliefs in terms of predictions and PEs about targets and cue validity levels. The HGF comprises an observer model, describing how the participants’ beliefs about various hierarchical aspects of the task are updated given trial outcome, and a response model, linking these beliefs to behavioural responses (Fig 2A). The observer model assumes that participants can form beliefs about 3 hierarchical aspects of uncertainty entailed by the task: (1) target location and/or latency in a particular trial (given the cue), (2) the current cue validity level, and (3) the current volatility (i.e., how fast cue validity changes over trials). Because task relevance was introduced in a deterministic (rather than probabilistic) way, we modelled relevance as a set of weights quantifying the contribution of predictions to the response in a given trial. By fitting the model to behavioural data, one can estimate the evidence for a particular model (quantifying how well the model explains the data, while penalizing for model complexity) and the model parameters. These parameters describe individual differences in learning and trial-by-trial expectations that generate predictions and PEs at various hierarchical levels.
We compared 5 alternative observation models: first, we specified 3-level HGFs (HGF3; in which participants’ beliefs at all levels are updated and can influence behaviour) in which the learning parameters could be either context-specific (i.e., the contribution of PEs to prediction updates could vary between task-relevant and task-irrelevant context) or nonspecific; in the same way, we specified 2-level HGFs (HGF2; in which changes in volatility are not inferred), again with context-specific or nonspecific learning parameters; finally, for a comparison with the Bayesian models we added a standard reinforcement-learning model (with a fixed learning rate). We also specified 2 alternative response models: task-specific (in which spatial/temporal predictions are used only to model responses in the spatial/temporal task) and task-general (in which both types of predictions contribute to responses in both tasks). Thus, our model space contained 10 models.
A random-effects Bayesian model comparison revealed that the winning model was HGF2, with context-specific learning parameters and observation parameters (Fig 2B; protected exceedance probability >95%, Bayesian omnibus risk p < 0.001, indicating very strong evidence for the winning model; cf. [32]). This suggests that our participants did not infer changes in volatility [33] and that their beliefs about target outcomes influenced learning and behaviour in task-relevant contexts only. The prior and posterior model parameters are provided in the Materials and methods section. Across tasks, the posterior learning parameter ωrel of the observation model (denoting the weight of context-relevant PEs in updating subsequent predictions; see Eqs 14 and 15 in Materials and methods) was the only significant predictor of individual participants’ mean accuracy out of the 4 free model parameters considered (stepwise regression: β = 0.25, p = 0.04; see Materials and methods and Fig 2E). This between-subject correlation provides an important validation of the within-subject model of behaviour, and suggests that the degree to which individuals learn predictions in task-relevant contexts is relevant for adaptive behaviour. Furthermore, the learning parameters ωrel and ωirrel were significantly different within participants (repeated-measures ANOVA with factors Relevance and Task; Relevance: F1,16 = 5.11, partial η2 = 0.15, p = 0.037; Task and Interaction: p > .25), providing further evidence that the learning parameters were sensitive to the observed behavioural effects of contextual relevance.
Example time-series of predictions and PEs (from one participant) are shown in Fig 2C. Beliefs about the most likely location and latency in a given trial (μ^1) track the evolving contingencies (pitch–location and composition–latency), suggesting that participants learn the objectively defined cue-validity level (although it is not an input to the model). Accordingly, predictions about cue validity μ^2 are highest for strongly predictable trials. Furthermore, PEs about target locations/latencies (δ1) increase when the outcome in a given trial does not match the prediction, and gradually decrease as the participant learns a new validity level and the respective precision ψ2 ramps up. This precision term is in turn used to weight the influence of the PEs on prediction updates. To relate predictions and PEs to neural activity, we used (unsigned) precision-weighted PEs εi = ψiδi-1 (see Eq 8), in addition to predictions about cue validity μ^2, as regressors to explain TF power of the MEG responses. The mean correlation between the regressors did not exceed r = 0.25, consistent with previous studies using the HGF ([27,29]; Fig 2D), and warranting their use as independent regressors in the analysis of neural activity.
To reduce MEG data dimensionality, analysis was performed in source space after localizing the principal sources involved in cue and target processing (Fig 3A; Table 1). Source reconstruction showed that (auditory) cues evoked activity in bilateral primary auditory cortex (A1) and middle temporal gyrus (MTG), whereas (visual) targets evoked activity in the region of calcarine cortex (V1). Additionally, cues induced more activity in the bilateral temporoparietal junction (TPJ) than targets. Source-level time-series were extracted from each source and transformed into TF estimates for the entire experimental session, averaging across hemispheres to avoid a multiple comparisons problem across unilateral regions in analysing the TF responses and in subsequent modelling. Thus, the main analysis focused on the modulation of the neural correlates of prediction and PE signalling independent of their possible lateralization. Participant-specific model-based sequences of predictions and PEs were used as regressors in a convolution general linear model (GLM) of TF responses [34]. The convolution model enabled us to detect significant parametric effects of predictions (|μ^2|) and PEs (|ε2|) on responses in each region. To test for modulatory effects of contextual relevance on prediction and PE signalling, regressors were entered separately for task-relevant and task-irrelevant contexts (thereby modelling an interaction).
As for the behavioural data, neural effects of predictions depended on task relevance. In the analysis of the simple main effects of predictions and PEs (i.e., ignoring their task relevance), no effect survived statistical significance testing (TF clusters thresholded at p < 0.05, uncorrected across TF points and Bonferroni-corrected across brain regions). However, testing for the effect of relevance on the TF correlates of predictions and PEs yielded several significant clusters of activity in cue- and target-processing regions (Table 2). Specifically, relevant predictions increased post-cue beta power in MTG and TPJ and decreased post-cue alpha power in A1 and V1 (Fig 3B), while PEs increased post-target gamma power in TPJ and V1 and decreased beta-band power in V1 (Fig 3C).
Beyond the modulation of prediction and PE signalling by relevance, cue-induced alpha-band responses in V1 differentiated between the spatial and temporal prediction estimates (Table 2 and Fig 3B). Furthermore, beta-band prediction signalling in MTG and PE signalling in V1 were modulated by task relevance and cue-target contingency, such that spatial predictions showed a stronger modulation by relevance than temporal predictions in MTG, whereas spatial PEs showed a weaker modulation by relevance than temporal PEs in V1 (Table 1 and Fig 3B and 3C). The latter finding might reflect a lateralization effect, whereby PE signalling in the temporal task is likely nonlateralized and as a result its modulation might be easier to detect in source activity averaged across hemispheres. Thus, we performed an additional control analysis to assess whether PE signalling is indeed more lateralized in the spatial task. To this end, we re-ran convolution modelling using a signed PE regressor, as opposed to the unsigned PE regressor used in the main analysis; thus, the signed PE regressors had positive values for unexpected targets on the left (or at early latencies), and negative values for unexpected targets on the right (or at late latencies). We reasoned that by using the signed PE regressor, in the spatial task source-level activity linked to PE signalling in different hemispheres in V1 should have the opposite sign (due to hemifield-specific PE signalling), whereas in the temporal task they should have the same sign (because PEs regarding target latency will be processed in both hemifields to a similar extent). We used a region-of-interest approach, focusing on the significant clusters identified in the main analysis (Fig 3C), whereby the mean spectral power was extracted from these clusters per hemisphere (left/right), task (spatial/temporal), context (relevant/irrelevant), and participant, and entered into an ANOVA with 3 factors (hemisphere, task, and context). As hypothesised, we did observe a significant hemisphere effect for spatial PEs in V1 (F1,64 = 8.29, p = 0.01), in addition to a main effect of relevance (F1,64 = 4.74, p = 0.04). In the temporal task, however, the main effect of hemisphere was not significant (F1,64 = 0.55, p = 0.46), although the effect of relevance was preserved (F1,64 = 4.33, p = 0.04). The remaining main or interaction effects were not significant.
To control for the possibility that the effects of predictions on neural responses might be contaminated by a differential processing of specific auditory cues (e.g., either their pitch or composition being more salient and therefore easier to process), we ran an additional control analysis testing whether source-level activity showed differential effects of pitch and/or composition. To this end, we repeated the analysis of cue-induced (i.e., prediction-related) responses with additional regressors coding for cue pitch (high versus low) and composition (ascending versus descending). Our rationale was that, in addition to treating pitch and composition as possible confounding factors on their own, any difference in variance explained by the 2 features respectively would be accounted for by these confound regressors and effectively removed from prediction-related activity. The inclusion of these regressors did not change the results in TF space: all the significant clusters identified before showed the same patterns of condition-specific differences as reported in Fig 3B, using identical significance criteria as in the original analysis (i.e., correcting for multiple comparisons across regions using Bonferroni correction, and for TF points using family-wise error ratio under random field theory assumptions).
To test the directionality of the effects identified above, we used dynamic causal modelling (DCM) for TF responses [35]. This phenomenological Bayesian modelling approach allows a quantification of the effective (directional) connectivity between different regions and frequency bands. Within- and cross-frequency amplitude–amplitude coupling was analysed in 2 time windows: 0–500 ms relative to the cue onset (in which TF activity was found to be modulated by prediction relevance; Fig 3B), and 0–500 ms relative to the target onset (in which TF activity was modulated by PE relevance; Fig 3C). In the analysis of the cue-processing period, we modelled connectivity in a network of 4 sources sensitive to prediction relevance: A1, MTG, TPJ, and V1. Similarly, when analysing target-induced activity, effective connectivity was modelled in a network of 2 sources: V1 and TPJ.
Fig 4 provides frequency-frequency coupling maps mediating (Fig 4A and 4B) the modulation of prediction and PE signalling by relevance, and (Fig 4C–4H) the significant modulatory parameter estimates quantifying the effects of relevance on cross-frequency coupling within and between regions. Regions involved in prediction (Fig 4A) and PE (Fig 4B) signalling showed the spectral asymmetry between ascending and descending connections, as suggested previously [5]. Specifically, in prediction signalling, ascending connections from A1 to TPJ, from V1 to MTG, and from V1 to TPJ showed strong excitatory effects in higher frequency ranges, whereas the respective descending connections showed net inhibitory effects in low-frequency ranges (Fig 4A). Similarly, in PE signalling, the ascending connection from V1 to TPJ mediated primarily excitatory effects (in both high and low frequency bands), whereas the descending connection mediated primarily inhibitory effects [5]. Upon closer inspection of the significant modulatory parameters of contextual relevance on prediction processing (Fig 4C–4E), task relevance primarily modulated the influence of low-frequency (alpha-beta) activity in A1 on low-frequency activity throughout the network, having a negative net effect on alpha-beta power in all regions. Additionally, contextual relevance modulated the influence of TPJ on MTG activity (frequency mode 1; Fig 4C), reflecting a further inhibition of alpha-beta activity in MTG. In contrast, PE relevance (Fig 4F–4H) primarily modulated the influence of V1 activity on the network, leading to a shift from lower to higher frequencies in V1, as well as to increased propagation of both high- and low-frequency activity to TPJ. Taken together, our DCM findings expand the previous results on low-frequency prediction signalling and high-frequency PE signalling by characterising the network-wide effective connectivity mediating these spectrally distinct effects.
The present study used a model-based MEG approach, assuming an ideal-observer model of behavioural data acquired in a task that orthogonally manipulated stimulus predictability and relevance. By fitting an HGF model [26] to each participant’s behaviour, we estimated the trial-by-trial predictions and PEs that best explained their performance. These estimates of subjective beliefs were then used as regressors in the analysis of the power of MEG responses. Crucially, we demonstrate an interaction between stimulus predictability and task relevance at the following 2 levels: task performance (accuracy) and neural activity. These converging results suggest that prediction and PE signalling are contextualized by current task goals.
The effects on accuracy extend previous findings, suggesting that stimulus predictability improves performance only when predictions pertain to task-relevant targets [24,36,37]. Accordingly, the validity of cues predictive of the target location (latency) improved accuracy in the spatial (temporal) task (Fig 1C), but not vice versa. The effects of irrelevant predictions were either abolished (in the spatial task; see [23], in which sensory predictions failed to affect processing of irrelevant stimuli) or nominally (but not significantly) reversed (in the temporal task) [38]. Finally, despite prior stimulus titration to 70% accuracy in both tasks, we observed differences in performance between the tasks, most likely due to an asymmetry between spatial and temporal discrimination; namely, successful temporal discrimination was necessarily associated with successful spatial discrimination, but inferring target location did not depend on inferring its latency. Nevertheless, the interaction of relevance and predictability was significant in both tasks and had similar effect sizes, suggesting robustness with respect to performance levels. Taken together, although previous studies suggest that the validity of task-irrelevant cues can be learned [20–22,39; but see 40] and spatiotemporal cues can work synergistically [3,16,17,18,19], we show that the effects of predictable cues on accuracy are strongly modulated by task set.
To explain this context sensitivity, we modelled behavioural data using HGF—an ideal Bayesian observer model of learning under uncertainty—allowing us to reconstruct subjective beliefs about experimental contingencies [26]. Hierarchical Bayesian models, such as the HGF, have been proven powerful in explaining behaviour in volatile and probabilistic tasks, by quantifying trial-by-trial inference. In previous work, the HGF has been applied to probabilistic attentional cueing paradigms [28,29] and used to delineate the functional anatomy [27] and neuromodulatory mechanisms [31] of encoding uncertainty at different hierarchical levels. In our task, uncertainty pertained to (1) target location/latency in a given trial, (2) cue validity level governing several trials, and (3) its volatility over multiple trials. In previous applications of HGF to cueing tasks (with a single cue-target contingency), the HGF3 (under which participants’ estimates at all levels influence behaviour) has typically been selected by Bayesian model comparison [27–30]. In our study, however, a comparison of several alternative observation models indicated a reduced HGF2 was the winning model, suggesting that performance of our participants was Bayes-optimal but not sensitive to changes in volatility [33]. Furthermore, the winning response model allowed only for a task-specific influence of relevant predictions on performance, consistent with the effects on accuracy—and establishing the construct validity of our modelling approach. Interestingly, the winning model implemented predictability and contextual relevance at hierarchically different levels: although predictability (cue validity) corresponds to the hidden state μ1 that the model successfully recovers from behaviour (Fig 2C, upper panels), contextual relevance is implemented at the level of weights ζrel and ζirrel that link these predictions to the simulated responses. Finally, the winning model included context-specific learning rates (separate for relevant and irrelevant contexts), optimised to each individual’s behavioural performance. At the group level, there was a significant difference between context-relevant and context-irrelevant learning rates. Furthermore, at the between-subject level, the context-relevant learning parameter quantifying the learning rates of spatial predictions in the spatial task, and of temporal predictions in the temporal task, correlated with individual participants’ mean accuracy, providing a further validation of the model.
Our model-based finding suggesting that learning rates depend on contextual relevance might explain the discrepancy between our results and several previous studies that reported the effects of predictability even in task-irrelevant contexts [20–22,39]. In these studies, predictability levels were fixed, unlike in our paradigm, in which cue validity varied over the course of the experiment with participants continuously updating their cue-based predictions of target features. In contrast, our results are fully consistent with previous work suggesting that task relevance facilitates learning of cue-target contingencies [24].
Further evidence for the interaction between predictability and relevance was seen at the level of neural responses. Here, we used the model-based trial-by-trial estimates of predictions and PEs as regressors in the analysis of MEG responses. This analysis revealed no main effect of predictions or PEs on cue- or target-induced responses. However, there was a significant interaction between task relevance and prediction (following cue onset) or PE estimates (following target onset). Thus, the neuronal responses were in line with the behavioural results and Bayesian model comparison described above, and suggested that the neuronal correlates of predictions and PEs are context sensitive and show an effect of task relevance.
Specifically, relevant predictions were associated with postcue beta-band synchronization and alpha-band desynchronization in auditory regions; most likely involved in the processing of the acoustic cues used in this paradigm. Here, although alpha-band activity was similar across tasks, beta-band modulation was seen predominantly in the spatial task (Fig 3B), possibly reflecting baseline performance differences between the 2 tasks. Both beta-band and alpha-band effects have previously been linked to the processing of predictive auditory stimuli (beta-band synchronization: [41]; alpha-band desynchronization: [42]), and interactive effects of expectation and attention have been identified in auditory beta-band activity [10]. Here, beyond the auditory regions, relevant predictions decreased cue-induced alpha-power in early visual cortex, consistent with previously reported alpha-band modulation due to anticipatory predictions in the visual [7,8,43,44] and other domains [11,12,45]. Although these findings suggest that prediction signalling in low-frequency bands might be a modality-general phenomenon [5], we show for the first time that this effect is modulated by contextual relevance.
Context-sensitive signatures of prediction signalling were seen following auditory cues but were not observed at any fixed latency before target onset, consistent with a recent study showing that the latency of beta-band synchronization is not predictive of the anticipated target latency, but instead locked to the cue onset [45]. Furthermore, prediction signalling was associated with activity in auditory (cue-processing) regions as well as in visual (target-processing) regions, and our DCM-based effective connectivity analysis suggested that the network-wide effects of contextual relevance on prediction processing are predominantly due to the influence of A1 activity on the neural responses throughout the network. Although studies on cross-modal orienting have shown that the effects of cue validity are primarily expressed in target-processing regions [46], previous work on cross-modal expectations suggests that predictions might be generated in 1 modality, but their effects manifest as PEs in another modality [47]. Because PEs are thought to be scaled by expected precision [48,49], our results suggest that predictions might be encoded shortly after the cue onset, but their effects on target processing will entail a modulation of target-induced PE activity.
In contrast to the neural correlates of prediction, relevant precision-weighted PEs were linked to target-induced, gamma-band modulation in posterior regions, which is in line with previous empirical work [9,14,15,50] and theoretical proposals [5]. Specifically, relevant PEs along both spatial and temporal dimensions were associated with increased target-induced, gamma-band responses in the posterior supramodal region TPJ/supramarginal gyrus (SMG), consistent with previous functional MRI (fMRI) correlates of HGF-based PE signalling [27]. Furthermore, relevant PEs decreased high beta-band power (approximately 34–35 Hz, below the range of typical visual gamma MEG responses [5]) in early visual regions, predominantly in the temporal task (Fig 3C) in which PE signalling was not as lateralized as in the spatial task, and thus its modulation easier to detect in source activity averaged across hemispheres (as identified in a control analysis; see TF responses in Results). A boost of gamma oscillations at the expense of lower frequencies [51,52] has often been reported as a correlate of predictability of targets [9,15,50,53,54]. In the effective connectivity analysis, we have identified that the observed modulatory effects of contextual relevance are explained by self-reinforcing alpha-band desynchronisation and gamma-band synchronization in V1. Such spectral shifts of neural responses towards higher frequencies are plausibly explained as a result of modulation of the excitability of principal cells and neuronal time constants that underwrite synaptic gain control [5,55]. The augmented high-frequency responses can then be propagated to higher regions, as suggested by stronger within-frequency coupling between V1 and TPJ identified in our DCM analysis.
The link between precision-weighted PEs and high-frequency responses is consistent with predictive coding, under which stimulus predictability (expected precision) is thought to increase postsynaptic gain of principal cells in superficial layers, typically associated with ascending output in high-frequency bands [5,56]. Previous work trying to disentangle the oscillatory signatures of predictions and PEs yielded evidence converging with our results. In a passive listening paradigm, in which the acoustic stimuli changed according to specific rules, sensory prediction violations (putative PEs) were linked to induced gamma activity, while prediction updates were manifest in the beta-band [9]. Similar results have been reported in an active attentional cueing paradigm, where anticipatory alpha/beta activity scaled with target predictability, while post-target gamma activity increased following sensory mismatch [50]. More generally, the spectral asymmetries between TF activity underlying prediction and PE signalling are consistent with previous postulates that predictions are propagated as descending signals from hierarchically higher to lower regions and mediated by low-frequency (e.g., beta-band) synchronization, while PEs are propagated as ascending signals from lower to higher regions and mediated by high-frequency (e.g., gamma-band) synchronization [5,13]. Here, because increased gamma-band power in the associative TPJ/SMG region preceded decreased lower frequency power in visual regions, our findings raise the possibility that the latter effect reflects a descending prediction update [9] following PE signalling in regions integrating cue and target processing.
Beyond showing spectrally and regionally specific responses corresponding to predictions and PEs, our results indicate that these responses are strongly modulated by their task relevance. Although this context sensitivity of prediction and PE signalling could be explained by an active inhibition of irrelevant features [57], recent evidence suggests that distractor suppression is less flexible than target facilitation [58], making it an unlikely explanation of the effects found in our relatively dynamic task. Alternatively, enhanced signalling of predictions and PEs relevant for the current context might reflect attentional prioritization (increased precision; cf. [48]) of the relevant (i.e., salient, uncertainty reducing) cue features and corresponding target features. This interpretation is consistent with recent evidence suggesting that relevant predictive cues attract gaze ([59]; but see [60] for evidence that stimulus regularity itself does not have to be salient), show enhanced working memory maintenance [61] and are judged more positively by viewers [62]. It is worth noting that, to ensure that the neural signalling of predictions and PEs (and its modulation by task relevance) is not mediated by nonspecific effects of attentional capture such as pupil size [38,63], we treated pupil size as a nuisance regressor in the analysis of TF responses. Thus, the observed neural effects were specifically related to the magnitude of predictions and PEs.
Although the interactive effects of cue validity and task relevance on behavioural and neural responses likely reflect that our participants deployed cue-based predictions to prioritize the contextually relevant visual target features, the same cues could have arguably been used to predict the most likely correct motor response. Although our task was not designed to specifically dissociate cue-target and cue-response mappings, we think that this is an unlikely scenario. First, because in both tasks feedback was given only at a block level and not at a single-trial level, it is unlikely that participants would show an effect of cue validity (interactive with contextual relevance) by learning the cue-response mapping alone. In other words, highly predictive and nonpredictive segments of the experiment could not be differentiated without the participants responding to the visual targets. Thus, our behavioural findings are unlikely to be due to participants dynamically updating their cue-to-response mapping on a trial-by-trial basis. Furthermore, in analysing the neural responses, possible confounds due to motor preparation were controlled for by using convolution modelling instead of more conventional TF analyses of epoched data. Specifically, a parametric regressor coding for which button was pressed in a given trial was included in each participant’s convolution model design matrix, effectively removing the effect of lateralised button press preparation up to 250 ms prior to motor response (i.e., overlapping with the latency of the observed neural correlates of PE processing).
Although several previous studies have used the terms “prediction” (or “expectation”) and “attention” interchangeably, here we followed previous conceptual distinctions [64] in treating expectation as the effect of likelihood of a given stimulus or event on its perceptual and neural processing, and attention as stimulus prioritisation based on its relevance. Although the common interchangeable use of the 2 terms in previous literature can to some extend be attributed to the popularity of classical paradigms (e.g., the Posner paradigm) confounding the 2 factors (i.e., the probability of a stimulus occurring and the probability of the required behavioural report), here we made sure to orthogonalise the probability and behavioural relevance of stimulus features (location and latency). However, besides this well-established distinction, there is a more subtle distinction to be made about expectations of particular stimulus features (e.g., of a particular target occurring on the left and late in a given trial) and the level of its predictability (manipulated here as cue validity). The latter distinction has been discussed more recently in the context of predictive coding [49], in which expectations of specific stimulus contents form first-order predictions, while the degree to which these expectations can be formed form second-order predictions (or predictions of precision). It is worth noting that this dissociation is captured by the HGF, in which both first-order predictions (at the lower level) and second-order predictions (precision ratio at the higher level) can influence behaviour. In the context of predictive coding, first-order predictions are thought to be mediated by descending (inhibitory) connections, silencing the ensuing PEs; second-order predictions, on the other hand, are thought to be mediated by modulatory connections increasing or decreasing the precision (gain) of PEs. As such, however, second-order predictions are akin to attention [16], which has also been linked to precision modulation under predictive coding [48]. At the level of TF responses, increased gain is typically associated with a shift from lower to higher frequency bands [5], as observed in our study. Thus, our findings can be interpreted as reflecting co-modulation of gain by predictability of specific stimulus features and the attentional prioritisation of these features.
We show that task relevance of cue and target features modulates performance accuracy, the influence of predictions on behavioural responses (as evidenced by Bayesian modelling), and the neural activity induced by both cues and targets. These findings are in line with the notion that the brain performs hierarchical perceptual inference by comparing sensory inputs with the predictions it generates about its own environment at multiple temporal scales. Crucially, we provide evidence that both the predictions and the ensuing PEs can flexibly adapt to dynamically changing goals.
This study was approved by the local ethics committee (Inter-divisional Research Ethics Committee, Medical Sciences, University of Oxford, approval ref. no. R48540/RE001) and all investigation has been conducted according to the principles expressed in the Declaration of Helsinki. Written informed consent has been obtained for each participant.
Healthy volunteers (N = 20, 12 females, 8 males; median age 22, range 18–49; all right-handed) were invited to participate in the experiment. All participants had normal hearing, no history of neurological or psychiatric diseases, and normal or corrected-to-normal vision.
Participants were asked to perform a speeded location (left versus right hemifield) or latency (approximately 0.75 s or approximately 1.25 s relative to an auditory cue) discrimination of visual targets. At the block level, an instruction screen specified which task (location or latency discrimination) should be performed next. Each block consisted of, on average, 48 trials (range 38–58). Participants received feedback about their average accuracy and RT after each block. Each participant completed 20 alternating blocks (10 per task), resulting in 960 trials in total. The duration of the whole experimental session was approximately 1 hour.
All visual stimulation was delivered using a projector (60-Hz refresh rate) in the experimenter room and transmitted to the MEG suite using a system of mirrors onto a screen located approximately 90 cm from the participants. Auditory stimulation was delivered by using MEG-compatible stereo ear tubes.
Each trial started with a display of 2 peripherally located placeholders on either side of a centrally presented fixation cross against a grey background. The placeholders were circles (radius 1.5° of the visual angle) consisting of random white dot patches (30% of the pixels within each circle; refreshed with every screen flip at 60-Hz refresh rate). The circles were located on a horizontal axis, with the centre of each placeholder 4° laterally from the fixation cross. After 500 ms (± 10 ms jitter) of placeholder presentation, an auditory cue was played. The cue consisted of 2 short (66 ms) gapless tone pips with carrier frequencies drawn from 4 possible values (400, 500, 800, and 1000 Hz). The cue was administered in a 2 x 2 factorial manner (factors: pitch and composition), and could therefore consist of pips that were either high (800 and 1000 Hz) or low (400 and 500 Hz) in pitch, forming either an ascending (400–500 Hz; 800–1000 Hz) or a descending pair (500–400 Hz; 1000–800Hz). After a variable delay, the cue was followed by a visual target—a white cardinally or diagonally oriented square (side length equals the radius of the placeholder; 50-ms display duration). The target was also administered in a 2 x 2 factorial manner (factors: location and latency), and could therefore be presented either within the left or the right placeholder, and either early (approximately 0.75 s) or late (approximately 1.25 s) relative to cue onset. The orientation of the target was a task-irrelevant feature introduced so that participants could not form a unique target template. The response buttons were counterbalanced across participants. Consecutive trials were separated by a jittered interval (1500–2500 ms).
Unbeknownst to the participants, cue features (pitch and composition) could predict 1 or both target features (location and latency) with varying validity (90, 70, 50, 30, or 10%) forming 2 contingency time-series: cue pitch could predict target location, while cue composition could predict target latency. The 2 contingency time-series were uncorrelated (r < .0001) and based on a predetermined arbitrary association. For instance, 90% spatial cue validity corresponded to 90% right (left) targets following high (low) pitched cues and 10% left (right) targets following high (low) pitched cues; in the 10% validity level, these proportions were reversed. Validity in each contingency time-series changed on average every 32 trials (range 8–54) and the consecutive validity levels varied pseudorandomly with no repetitions. Additionally, over the course of the experiment, validity could change in a more or less volatile way (on average every 16, 32, or 48 trials—with volatility updated after each 5 validity changes). To facilitate subsequent modelling, the validity time-series were precalculated for 2 runs of 480 trials each and fixed for all participants. The order of the 2 runs was counterbalanced across participants. In the behavioural and MEG analyses of both spatial and temporal predictability we a priori collapsed the 2 strongly predictable (90 and 10%) and the weakly predictable (70 and 30%) validity levels (cf. [33]). Thus, the main factors of interest in our experimental design and analysis were: spatial predictability, temporal predictability (each with 3 levels: strongly predictable, weakly predictable, and unpredictable), and task relevance (2 levels: spatial versus temporal task).
Prior to the main experimental session, we ran a short cue training session in which participants discriminated the pitch or composition of the cue until they reached >95% performance. We then trained them on the main experimental tasks (spatial and temporal discrimination of visual targets presented after the auditory cue, with cue validity changing dynamically just as in the main experiment; min. 50 trials per task). During this training, we administered a target stair-casing procedure; in which we adjusted the contrast of the visual targets to approximately 70% performance (1 up, 2 down procedure with an adaptive step size) in the spatial task, and the relative onset of the early versus late targets to approximately 70% performance in the temporal task. As a result, the mean target contrast was 0.28 relative to the placeholder (SD 0.08, range 0.18–0.53), and the mean asynchrony between early and late targets was 541.6 ms (SD 184 ms; mean early and late latencies 729.2 ms and 1270.8 ms postcue respectively; range of early latencies 600–877.6 ms postcue; range of late latencies 1122.4–1400 ms postcue). Prior to analyzing the behavioral and neural effects of stimulus predictability and task relevance, we excluded data from 1 participant who could not maintain central fixation, and 2 further participants whose mean accuracy in either task was below 55% or above 95%.
We analysed accuracy in two 3 x 2 repeated-measures ANOVAs, separately for each predictability manipulation (spatial versus temporal), with the main factors predictability (3 parametrically defined levels: strongly predictable, weakly predictable, and unpredictable) and task relevance (2 levels: relevant and irrelevant). The task-relevant trials corresponded to the spatial (temporal) task when analysing spatial (temporal) predictability; the remaining trials were treated as task-irrelevant. Because cue validity changed unbeknownst to the participants, and thus predictability effects could be offset by the initial trials in each validity level in which the previously learned contingency could be used, mean accuracy scores were calculated based on the second half of trials within each run with stable cue validity. Although synergistic effects between spatial and temporal predictability might have contributed to task performance [3,16,17], a 3 x 3 x 2 repeated-measures ANOVA was not conducted due to a low number of trials (<20) in some cells. Furthermore, trials with RTs longer than median +2 SD were discarded from behavioural and neural analyses.
Beyond testing for the behavioural effects of predictability and relevance, we used individual participants’ behavioural data to infer their beliefs about the targets and validity levels on a trial-by-trial level. Specifically, we modeled individual time-series of responses using a HGF (implemented in a Matlab toolbox available as an open source code: http://www.translationalneuromodeling.org/tapas) that models evidence accumulation or learning at multiple levels, and reconstructs an agent’s beliefs about the causes of their sensory inputs [26]. The model uses a variational approximation to an ideal hierarchical Bayesian observer. By fitting the model to behavioural data, one obtains participant-specific parameters of the model (determining the coupling of hierarchical levels, and thus individual learning time-series) and single-trial predictions and precision-weighted PEs at each level of the model hierarchy.
By design, our task introduced uncertainty at 3 levels: (1) where and when the target will appear in a particular trial; (2) which cue-validity level governs the given trial; and (3) how quickly the cue-validity level changes over time. Accordingly, the model estimates the participants’ beliefs at 3 different levels, corresponding to (1) the location xs1 and latency xt1 of the target, (2) the pitch-location contingency xs2 and the composition-latency contingency xt2, and (3) the volatility of these contingencies xs3 and xt3, respectively. These inferred beliefs are hidden states of the observation model, evolving as a Gaussian random walk, with the hidden states at a given level determining the variance of the random walk at the level below:
p(xs1,t1|xs2,t2)=s(xs2,t2)xs1,t1(1-s(xs2,t2))1-xs1,t1=Bernoulli(xs1,t1;s(xs2,t2)),
(1)
p(xs2,t2(k)|xs2,t2(k-1),xs3,t3(k))=N(xs2,t2(k);xs2,t2(k-1),exp(κxs3,t3(k)+ω)),
(2)
p(xs3,t3(k)|xs3,t3(k-1),ϑ)=N(xs3,t3(k);xs3,t3(k-1),ϑ).
(3)
At the lowest level (Eq 1), the prediction of the target location xs1 (or latency xt1) in a particular trial takes possible values {0; 1} arbitrarily defined in contingency space: i.e., left targets following high-pitch cues and right targets following low-pitch cues are defined as 1, while the opposite locations are defined as 0; similarly, early targets following ascending cues and late targets following descending cues are defined as 1, while the opposite latencies are defined as 0. These low-level predictions are described as a logistic sigmoid function of the respective inferred contingency between the cue and the target xs2 (xt2), such that if the inferred contingency xs2,t2 = 0, both targets (xs1,t1 = 1 and xs1,t1 = 0) are equiprobable. At the middle level (Eq 2), the inferred cue validity level in a given trial xs2,t2(k) is normally distributed around the validity level from the previous trial xs2,t2(k-1), with the variance of this distribution depending on the inferred volatility xs3,t3(k). Here, the free parameter κ describes how strongly the estimated volatility will influence validity level learning, and ω is a constant component of the learning step size. Finally, at the highest level (Eq 3), the inferred volatility xs3,t3 is normally distributed around the inferred volatility from the previous trial, with the variance of this distribution (i.e., the speed of learning about the volatility) described by a free parameter ϑ.
During the fitting of the model to the data, one can estimate the trial-by-trial time-series (at each level i) of the participants’ beliefs μi(k)(i.e., posterior means of states xi(k)) and the updates on these beliefs εi(k)(precision-weighted PEs) after observing a target. The variational approximation in the HGF provides analytic update equations describing these time-series:
μi(k+1)-μi(k)~ψi(k)δi-1(k)=εi(k),
(4)
ψi(k)=π^i-1(k)πi(k),
(5)
πi(k)=1σi(k).
(6)
As shown in Eqs 4–6, in each trial, a belief update μi(k+1)-μi(k) is proportional to the PE at the level below δi-1(k), weighted by a precision ratio ψi(k). This precision ratio depends on the precision (inverse variance) of the prediction at the level below π^i-1(k) and at the current level πi(k). The superscript ^ denotes “prediction”: μ^1(k) is the prediction on trial k before observing the trial outcome, and π^i(k) is the precision of this prediction. After applying these update equations to specific hierarchical levels, we obtain:
μ3(k+1)-μ3(k)~ψ3(k)δ2(k)=ε3(k),
(7)
δ2(k)=σ2(k)+(μ2(k)-μ2(k-1))2σ2(k-1)+eκμ3(k-1)+ω-1,
(8)
μ2(k+1)-μ2(k)~ψ2(k)δ1(k)=ε2(k),
(9)
δ1(k)=μ1(k)-μ^1(k),
(10)
μ^1(k)=s(μ2(k-1)).
(11)
At the lower level, the PE about the observed target δ1(k) is simply the difference between the actual and the predicted target (Eq 10), in which the prediction is a sigmoid function of the previous trial’s prediction about the validity level (Eq 11). This PE, weighted by the corresponding precision ratio, is used to update the predictions about the validity level in the next trial (Eq 9). At the higher level, the PE about the validity level (Eq 8; cf. [23] for a detailed explanation) is used to update the prediction of volatility in the next trial (Eq 7). These HGF-derived time-series (specifically, |μ^2(k)| and |ε2(k)|)–fitted to each participant’s behavioural data—were used as regressors in subsequent analysis of MEG data. Prior variance log(σ2(0)) was treated as a free parameter.
Finally, to map the agent’s beliefs onto the observed behavioural data, we specified a response model for categorical outcomes (a binary softmax function of the agent’s predictions), where the probability of a particular outcome y = {0; 1} is described by the logistic sigmoid:
p(y|μ^1,ζ)=s(-ζ(2μ^1-1)(2y-1)).
(12)
The free parameter ζ encodes the decision noise. Here, because we had 2 time-series of predictions corresponding to the target location and latency (μ^1s and μ^1t), and they could be either relevant (e.g., μ^1s in trials corresponding to the spatial task) or irrelevant (e.g., μ^1s in trials corresponding to the temporal task), we parameterized the response model such that both relevant and irrelevant predictions could explain behaviour with different weights:
p(ys,t|μ^1s,t,ζ)=s(-ζs,trel(2μ^1s,t-1)(2ys,t-1)-ζs,tirrel(2μ^1t,s-1)(2ys,t-1)).
(13)
Thus, in a spatial task, both spatial predictions μ^1s (via ζsrel) and temporal predictions μ^1t (via ζtirrel) may have been used to model the observed response. Responses y were coded in contingency space, thus the mapping of y onto its possible values {0; 1} was identical to the mapping of xs1,t1 onto {0; 1}.
To select a model that best describes our observed data, we designed 5 alternative observation models (HGF3s with context-specific or nonspecific learning parameters ω; HGF2s, where changes in volatility are not inferred as κ = 0, again with context-specific or nonspecific learning parameters ω; and a standard reinforcement-learning model with a fixed learning rate) and 2 response models (Eq 13) in a factorial manner. Thus, our HGF observation model could consist of 3 levels (with a free parameter κ) or 2 levels (with κ = 0). As a result, in the reduced HGF2, the volatility estimates were decoupled from the lower levels and did not influence behaviour. Furthermore, the HGF learning parameters ω were either context-specific (i.e., with free parameters ωrel and ωirrel estimated for relevant and irrelevant contexts respectively, see Eqs 14 and 15 below) or nonspecific (whereby a single free parameter ω was estimated for both contexts, as in Eq 8).
Similarly, our response model could include both the relevant and irrelevant predictions (with free parameters ζrel and ζirrel), or only the relevant predictions (with ζirrel = 0). Additionally, as an alternative observation model not based on the HGF, we considered a standard reinforcement learning based on the Rescorla-Wagner formulation [65], with 2 free parameters representing fixed learning rates of location and latency, respectively. Models were compared using their free-energy approximation to log-model evidence in a random-effects Bayesian model selection procedure [32]. The prior and posterior means ± SD for all free parameters of the winning model are shown in Table 3.
MEG data were acquired using a 275-channel whole-head setup with third-order gradiometers (CTF MEG International Services LP, Coquitlam, British Columbia, Canada) at a sampling rate of 1200 Hz. Eye movements and pupil size data were recorded using a nonferrous infrared eye-tracking system (SR Research, Ottawa, Ontario, Canada). All subsequent analyses were performed in SPM12 (Wellcome Trust Centre for Neuroimaging, University College London), except where noted.
Continuous data were high-pass filtered at 0.1 Hz and notch-filtered at 50 Hz to remove slow drifts and line artefacts, and downsampled to 300 Hz. The vertical eye-tracker data were used to detect blinks. Sensor data were corrected for blink artifacts by subtracting their 2 principal modes [66]. To reduce the dimensionality of the data for subsequent analysis, we identified the main sources involved in processing task stimuli using multiple sparse priors under group constraints [67]. Here, artefact-corrected data were epoched between −1000 and 1000 ms relative to cues and targets, low-pass filtered at 48 Hz and baseline-corrected relative to the last 100 ms before cue or target onset by subtracting the average of the baseline period. Epoched data (960 trials per condition) were averaged per channel and condition using robust averaging [68]. Per participant, we calculated 3D source activity maps corresponding to the evoked activity in the 0–400 ms (Hanning) window relative to cue and target onset as well as their respective baselines (−400 to 0 ms). The primary sources involved in cue (target) processing were identified as clusters of significant differences between postcue (post-target) and precue (pretarget) source activity maps using GLMs with factors participant and epoch (post versus pre), after thresholding and correcting the statistical parametric maps at a peak-level pFWE < 0.05. Additionally, to identify sources involved in differential processing of cues and targets, we calculated 3D source activity maps of total (evoked and induced) activity (0–400 ms relative to cue and target) present in the data after band-pass filtering the entire epochs between 1 and 48 Hz, and contrasted the ensuing activity maps related to cue- versus target-processing in a GLM with factors participant, stimulus (cue versus target), and epoch (post versus pre), thresholding and correcting the statistical parametric maps at a peak-level pFWE <0.05. Sources were labeled using the SPM12 atlas provided by Neuromorphometrics, Inc. Significant clusters were then used to extract individual participants’ source-level time-series using a linearly constrained minimum variance beamformer [69], as implemented in the Data Analysis in Source Space (DAiSS) toolbox for SPM12 (https://github.com/SPM/DAiSS). Source-level time series, extracted from continuous data after high-pass and notch filtering, but before the remaining preprocessing steps, were transformed into a TF representation (frequency range: 8–48 Hz, frequency step: 2.5 Hz, frequency smoothing: ±2 Hz) using a sliding Hanning tapered window (length: 400 ms, time step: 20 ms) as implemented in the mtmconvol function of the FieldTrip toolbox (http://www.fieldtriptoolbox.org/). TF data were log-transformed, averaged per source across hemispheres, and entered into a convolution analysis for TF responses [34].
|
10.1371/journal.pmed.1002502 | Perfluoroalkyl substances and changes in body weight and resting metabolic rate in response to weight-loss diets: A prospective study | The potential endocrine-disrupting effects of perfluoroalkyl substances (PFASs) have been demonstrated in animal studies, but whether PFASs may interfere with body weight regulation in humans is largely unknown. This study aimed to examine the associations of PFAS exposure with changes in body weight and resting metabolic rate (RMR) in a diet-induced weight-loss setting.
In the 2-year POUNDS Lost randomized clinical trial based in Boston, Massachusetts, and Baton Rouge, Louisiana, that examined the effects of energy-restricted diets on weight changes, baseline plasma concentrations of major PFASs were measured among 621 overweight and obese participants aged 30–70 years. Body weight was measured at baseline and 6, 12, 18, and 24 months. RMR and other metabolic parameters, including glucose, lipids, thyroid hormones, and leptin, were measured at baseline and 6 and 24 months. Participants lost an average of 6.4 kg of body weight during the first 6 months (weight-loss period) and subsequently regained an average of 2.7 kg of body weight during the period of 6–24 months (weight regain period). After multivariate adjustment, baseline PFAS concentrations were not significantly associated with concurrent body weight or weight loss during the first 6 months. In contrast, higher baseline levels of PFASs were significantly associated with a greater weight regain, primarily in women. In women, comparing the highest to the lowest tertiles of PFAS concentrations, the multivariate-adjusted mean weight regain (SE) was 4.0 (0.8) versus 2.1 (0.9) kg for perfluorooctanesulfonic acid (PFOS) (Ptrend = 0.01); 4.3 (0.9) versus 2.2 (0.8) kg for perfluorooctanoic acid (PFOA) (Ptrend = 0.007); 4.7 (0.9) versus 2.5 (0.9) kg for perfluorononanoic acid (PFNA) (Ptrend = 0.006); 4.9 (0.9) versus 2.7 (0.8) kg for perfluorohexanesulfonic acid (PFHxS) (Ptrend = 0.009); and 4.2 (0.8) versus 2.5 (0.9) kg for perfluorodecanoic acid (PFDA) (Ptrend = 0.03). When further adjusted for changes in body weight or thyroid hormones during the first 6 months, results remained similar. Moreover, higher baseline plasma PFAS concentrations, especially for PFOS and PFNA, were significantly associated with greater decline in RMR during the weight-loss period and less increase in RMR during the weight regain period in both men and women. Limitations of the study include the possibility of unmeasured or residual confounding by socioeconomic and psychosocial factors, as well as possible relapse to the usual diet prior to randomization, which could have been rich in foods contaminated by PFASs through food packaging and also dense in energy.
In this diet-induced weight-loss trial, higher baseline plasma PFAS concentrations were associated with a greater weight regain, especially in women, possibly explained by a slower regression of RMR levels. These data illustrate a potential novel pathway through which PFASs interfere with human body weight regulation and metabolism. The possible impact of environmental chemicals on the obesity epidemic therefore deserves attention.
ClinicalTrials.gov NCT00072995
| Although many approaches can be used to achieve a short-term weight loss, maintenance of weight loss has become a key challenge for sustaining long-term benefits of weight loss. Accumulating evidence has suggested that certain environmental compounds may play an important role in weight gain and obesity development.
The potential endocrine-disrupting effects of perfluoroalkyl substances (PFASs) have been demonstrated in animal studies, but whether PFASs may interfere with body weight regulation in humans is largely unknown.
In a 2-year diet-induced weight-loss trial (the POUNDS Lost trial), we measured plasma concentrations of PFASs at baseline in 621 overweight and obese men and women and collected information on changes in body weight, resting metabolic rate (RMR), and other metabolic parameters during weight loss and weight regain over the 2 years the participants were on the study diet.
Higher baseline levels of PFASs were significantly associated with a greater weight regain, primarily in women. On average, women in the highest tertile of PFAS concentrations regained 1.7–2.2 kg more body weight than women in the lowest tertile.
Higher baseline plasma concentrations of PFASs, especially perfluorooctanesulfonic acid (PFOS) and perfluorononanoic acid (PFNA), were significantly associated with greater decline in RMR during the first 6 months and less increase in RMR during the period when participants on average regained weight (6–24 months).
In this diet-induced weight-loss trial, higher baseline PFAS concentrations were associated with a greater weight regain, especially in women, possibly explained by a slower return of RMR levels. These data provide initial evidence suggesting that PFASs may interfere with human body weight regulation and counteract efforts to maintain weight loss in adults.
| Obesity has become a worldwide public health concern [1,2]. Based on recent US data, the prevalence of obesity is 37.7% in adults and 17.0% in children and adolescents, with no sign of a reduction in the foreseeable future [3–5]. Although many approaches can be used to achieve short-term weight loss, its maintenance remains a key challenge [6,7]. Meanwhile, given the same intervention strategies, apparent within-group variability in weight loss and weight regain has been demonstrated [7,8]. Although the exact reasons for the variability are largely unknown, accumulating evidence has suggested that certain environmental compounds may play an important role in weight gain and obesity development [9,10].
Perfluoroalkyl substances (PFASs), especially perfluorooctanoic acid (PFOA) and perfluorooctanesulfonic acid (PFOS), have been identified as plausible endocrine disruptors with the potential to perturb weight regulation [9,11–14]. Evidence from animal studies has suggested that PFASs may be involved in altering energy metabolism and thyroid hormone homeostasis [15–17], likely through the activation of various transcriptional factors, such as the peroxisome proliferator-activated receptors (PPARs) [18–20]. However, given the species-specific toxicokinetics and tissue distribution of PFASs [18], extrapolation from animals to humans has yet to be supported. Although some human studies have examined the potential intergenerational effects of PFASs on body weight, the findings were somewhat inconsistent [21–27]. To our knowledge, no prospective study has explored the association between PFAS exposure and weight change in adults under controlled circumstances. Furthermore, it is largely unknown whether resting metabolic rate (RMR) or thyroid hormones, factors that can influence energy expenditure [28], might be also involved in the potential effects of PFASs on weight regulation [29,30].
PFASs are extensively used in many industrial and consumer products, including food packaging, paper and textile coatings, and non-stick cookware [31–34]. A recent study reported that the drinking water supplies for at least 6 million US citizens may exceed the US Environmental Protection Agency’s health advisory limit for lifetime exposure to PFOS and PFOA from drinking water [35]. In addition, these compounds are extremely stable in the environment and have a long elimination half-life in the human body [36], thus rendering PFASs a possible threat to human health. Due to the potential metabolic abnormalities associated with elevated PFAS levels, we aimed to examine the associations of PFAS exposure with changes in body weight and RMR in the well-designed and rigorously conducted POUNDS (Preventing Overweight Using Novel Dietary Strategies) Lost trial [37].
The protocol was approved by the institutional review boards at the Harvard T.H. Chan School of Public Health, Brigham and Women’s Hospital, and the Pennington Biomedical Research Center of the Louisiana State University System, as well as by a data and safety monitoring board appointed by the National Heart, Lung, and Blood Institute. All participants provided written informed consent. The trial was registered at ClinicalTrials.gov (NCT00072995).
The POUNDS Lost study, a 2-year randomized clinical trial, was designed to compare the effects of 4 energy-reduced diets with different macronutrient (i.e., fat, protein, and carbohydrate) compositions on body weight, as previously described [37]. At baseline, 811 overweight and obese men and women aged 30–70 years were randomly assigned to 1 of 4 diets that consisted of different compositions of similar foods and met the guidelines for cardiovascular health. Eighty percent of the participants (n = 645) completed the trial. Each participant’s caloric prescription for the 2-year period represented a deficit of 750 kcal per day from baseline, as calculated from each individual’s resting energy expenditure and activity level [37]. All participants had normal thyroid function at study baseline [38]. The main findings of this trial were that most of the weight loss was observed in the first 6 months, followed by a gradual weight regain through to 24 months, and that the weight changes (both weight loss and weight regain) did not differ significantly between the diet groups [37].
The current analysis included 621 participants with available fasting plasma samples collected at baseline. Of these individuals, 592 and 460 participants also provided blood samples at 6 months and 2 years, respectively.
In the morning before breakfast and after urination, body weight and waist circumference were measured at baseline and 6, 12, 18, and 24 months. Body mass index (BMI) was calculated as body weight in kilograms divided by height in meters squared. At baseline and 6 and 24 months, body fat mass and lean mass (n = 424) were measured using dual energy X-ray absorptiometry (DXA) (Hologic QDR 4500A bone densitometer; Hologic); visceral and subcutaneous abdominal fat (n = 165) were measured using a computed tomography (CT) scanner [39]; and blood pressure was measured by an automated device (Omron HEM907XL; Omron). RMR was assessed at baseline and 6 and 24 months using a Deltatrac II Metabolic Monitor (Datex-Ohmeda) after an overnight fast [40]. Briefly, after a 30-minute rest, a transparent plastic hood was placed over the head of the participant for another 30 minutes. Participants were required to keep still and awake during the test, and the last 20 minutes of measurements were used for the calculation of RMR [40].
Plasma concentrations of PFOS, PFOA, perfluorononanoic acid (PFNA), perfluorohexanesulfonic acid (PFHxS), and perfluorodecanoic acid (PFDA) were measured at baseline only, using a sensitive and reliable method based on online solid phase extraction and liquid chromatography coupled to a triple quadropole mass spectrometer [41], with minor modifications. Due to the long elimination half-lives of the PFASs and incomplete samplings, we did not measure plasma PFAS levels during the trial. For all major PFASs, the concentrations were above the limit of detection (0.05 ng/ml), and the inter- and intra-assay coefficients of variation were <6.3% and <6.1%, respectively.
In our pilot study evaluating the within-person stability of PFAS concentrations, intra-class correlation coefficients (ICCs) between concentrations in 2 blood samples collected 1–2 years apart from 58 participants in the Nurses’ Health Study II demonstrated excellent reproducibility of PFAS concentrations in blood: the ICCs were 0.91 for PFOS, 0.90 for PFOA, 0.94 for PFHxS, 0.87 for PFNA, and 0.82 for PFDA (all P < 0.001).
At baseline, 6 months, and 24 months, fasting plasma glucose, insulin, total cholesterol, high-density lipoprotein (HDL) cholesterol, low-density lipoprotein (LDL) cholesterol, and triglycerides were measured on the Synchron CX7 (Beckman Coulter), and hemoglobin A1C (HbA1c) was measured on a Synchron CX5 (Beckman Coulter); plasma leptin and soluble leptin receptor were measured by an ultrasensitive immunoassay (R&D Systems); and serum free triiodothyronine (T3), free thyroxine (T4), total T3, total T4, and thyroid stimulating hormone were measured using a competitive electrochemiluminescence immunoassay on the Roche E modular system (Roche Diagnostics), as previously described elsewhere [37]. The homeostatic model assessment of insulin resistance (HOMA-IR) value was calculated using the updated HOMA model (HOMA2) described by Levy et al. [42]. Adipose tissue was obtained from 103 participants at baseline and at 6 months. Gene expression was measured by direct hybridization using the Illumina HumanHT-12 v3 Expression BeadChip (Illumina) (details in S1 Text).
Using standardized questionnaires, we obtained information on age, sex, race (white, black, Hispanic, or other), educational attainment (high school or less, some college, or college graduate or beyond), smoking status (never, former, or current smoker), alcohol consumption (drinks/week), menopausal status (yes or no), and hormone replacement therapy use (yes or no). At baseline, 6 months, and 24 months, physical activity was assessed using the Baecke physical activity questionnaire, which included 16 items inquiring about levels of habitual physical activities (i.e., physical activity at work, sports during leisure time, and other physical activity during leisure time) [43]. All responses were pre-coded on 5-point scales. Total physical activity was expressed as the average of the sum of the individual responses, with a score ranging from 0 to 5 [43].
The comparisons between participants included in the current analysis and those excluded were evaluated by the Student’s t test for normally distributed variables, the Wilcoxon rank-sum test for skewed variables, and the chi-squared test for categorical variables. The associations between baseline PFASs and changes in body weight and RMR during the period of weight loss (first 6 months) or weight regain (6–24 months) were examined using linear regression. The least-square means of changes in body weight (at 6, 12, 18, and 24 months) and RMR (at 6 and 24 months) according to tertiles of baseline PFAS concentrations were calculated. In addition, the relationship between PFASs and other potential mediators including thyroid hormones and leptin were further evaluated using linear regression. Covariates considered in multivariate adjustments included baseline age (continuous), sex, race, educational attainment (high school or less, some college, or college graduate or beyond), smoking status (never, former, or current smoker), alcohol consumption (continuous), physical activity (continuous), the 4 diet groups, and baseline BMI (or baseline RMR for the analysis of RMR change). Moreover, menopausal status and hormone replacement therapy (women only) were also entered into the model in a sensitivity analysis. To test the linear trend of the associations of baseline PFAS concentrations with changes in body weight and RMR, we assigned a median value to each tertile of PFAS concentration and treated it as a continuous variable. We also tested the linear trend using the PFAS concentrations as continuous variables (log10-transformed). In an exploratory analysis, factor analysis was used to explore the potential exposure patterns of PFASs.
To investigate the associations of baseline PFASs with baseline values of and changes in other metabolic parameters (including glucose, lipids, thyroid hormones, and leptin), Spearman correlation coefficients (rs) were calculated with adjustment for the potential confounders mentioned above. Stratified analysis was also conducted according to sex, and a likelihood ratio test was performed to test for potential interactions. In sensitivity analyses, body weight or RMR at 6 months (or changes during the first 6 months), instead of the baseline value, was included in the multivariate models when examining the associations between baseline PFASs and changes in body weight or RMR during the period of 6–24 months. We also stratified the analyses by dietary intervention group. In addition, to account for the correlations between measurements on the same individuals, linear mixed-effects models were also used to examine the associations between baseline PFAS concentrations and weight regain (weight measurements at 6, 12, 18, and 24 months), with an unstructured covariance matrix. To assess confounding patterns, in another sensitivity analysis, the covariates were entered into the model in a stepwise manner. In an exploratory analysis, we also examined the associations of PFAS exposures with the gene expression profile in adipose tissue (S1 Text).
A 2-sided P < 0.05 was considered statistically significant. The statistical analyses were performed with SAS software, version 9.4 (SAS Institute).
The mean (SD) age of the 621 participants was 51.4 (9.1) years, with a mean (SD) baseline BMI of 32.6 (3.8) kg/m2. Participants lost an average of 6.4 kg of body weight during the first 6 months and subsequently regained an average of 2.7 kg during the remaining study period. In comparison with the POUNDS Lost participants not included in the current study due to the lack of plasma samples at baseline, the participants included were slightly older (51.4 versus 49.1 years, P = 0.01), but there were no significant differences in other characteristics, including body weight and RMR (S1 Table).
Table 1 shows the baseline characteristics of the study participants. PFOS and PFOA were the dominant PFASs. The median (interquartile range) plasma concentration was 24.5 (16.2–37.0) ng/ml for PFOS, 4.5 (3.3–6.3) ng/ml for PFOA, 2.4 (1.5–3.6) ng/ml for PFHxS, 1.5 (1.0–2.4) ng/ml for PFNA, and 0.37 (0.27–0.52) ng/ml for PFDA. At baseline, significant inter-correlations were observed between PFOS, PFOA, PFHxS, PFNA, and PFDA (rs ranged from 0.38 to 0.85) (S2 Table), although no particular pattern of PFAS mixture was identified in the factor analysis. After multivariate adjustment, PFOS, PFOA, and PFNA concentration were all positively associated with insulin, HOMA-IR, diastolic blood pressure, and free T3 (rs ranged from 0.10 to 0.18, all P < 0.05) at baseline. In addition, certain PFASs (e.g., PFHxS and PFDA) were positively associated with some of the variables, including visceral fat mass, systolic blood pressure, glucose, triglycerides, LDL cholesterol, free T4, total T4, and leptin (rs ranged from 0.08 to 0.24, all P < 0.05) (S2 Table). No PFASs were correlated with body weight, waist circumference, BMI, or RMR at baseline.
After multivariate adjustment including smoking status, physical activity, baseline BMI, and dietary intervention group, baseline PFAS concentrations were not associated with weight loss in the first 6 months (Table 2). The crude positive associations between certain PFAS levels and weight loss were abolished after multivariate adjustment (Table 2). In contrast, after multivariate adjustment, baseline PFOS and PFNA concentrations were positively associated with greater weight regain in the total study population. Comparing the highest to the lowest tertiles, the least-square means (SEs) of weight regain were 3.3 (0.6) versus 1.8 (0.6) kg for PFOS (Ptrend = 0.009) and 3.4 (0.6) versus 2.0 (0.6) kg for PFNA (Ptrend = 0.01) (Model 2 in Table 2). The results were similar when PFAS concentrations were treated as continuous variables (the beta coefficients for per-unit log10-transformed PFOS and PFNA increment were 0.80 and 1.02, respectively; both Pcontinuous < 0.05) (Table 2). After further adjusting for baseline thyroid hormones (Model 3 in Table 2), the associations remained significant. In sensitivity analyses, when body weight at baseline or 6 months (instead of BMI at baseline) was adjusted for in the models, the results were largely unchanged. When changes in body weight or changes in thyroid hormones or leptin during the first 6 months were also included as covariates, the results did not change materially. In addition, similar results were obtained when using linear mixed-effects models. When PFAS levels were categorized into quartiles, the results were largely similar.
In an analysis stratified by sex, significant associations with weight regain were observed for all individual PFASs in women, but not in men. Comparing the highest to the lowest tertiles, the least-square means (SEs) of weight regain in women were 4.0 (0.8) versus 2.1 (0.9) kg for PFOS (Ptrend = 0.01); 4.3 (0.9) versus 2.2 (0.8) kg for PFOA (Ptrend = 0.007); 4.9 (0.9) versus 2.7 (0.8) kg for PFHxS (Ptrend = 0.009); 4.7 (0.9) versus 2.5 (0.9) kg for PFNA (Ptrend = 0.006); and 4.2 (0.8) versus 2.5 (0.9) kg for PFDA (Ptrend = 0.03) (Table 3). Significant interactions with sex were demonstrated for PFOA and PFHxS (Pinteraction = 0.04 and 0.01, respectively). When the covariates were entered into the model in a stepwise manner, these results did not change materially (S3 Table). The trajectory of changes in body weight in men and women according to tertiles of PFAS concentrations is shown in Fig 1. The trajectory of changes in body weight among total participants is shown in S1 Fig.
After multivariate adjustment, including baseline RMR and dietary intervention group, baseline plasma PFAS concentrations, especially for PFOS and PFNA, were significantly associated with a greater decline in RMR during the weight-loss period (first 6 months) and a lower increase in RMR during the weight regain period (6–24 months). During the first 6 months, comparing the highest to the lowest tertiles, the least-square means (SEs) of RMR change were −45.4 (15.5) versus −5.0 (16.3) kcal/day for PFOS (Ptrend = 0.005) and −49.8 (15.9) versus −3.3 (16.1) kcal/day for PFNA (Ptrend = 0.002) (Model 3 in Table 4). During the period of 6–24 months, comparing the highest to the lowest tertiles, the least-square means (SEs) of RMR change were 0.9 (26.2) versus 94.6 (27.5) kcal/day for PFOS (Ptrend < 0.001); 12.7 (28.1) versus 69.3 (27.3) kcal/day for PFOA (Ptrend = 0.03); 24.6 (28.5) versus 81.5 (27.5) kcal/day for PFHxS (Ptrend = 0.03); 14.1 (27.7) versus 73.7 (27.6) kcal/day for PFNA (Ptrend = 0.02); and 23.1 (27.6) versus 66.5 (28.2) kcal/day for PFDA (Ptrend = 0.09) (Model 3 in Table 4). The results were similar when PFAS concentrations were treated as continuous variables (Table 4). When adjusting for RMR at 6 months (instead of RMR at baseline), the results maintained statistical significance. When changes in RMR or changes in thyroid hormones during the first 6 months were further adjusted for, the results remained largely unchanged. In the sex-stratified analysis, similar results were observed, although some associations did not reach statistical significance, possibly due to diminished power (S4 Table). No interaction between PFASs and sex on RMR changes was detected. The trajectory of changes in RMR among total participants according to tertiles of PFAS concentrations is shown in Fig 2. In addition, similar results were demonstrated when analyses were stratified by dietary intervention group.
During the weight-loss period, after multivariate adjustment including baseline levels of each metabolic parameter, plasma concentrations of PFOS, PFNA, and PFDA were inversely associated with changes in visceral fat mass (rs ranged from −0.19 to −0.27, all P < 0.05), and baseline PFOA was inversely associated with changes in HDL cholesterol (rs = −0.12, P < 0.01) (S5 Table). During the weight regain period, baseline PFOS, PFNA, and PFDA levels were positively associated with changes in some of the parameters, including waist circumference, insulin, and leptin (rs ranged from 0.10 to 0.15, all P < 0.05), and baseline PFOA and PFHxS were associated with a greater increase in visceral fat mass (rs = 0.30 and 0.27, respectively; both P < 0.05) (S5 Table). The results were largely similar when analyses were stratified by sex. In sensitivity analyses, the results did not materially change when further adjusting for study location (Boston or Baton Rouge) or participant compliance (number of sessions participants attended). The table in S1 Text shows the associations of baseline PFASs with gene expression in adipose tissue.
In this 2-year randomized weight-loss trial, we found that higher baseline plasma PFAS concentrations were not associated with weight loss induced by energy restriction, but were significantly associated with a greater weight regain, primarily among women, during the follow-up period between 6 and 24 months. In addition, after multivariate adjustment, higher baseline PFAS levels were significantly associated with a greater decrease in RMR during the weight-loss period and a lower increase in RMR during the weight regain period.
To date, evidence on the influence of PFAS exposure on body weight change and metabolic parameters has been limited and has been primarily generated from cross-sectional studies that could not establish causal relationships [30,44–47]. In addition, the causes of weight change are likely heterogeneous (including diet, physical activity, and medications) and often not well understood in observational studies. Prospective evidence linking PFAS exposure with body weight regulation was primarily from studies that examined prenatal or early life exposures to PFASs in relation to body weight later in life, and the results were somewhat mixed [21–27,48,49]. For example, in 3 birth cohort studies conducted in European populations, maternal concentrations of PFASs were significantly associated with offspring body weight and other anthropometric and metabolic traits, primarily among girls [21,23,25]. However, other studies generated inconsistent findings regarding maternal PFAS exposure and offspring BMI or obesity risk, with no sex difference [22,24,49]. In addition, recently, in the European Youth Heart Study, Domazet et al. demonstrated that higher plasma PFOS concentrations during childhood, but not adolescence, were associated with greater adiposity in adolescence and young adulthood [48].
To our knowledge, the current investigation is among the first studies in adults to evaluate the associations of PFAS exposures with changes in body weight and metabolic parameters induced during a controlled weight-loss trial. All individual PFASs were significantly associated with more weight regain in women, but not in men, which was in agreement with previous studies in which the intergenerational effects of PFASs on body weight were observed only in girls and not in boys [21,25,26]. Although the reasons for these gender-specific findings are still unclear, accumulating evidence from experimental research suggests that PFASs are able to interfere with estrogen metabolism and functionalities [12,50,51]. As potential endocrine disruptors, PFASs might reduce estradiol production and the expression of some key genes related to estrogen synthesis [12], or influence estradiol concentrations through pathways such as hepatic aromatase induction, with an initial inhibition and a later stimulation [50]. Using in vitro and in silico species comparison approaches, Benninghoff et al. reported that PFASs may interact directly with estrogen receptors, suggesting that PFASs could act as weak environmental xenoestrogens [51]. The experimental evidence implies that the detrimental effects of PFASs can be sex-specific, thus supporting the notion that women may be particularly vulnerable to obesogenic effects of PFASs. In addition, it is worth noticing that women generally have a higher percentage of body fat than men [52]. Given that fat-free mass could substantially influence RMR, the difference in body composition between men and women could result in significant differences in energy homeostasis dynamics [52].
In addition to the adverse effects of PFASs on estrogen-related pathways, animal studies suggest that PFOA and PFOS may also interfere with energy homeostasis and the endocrine system through other mechanisms [14,15,18,53], including the activation of PPARα and PPARγ [18,19], key regulators in fatty acid oxidation, differentiation and normal function of adipocytes, and glucose metabolism [20,54]. An experiment on human liver cells suggested that PFOA could alter the expression of proteins regulated by hepatocyte nuclear factor 4α [55], which is a key regulator of lipid metabolism and gluconeogenesis [56]. In addition, some animal studies have suggested that PFAS exposure might disrupt thyroid hormone homeostasis, possibly via influencing uridine diphosphoglucuronosyl transferases and type 1 deiodinase [17,57]. Of note, due to the species-specific toxicokinetics (e.g., the elimination half-lives are 3–8 years in humans and 17–30 days in mice and monkeys) and tissue distribution of PFASs [18], caution is needed when extrapolating findings from animal studies to humans. In addition, mechanisms need to be elucidated to interpret the findings that higher baseline PFASs, especially PFOS and PFNA, were associated with changes in RMR, which is a major determinant of weight maintenance, in both men and women [58,59]. Finally, whether the 5 major PFASs might have different biological mechanisms and perhaps exert additive or synergistic effects also warrants further exploration.
The primary strength of the current study is that the cause of weight changes was well characterized. Unlike previous observational studies in which reasons for weight changes were usually unknown, this weight-loss trial applied energy restriction to induce the weight changes. Moreover, repeated measurements of body weight, RMR, thyroid hormones, leptin, and other metabolic biomarkers allowed documentation of longitudinal associations between PFAS exposures and changes in these parameters during the weight-loss and weight regain periods.
Several limitations should be considered as well. First, although we included men and women with a wide range of ages (30–70 years), participants in the current study were otherwise relatively homogeneous in terms of health status and body fatness because they were selected following narrow inclusion criteria. Therefore, it is unclear whether our findings can be extrapolated to more general populations. Second, we measured only the baseline plasma PFAS concentrations. However, given the long elimination half-lives (3–8 years) of these chemicals [36] and a strong stability over time observed in our pilot study, concentrations in the blood likely reflect relatively long-term PFAS exposures. Moreover, unlike many other persistent organic pollutants, PFASs are not lipophilic, and blood concentrations are therefore not affected by changes in the size of the lipid compartment [60]. Third, we did not measure ghrelin, an orexigenic hormone regulating appetite, RMR, and other key physiological processes related to weight changes [61], and the interrelationship between PFASs and ghrelin during weight changes needs to be elucidated. Fourth, we did not apply Bonferroni correction in the analyses given the inter-correlation between the PFASs (rs ranged from 0.4 to 0.9), and the role of multiple testing could not be entirely excluded. Fifth, physical activity was assessed using the Baecke questionnaire, which might be subject to measurement errors, although a validation study conducted in US adults has shown reasonable validity of this questionnaire [62]. In addition, although some covariates including education, smoking status, and physical activity were adjusted for in our study, we could not entirely exclude the possibility that unmeasured or residual confounding by socioeconomic and psychosocial factors, as well as participants’ usual diet, might partially account for the associations we observed. One particular concern is that PFASs are extensively used in food packaging due to their oil- and water-repellant characteristics [32]. If some participants relapsed to their usual pre-randomization diet and this diet was rich in foods that are contaminated by PFASs through food packaging and are also dense in energy, they might thus have gained weight faster. However, when we further controlled for the frequency of craving hamburgers, French fries, or donuts at baseline assessed using a questionnaire, the results were largely unchanged. In addition, humans are exposed to PFASs through multiple pathways, including drinking water and contaminated seafood [31], although these factors are not established risk factors for weight gain. Moreover, we adjusted for the number of study sessions that participants attended, which is a measurement of compliance to the prescribed diet. Finally, lipophilic persistent pollutants with obesogenic effects (such as hexachlorobenzene [HCB] and dichlorodiphenyldichloroethylene [DDE]) might have confounded the associations of PFASs with changes in body weight and RMR. However, in 793 women participating in the Nurses’ Health Study II, weak associations were observed between PFASs and lipophilic persistent pollutants (e.g., the rs of PFOA and PFOS with HCB was 0.07 and 0.06, respectively, and the rs of PFOA and PFOS with DDE was 0.05 and 0.06, respectively), suggesting that confounding by these pollutants would not be substantial.
Our study provides the first piece of evidence from a controlled weight-loss trial that higher baseline plasma PFAS concentrations in adults are associated with a greater weight regain, especially in women, possibly due to suppressed RMR levels. These findings imply that overweight and obese individuals with relatively low PFAS exposures might potentially benefit more from weight-loss interventions. Although the production of PFOS and PFOA in the US has largely been phased out [31,63], the production of other PFASs, such as PFNA, may continue or even increase, especially in developing countries [64]. Given the persistence of these PFASs in the environment and the human body, their potential adverse effects remain a public health concern.
In a diet-induced weight-loss setting among overweight and obese individuals, higher baseline plasma PFAS concentrations were significantly associated with greater weight regain, especially in women, accompanied by a slower regression of RMR. These findings suggest that environmental chemicals may play a role in the current obesity epidemic. More studies are warranted to elucidate the mechanisms underlying the link between PFAS exposure and weight regulation in humans.
|
10.1371/journal.pcbi.1004365 | Neutral Models of Microbiome Evolution | There has been an explosion of research on host-associated microbial communities (i.e.,microbiomes). Much of this research has focused on surveys of microbial diversities across a variety of host species, including humans, with a view to understanding how these microbiomes are distributed across space and time, and how they correlate with host health, disease, phenotype, physiology and ecology. Fewer studies have focused on how these microbiomes may have evolved. In this paper, we develop an agent-based framework to study the dynamics of microbiome evolution. Our framework incorporates neutral models of how hosts acquire their microbiomes, and how the environmental microbial community that is available to the hosts is assembled. Most importantly, our framework also incorporates a Wright-Fisher genealogical model of hosts, so that the dynamics of microbiome evolution is studied on an evolutionary timescale. Our results indicate that the extent of parental contribution to microbial availability from one generation to the next significantly impacts the diversity of microbiomes: the greater the parental contribution, the less diverse the microbiomes. In contrast, even when there is only a very small contribution from a constant environmental pool, microbial communities can remain highly diverse. Finally, we show that our models may be used to construct hypotheses about the types of processes that operate to assemble microbiomes over evolutionary time.
| Microbial communities associated with animals and plants (i.e., microbiomes) are implicated in the day-to-day functioning of their hosts. However, we do not yet know how these host-microbiome associations evolve. In this paper, we develop a computational framework for modelling the evolution of microbiomes. The models we use are neutral, and assume that microbes have no effect on the reproductive success of the hosts. Therefore, the patterns of microbiome diversity that we obtain in our simulations require a minimal set of assumptions relating to how microbes are acquired and how they are assembled in the environment. Despite the simplicity of our models, they help us understand the patterns seen in empirical data, and they allow us to build more complex hypotheses of host-microbe dynamics.
| Microbial communities associated with animals and plants (i.e., microbiomes) are implicated in the day-to-day functioning of their hosts in a variety of ways; microbes provide hosts with access to nutrients [1–4], they protect against pathogens [5,6], confer drought resistance [7], mediate hosts’ social interactions [8], and modulate host behavior [9–11]. Not surprisingly, many studies have related human health to the human microbiome. Researchers have proposed that perturbations in human microbiome composition may be associated with a range of disorders, including inflammatory bowel disease [12–14], anxiety and depression [15], obesity [16,17], autism [18,19], allergic responses [20] and respiratory ailments [21–23]. There is a growing number of large-scale projects underway to characterize and analyze the collection of microbes linked to human health and disease using advanced sequencing methods and technologies, including the Human Microbiome Project [24] and the European Metagenomics of the Human Intestinal Tract [25].
Many microbiome studies focus on descriptions of microbial dynamics within hosts over short periods [26–28], or on the composition and diversity of microbial communities amongst hosts with different phenotypes or under different treatment regimes [17,29], but there is increasing recognition that host-microbe dynamics need to be studied within the context of a unified ecological or evolutionary framework. In a recent paper, Costello et al [30] listed four ecological processes that mediate the diversity of human microbiota: environmental selection, whereby the host environment favors the presence and persistence of certain microbial taxa; historical contingency, in which differences in the timing and order of microbial colonization leads to differences in succession and climax communities; random sampling, where stochastic factors influence community assemblages; and dispersal limitation, where the availability of microbial taxa is restricted by the local structure of host communities and environments. Yeoman et al [31] discuss many of these same factors, and frame the challenge of understanding microbial diversity through the lens of evolution, identifying selection induced by competition between individuals of the same or different microbial species, and host influences including phylogeny, as important drivers of host-microbial variation.
In this paper, we take the first steps in developing an explicit framework for modelling the evolutionary and ecological dynamics of microbial communities within a population of hosts. The models we use are related to some of those that metacommunity theorists work with [32,33], specifically, neutral models of ecology and biodiversity [34], but there is one significant difference. Standard spatial models of community assembly, biogeography and biodiversity assume that areas available for colonization remain indefinitely and are static. These models do not apply to the evolution of microbiomes because the “spaces” these microbial communities colonize–that is, the hosts–share an evolutionary history characterized by lineages that persist or go extinct. For this reason, the pairing of host genealogy with microbiome assembly may lead to different patterns of biodiversity than those expected under existing models of microbial community assembly and metacommunity theory.
We develop an agent-based framework, similar to that used recently by Hellweger et al [35], to capture the emergent patterns of microbial diversity over many host generations. In contrast to Hellweger et al, we do not model mutations and speciation in microbial lineages but focus instead on how ecological and evolutionary sampling processes affect the standing variation of microbes in the environment and hosts. The neutral processes that are encapsulated in our framework are based on those that have been proposed by others [30,31] with the added dimensions of an evolutionary timescale and a genealogy of hosts, both of which contribute to shaping the eventual composition and variation of microbial communities within and between hosts. Our models do not assume that hosts differ in their reproductive success as a consequence of their microbial communities; nor do our models assume differences amongst microbes in their propensities to persist in the host or the environment, or in their abilities to disperse. Consequently, whereas our models do not capture the complexity of processes that are likely to mediate the evolution of microbiomes, they serve as minimalist null models against which empirical patterns may be compared.
Our framework is a generic one, and we have not developed it specifically for any one species of host, or indeed, any one systemic compartment within a host. Instead, our models incorporate simple host mechanisms for microbial acquisition, and we explore the effects of parental inheritance of the microbiome, and microbial recruitment from the environment. Our results indicate that parental inheritance tends to reduce microbial diversity and increase homogeneity within hosts, while ongoing environmental acquisition works to maintain microbial heterogeneity within hosts. Interestingly, we observe a non-linear relationship between the degree of parental inheritance and between-host differences in microbiome composition. Finally, we show how these neutral patterns allow us to make predictions about the processes that are important in shaping microbiome diversity when applied to empirical data.
Our framework applies to a population of hosts and an available pool of microbial colonists. As a first step, we assume that hosts do not exert any preferences on the microbial taxa they acquire. Similarly, we assume that microbes do not interfere with host reproductive capacity, or the survivorship and reproductive success of other microbes in the community. As will become clear, the only indirect effects that influence microbial recruitment and persistence from one generation of hosts to the next are competition for space within hosts and the relative abundance of microbial taxa. Simply put, we assume that the ecological and evolutionary processes that operate on hosts and their microbiomes are neutral; in this regard, our framework is analogous to neutral theories in evolutionary biology [36,37], ecology and biodiversity [34]. We expand on this analogy later, but for now, we note that neutral theories provide parsimonious accounts of the types of patterns that can emerge in complex systems, they serve as null models for statistical hypothesis tests, and they provide platforms upon which we may construct more elaborate representations of these same systems [38].
In this framework, hosts reproduce asexually in discrete generations, following a neutral Wright-Fisher process [39,40], where each individual in a succeeding generation chooses a parent randomly from the preceding generation. Hence, with a population of hosts of constant size N, all asexual individuals will share a common ancestor after 2N generations, on average. In our models, asexual reproduction is a computational convenience, and can be replaced with sexual reproduction without changing the essential patterns that we observe.
We model how hosts acquire their microbiomes in three ways (Fig 1). First, under a strict “parental-acquisition” (PA) process, all hosts acquire their microbial communities directly from their parents. Second, with strict “environmental-acquisition” (EA), hosts acquire their microbiomes solely from the environment. Between these two extremes, we also allow a third “mixed-acquisition” (MAx) process, whereby hosts acquire some percentage, x%, of their microbiomes from their parents and (100-x)% from the environment. MA0 is exactly equivalent to EA, and MA100 to PA; as such, EA and PA designate boundary conditions of the ecological processes that mediate microbial acquisition in hosts.
It is worth pausing at this point to clarify what we mean when we say that hosts acquire their microbiomes from their “parents” or their “environments”. Our models do not explicitly take account of the life events–illness, infections, changes in environments or diets–of each host within a generation, nor does it consider microbial fluxes within the lifespan of each host. Instead, the microbial composition of each host is essentially measured as an aggregate over the single generation that the host exists. Consequently, when we quantify the percentage of microbes from parents and environment using, say, MA10, we mean that over the life of the host, 90% of its microbes come from the environment and 10% from its parent. In our models, it is possible that the parental contribution happened in the first 10% of the host’s life, or it may be that over the entire lifespan of the host, there was an ongoing contribution by the parent that amounted to 10% of the microbial composition.
Since we allow hosts to recruit microbes from an “environment”, we need to define how the microbial content of this environment is constituted. In simulations, we characterize microbial composition using a distribution of taxa’ relative abundances. We propose three processes that determine the composition of the pool of microbes available for recruitment. First, we assume that the environment has a microbial composition that remains fixed over time. For the “fixed environment” (FE), all taxa are present in the environment throughout the simulation, and are available to every generation of hosts. The second process we propose involves a changing environmental microbial profile, whereby the relative abundance of each microbial taxon available to the hosts in a given generation, is an aggregate of their abundances from all hosts of the preceding generation. Under this “pooled-environment” (PE), microbial composition is reflection of what was present in the parents of the current generation of hosts. A third, intermediate, process is a combination of the previous two “environments”: the environmental microbial pool available for recruitment contains a percentage, y%, from the parental pool of microbes, and (100-y)% of microbes from the fixed environment. Under this “mixed environment” (MEy), the proportion of contribution from host microbiomes is given by y. As with our acquisition models, ME0 and ME100 are equivalent to the boundaries FE and PE, respectively.
Our framework allows us to combine different host-acquisition processes with different ways of constructing the pool of available microbes in the environment. Conceptually, each of these combinations is a particular neutral model, capturing some of the elements previously discussed in the literature. For instance, PA or MAx incorporate the phylogenetic dependencies that Yeoman et al [31] discuss, and EA x PE is equivalent to what Costello et al [30] call dispersal limitation, whereby the local host community influences microbial composition. It is worth noting that the combinations PA x (FE, MEy, PE)–read as “PA in combination with FE, with MEy or with PE”–will give identical results. This is because, in all cases, the environment contributes nothing to host microbial content (see the first row in Fig 1).
We model the construction of the microbial community in each host by competitive random sampling with replacement. Under this process, each host allows only a fixed and limited number of microbes to populate its microbiome. If microbial acquisition occurs under EA, each host samples randomly from the available pool of taxa according to the relative abundance of each taxon in the environment. In the case of MAx, x% of microbes are selected from the parent and (100-x)% from the environment. If hosts acquire their microbial taxa under PA, then all microbes are inherited from the hosts’ parents, although the relative abundance of each taxon fluctuates multinomially. By constructing microbial communities in this way, we allow stochastic factors and indirect competition to modify taxon composition within and between hosts, as proposed by Costello et al [30].
By simulating combinations of PA, MAx and EA against FE, MEy and PE forward in time over many host generations and over a range of conditions, we are able to recover data on the behavior of individual microbial taxa, as well as a variety of summary statistics, including the expected time it takes individual taxa to invade all hosts or go extinct in the host population, and the trajectories of microbial taxonomic richness (measured simply as the total number of microbial taxa) and microbial taxonomic evenness (measuring the similarity in the frequency of each taxon), microbial diversity within hosts (α-diversity), inter-host variation in microbial composition (β-diversity) and the aggregate microbial diversity from all hosts in the population (γ-diversity). Here, we report only on the latter three measures of diversity.
Microbial diversity within the host population is a function of the proportion of microbes that parents contribute directly to offspring and the proportion they contribute to the environment. Fig 2 illustrates how population-level taxon abundances change under various combinations of these proportions. In our simulations, the distributions of taxon abundances under high levels of parental contributions are skewed, and may be approximated by commonly-applied distributions, including the log-normal distribution and the Dirichlet multinomial (DM) distribution [41] (Fig 3; the DM distribution has the advantage of allowing α-, β- and γ-diversities to be simulated–see S2 Fig). The ability to recover skewed abundance distributions is interesting, because we begin our simulations with a uniform distribution of microbial taxa, and we retain this uniform distribution in the fixed environment throughout the evolutionary history of the host population. Consequently, the emergence of dominant and rare taxa is a consequence of repeated parental contributions either directly to the next generation of hosts or indirectly to the environment.
In fact, all simulations in which there was complete parental acquisition of microbes (i.e., PA) resulted in the loss of all but one microbial taxon in the host population. Similarly, when the environment was reconstituted each generation exclusively with microbes from the parents (i.e., PE), the same pattern was observed with only a single microbial taxon remaining. These result are consistent with predictions made under neutral models of community ecology [42], and highlight the strong depressive effect of parental transmission, either directly from parent to offspring or via parental contributions to a local pool of microbes, on population-level microbiome diversity.
With EA x FE, microbes are obtained randomly from a fixed environment that persists over the evolutionary history of the hosts; unsurprisingly, the host population retains all microbes found in the environment. Interestingly, when microbes are obtained both from parents and a fixed environment (MA x FE), we still see the persistence of all or almost all microbes in the host population (see Fig 2A and 2B, first column of each bar chart; ME(0) is equivalent to a fixed environment with no microbial contributions from parents). This is true even when the proportion of microbial taxa that an individual host acquires from the fixed environment at each generation is very small, on the order of 0.001. Therefore, a very small contribution from a constant environmental source of microbes is sufficient to retain high levels of microbial diversity in the host population.
Microbial diversities are frequently measured in three ways: α-diversity, β-diversity, and γ-diversity. Our simulations indicate that all three measures depend on the percentage of parental contribution to offspring microbiomes and the composition of the environmental microbial pool (see S1–S3 Tables for simulation means and standard deviations).
Under our neutral model, in which the absence of host sub-population structure means that all hosts sample their microbes from the same environment, α- and γ-diversities remain high, and β-diversity remains low, for a large part of the range of direct or indirect parental contributions (i.e., to offspring or to the environment, respectively). Nonetheless, at high values of parental contributions, there are discernible differences in diversities, and we have also focused our simulations in these areas (Fig 4; see S4–S6 Tables for simulation means and standard deviations). In general, α-diversity (average diversity within hosts) and γ-diversity (overall diversity within the entire population of hosts) increase as we increase the fixed environmental contribution because a fixed environment helps maintain a uniform distribution of taxon abundances and delays the loss of microbial taxa during evolution. Conversely, when hosts acquire increasing proportions of their microbiomes from their parents directly, or indirectly from a pooled environment, the variation of taxon abundance increases and taxon richness tends to decrease, thus lowering both α- and γ-diversities (Fig 4A and 4B; S1, S2, S4 and S5 Tables). Inter-host variation in microbial composition, or β-diversity, also depends on the degree of parental inheritance, and the ratio of fixed-to-pooled environmental components (Fig 4C; S3 and S6 Tables).
Under the combination of PA x (FE, ME or PE), β-diversity tends to zero, because all hosts descend from a single common ancestor and, as noted above, only a single microbial taxon remains in all hosts. When we have the parental microbiome as the only source of microbiomes in the next generation, ultimately, all lineages will have acquired their microbiomes from the most recent common ancestor (MRCA) of the population of hosts. Additionally, from one generation to the next, stochastic sampling of microbes over evolutionary time will result in the loss of all but one microbial taxon. Interestingly, with a high percentage of environmental acquisition, β-diversity is also relatively low, because all hosts acquire a large proportion of their microbial taxa from the same environmental pool, and consequently, will tend to acquire the same set of taxa. As noted above, the highest β-diversity occurs in a relatively narrow range of values of pure parental acquisition (between 87–99% of direct parental transmission; S6 Table).
If we focus on the relationship between β-diversity and the environmental pool, we see that its behavior is similar to that of α- and γ-diversities: it decreases as we increase the pooled environmental contribution to offspring microbiome. This is because a pooled environment, with contributions from the parental generation, tends to give rise to a non-uniform distribution of microbial taxa. As the degree of parental contribution increases, the environmental community will be dominated by few highly abundant species which are likely shared by most or all hosts within the population, accounting for high between-host similiarity in microbial composition (Fig 4C; S3 and S6 Tables).
It is important to note that our simulations have not been performed with inference or prediction in mind: the number of hosts, the number of microbes, and the number of taxa in our simulations are not necessarily equivalent to those of real-world microbial communities and their hosts, nor have we necessarily chosen the appropriate diversity indices or taxonomic resolution to optimize prediction/inference. Nonetheless, it is helpful to examine how the simulated values of diversity compare to empirical observations, and what these comparisons might tell us about the evolutionary processes that are acting on microbiomes. As an example, we used genus-level taxonomic data from the NIH Human Microbiome Project (HMP) [43], specifically, a table of relative abundance found in different compartments of the human body (http://www.hmpdacc.org/HMSMCP/; see S1 Dataset). Several large samples from the anterior nares, vaginal posterior fornix, stool, buccal mucosa, tongue dorsum and supragingival plaque were chosen to calculate α-, β- and γ-diversities on genus level (Table 1).
Values of α- and γ-diversity obtained from all sampled sites of the human microbiome are low, in comparison to most of the values we obtained in our simulations. In fact, human microbiome diversities are generally lower than those of other non-human primates [44,45]. If we compare the empirical diversities to those obtained in our simulations (S1–S6 Tables), we would have to posit very high parental contributions, both direct and indirect (>90%), to account for the α- and γ-diversities across all human body sites. In contrast, values of β-diversity appear to provide a little more discrimination amongst body sites: the site with the lowest β-diversity is the vaginal posterior fornix, and its value is consistent with a very low degree of direct parental contribution in our simulations (approximately between 0–15%). The β-diversities at other sites appear to suggest higher levels of parental contribution (again, >90%). In the next section, we discuss the implications of these results, as they relate to human microbiome evolution and how the neutral model may be used to construct hypotheses about relevant evolutionary and ecological processes.
In this paper, we introduce a simple and flexible framework to model the evolution of microbiomes within a population of hosts, which takes account of different modes of microbiome acquisition and environmental microbial composition. Under our neutral model, microbiome composition is affected by sampling effects. Stochastic changes in microbial abundances may affect the persistence of microbial taxa in the microbiome over one or a few generations (i.e., ecological drift), or over many generations. The latter may occur because host lineages die out; when this happens, changes in microbial abundance across the whole population of hosts are essentially equivalent to changes in allele frequencies (i.e., genetic drift).
The constitution of the microbial community in the environment also plays a considerable role in determining the ultimate fate of microbial taxa within host microbiomes. With a fixed environment, when there is a constant pool of the same microbial taxa from one generation of hosts to the next, microbial taxa never go extinct from the host population as long as hosts obtain some fraction of their microbiome from the environment. This is true, even when that fraction that the environment contributes to each host’s microbiome is very small (e.g., 0.1%). In contrast, when the environmental composition of microbes reflects the microbial content of the hosts in previous generations (i.e., PE, the “pooled” environment in our model), microbial diversity of the environment shrinks, as does the diversity of host microbiomes. Therefore, the extent to which parents contribute to the microbiomes of their offspring (either directly or through their contributions to the pooled environment) plays a crucial role in shaping microbiome diversity and constitution. In our simulations, values of α- and γ- diversities are at their lowest when parental contribution to the microbiome is high. Inevitably, microbial taxa are lost from the population as host lineages are lost.
Thus, under our neutral model, it is possible to recover skewed microbial abundance distributions reminiscent of those obtained with real data, despite a fixed environmental component that remains uniform and constant throughout our simulations. Increasing skewness–essentially, decreasing eveness–is obtained as we increase the degree of parental inheritance. Of course, we don’t claim any deep insight here: no one should be surprised that we are able to recover skewed abundance distributions with our models, because there is a large body of literature on the mechanisms–both neutral or otherwise–that may lead to the emergence of skewed abundance distributions (see [46] for an excellent synthesis). Our results reinforce what others [34,47,48] have found, by adding yet another neutral mechanism to account for the emergence of skewed abundance distributions.
Our framework includes sampling effects on an undivided host population, which evolves under a Wright-Fisher process. Consequently, our models have some points of similarity with those that have been developed in population genetics. For instance, Orive et al [49] analyze the evolutionary dynamics of endosymbionts using a discrete-time Moran population genetic model. In their model, endosymbionts are acquired either vertically, passed on from parent to offspring, or horizontally from the environment. This corresponds to our MA x FE model and, in agreement with our results, Orive et al find that increasing the environmental contribution of endosymbionts to host cells results in greater diversity within cells and less diversity between cells.
Our models do not include any mutational process or speciation acting on the microbes, as time moves forward. In reality, of course, microbes acquire mutations in their genomes at a rapid rate, but the measures of diversity we use in our analyses capture differences in taxonomic composition, not genetic diversity. In our models, it is implied that no cladogenetic events have occurred over the course of the simulations.
The models presented here provide an opportunity to construct hypotheses, and make qualitative predictions, about the patterns of diversity we can expect to find in different biological situations. For example, the effects of “pure” pooled versus fixed environments on microbiomes can be found in a comparison of social and solitary bees. Social bees exhibit behaviors that are likely to result in the transmission of microbes from a microbial pool within the colony [50]. In contrast, solitary bees acquire their microbiomes from the environment, through feeding or burrowing. Our model would predict that social bees would have lower α-diversity and lower taxonomic richness than solitary bees. This is consistent with the results obtained by Martinson et al [51] who surveyed the microbiomes of eusocial bee species Apis spp. and Bombus spp. and non-social bees (11 species) and wasps (3 species): they found depauperate microbiomes in social bees compared to non-social bees.
Whereas it is reassuring to obtain empirical corroboration for our models, arguably neutral models are most useful when real-world observations run counter to predicted outcomes. Falsification of neutral models provides a justification for augmenting these models to include additional processes that account for the phenomena under study. In this regard, our analysis of the Human Microbiome Project data is instructive. As we have noted above, our simulations should not be used for inference, and we should be cautious about reading too much into the comparisons between empirical and simulated patterns of diversity. Nonetheless, at least for some sites, i.e., the stool, the tongue dorsum, the supragingival plaque, the anterior nares and the buccal mucosa, the low empirical values of α- and γ- diversities appear to point consistently to a high level of parental inheritance when compared against values obtained in our simulations. There is evidence that the human microbiomes at various sites are seeded at birth by the mother, particularly if this birth is through the vaginal tract [52]. There is also reasonably strong evidence that families share microbes to a greater extent than unrelated individuals in a population [53], and at least in some human populations, mothers share more microbes in common with their offspring than with unrelated children [54]. It is not clear, based on the studies that have been done to date [55], whether the values of direct or indirect parental contribution we obtain when we compare empirical and simulated diversities are significantly higher than would be obtained in real populations, but we expect that the intuition of mosts microbial ecologists is that percentages of direct and pooled parental contributions > 90% are likely to be too high. Putting to one side the caveats about inference, we accept that while this intuition does not constitute evidence against the neutral model, it is likely to engender scepticism about the model’s correctness. If it is, in fact, true that direct or indirect parental contributions to the next generation’s microbiomes are not as high as our simulations suggest, how do we account for the apparent depression in α- and γ-diversities, and elevation of β-diversities at these sites? One hypothesis that explains low α- and γ-diversities, and high β-diversities, and does not require the action of non-neutral processes, is the existence of local host subpopulations. The existence of subpopulations of hosts, with limited immigration and sharing of microbes between subpopulations, is likely to give the appearance of high parental contribution from one generation to the next. Certainly, this is a plausible explanation for patterns of microbiome diversity in the oral cavity (i.e., the buccal mucosa, tongue dorsum and supragingival plaque) and stool samples, because of the likely influence of familial [53] or cultural dietary preferences/practices [56] or lifestyles [57] on these microbiomes. A similarly explanation may account for patterns of microbiome diversity of the anterior nares.
The vaginal posterior fornix presents an interesting contrast to the other body sites because the α- and γ-diversities suggest high parental contributions (although they cannot distinguish between direct or indirect contributions), whereas β-diversity suggests a low direct parental contribution. This inconsistency may again cause us to reject the neutral model in favor of an alternative explanation, but in this case, subpopulation structure may play a minor role relative to selection for a vaginal microbial community that is common amongst hosts. Such a selective filter is likely a consequence of a complex suite of factors including host immune defences, hormonal cycling, pregnancy, and the presence of apparently beneficial microbial species (e.g., Lactobacillus spp.) [58]. This hypothesis explains both the high level of α- and γ-diversity (i.e., a few abundant species with many rare species), and the low β-diversity.
For the human microbiome, neutral models have the potential to help identify additional processes that may account for patterns of diversity. As noted, of the two processes identified above–host subpopulations and selective filters–the former still remains part of an underlying neutral process, and a plausible extension to the neutral framework presented here. Rejection of a simple neutral model therefore allows us to identify incremental additions that may increase explanatory power.
Another example of empirical data that appears to contradict the expectations of our models is the comparison of microbiome diversities in high microbial abundance (HMA) and low microbial abundance (LMA) sponges [59]. HMA sponges have large numbers of associated microbes, in contrast to LMA sponges. Additionally, researchers have shown that microbial diversity in LMA sponges is lower than that of HMA sponges [60,61]. Based on our results, we would predict that there is a greater degree of vertical transmission in LMA sponges, but it turns out that this is not the case: Schmitt et al [62] have found that “vertical transmission, as a mechanism to obtain bacteria, seems to occur mainly in HMA sponges”. Giles et al [60] propose two possible reasons to account for the low diversity in LMA sponges. First, there may be selective filters that permit only certain microbial taxa to colonize the sponges; second, the initial colonization event is stochastic, but serves to constrain or exclude successive colonizations. As with the human microbiome data, the sponge example is important because it does not rely a priori on non-neutral processes to account for the low diversity in LMA sponges; instead, selection (or other ecological and/or evolutionary processes) is invoked only after it is shown that vertical transmission in LMA sponges is unlikely, thus indicating that our neutral models are an inadequate explanation for the observed data.
Riffing on the theme that “Essentially, all models are wrong, but some models are useful” [63], Hubbell, writing about models in community ecology, says “Probably no ecologist in the world with even a modicum of field experience would seriously question the existence of niche differences among competing species on the same trophic level” [64]. But, he continues, “[Neutral theory] begins with the simplest possible hypothesis one can think of … and then adds complexity back into the theory only as absolutely required to obtain satisfactory agreement with the data”. We agree with Hubbell: to paraphrase, given what we know about the interplay between hosts, their microbial communities, and the environment, we would hesitate to put money on the table and bet that many microbiomes have evolved under the simplest neutral models that we have constructed here. But we would be equally hesitant betting in favor of the null hypotheses evaluated in statistical tests of significance. The value of these hypotheses resides not in their rightness or wrongness but in their ability to protect against overconfidence in our favorite, more complex model. Whereas it is true that biological processes are frequently complex, Occam’s Razor dictates that we construct as simple explanations (or models) as possible. In this way, we remain vigilant against the addition of unnecessary and unjustifiable complexity. Much as we do with statistical hypothesis tests, we accept stronger alternative explanations only when we are sufficiently confident that our neutral hypotheses are unlikely.
This is not to say that neutral models only serve as strawmen; in molecular evolution, for instance, neutral models are frequently effective at explaining molecular variation [65]. And even in cases when the assumption of neutrality is questionable, the use of neutral models of substitution applied in molecular phylogenetics does not appear to jeopardize the accuracy of tree reconstruction [66]. Consequently, without taking account of the evolutionary processes of mutation, speciation, selection or recombination, or the ecological processes that operate in the context of spatial, environmental, and temporal heterogeneity, what we have developed is a framework on which we can begin to evaluate empirical patterns of diversity, and where necessary, add more elaborate ecological and evolutionary scenarios. We believe that even this simple framework, devoid as it is of all the embellishments afforded by evolution and ecology, can serve a useful purpose: it is a suitable staging ground on which we can construct null models of microbiome diversity in populations of hosts and it allows us to make strong, testable predictions.
Simulated host populations consisted of a fixed number of virtual host individuals (N = 500). Each host was allocated a virtual microbiome with a limited capacity or "slots" of microbes (n = 1000). The environmental pool consisted of 150 microbial taxa. Large number of hosts (N = 2000), microbes (n = 100000) per host and microbial taxa (m = 500) were also simulated with our neutral model, and similar patterns of diversity were observed (S1 Fig).
The microbiomes of the initial generation of host individuals were seeded randomly, with bacteria sampled from a uniform distribution of taxon abundances. We used an initial uniform distribution of taxa because we wanted to ascertain whether the equilibrium distribution of abundances obtained at the conclusion of our simulations would recover patterns seen in natural microbiomes.
For each subsequent generation, the microbiome of each individual host was simulated by populating each of the available "slots" in the individual's microbiome by sampling microbial taxa with replacement (multinomial choice) from either the environment (with probability given (1-x)) or from the microbiome of a "parent" host individual (selected with uniform random probability from the population of the previous generation).
When sampling from a parental/environmental microbial community, the probability that the new host microbiome will acquire a particular microbial taxon is given by the relative abundance of that taxon within the community (see below for details on how environmental microbial taxon abundances were calculated).
The probability, x, that a particular "slot" in a new individual host's microbiome was occupied by a microbial taxon sampled from a randomly selected parent was varied across simulations. Two sets of simulations were performed: (1) x and y varied linearly, between 0 and 1, with increments of 0.1 (see S1–S3 Tables for means and standard deviations of diversities); and (2) with values of x, y ∈ (0.0, 0.5[0,1,2…10]) (see S4–S6 Tables for diversities). When x = 0.0, a host’s microbiome was sampled directly from the environment, i.e., the probability of a microbial taxon being selected was equal to the relative frequency of the microbial taxon in the environment. When x = 0.50 = 1, a host's microbiome was sampled multinomially from the microbiome of its "parent" who was selected with uniform random probability from the previous generation. Similarly, when y = 0.0, there is no contribution from the previous host generation to the environmental microbial resource, whereas when y = 0.50, all microbes are replaced each generation by the pool of microbes resident in the hosts of the previous generation.
Three different models were used to model the relative abundance of taxa in the environment. First, for FE ("fixed environment"), the abundances of the microbial taxa in the environment were fixed to the initial uniform distribution and did not vary over the course of the simulation. Second, under PE ("pooled environment") the abundances of the microbial taxa in the environment were composed of the pooled microbiomes of all the hosts of the previous generation, i.e., by summing over the abundances of the respective microbial taxa in hosts’ microbiomes, and renormalizing to relative abundances. Third, under MEy (“mixed environment”), the abundances of the microbial taxa in the environment were calculated by combining the fixed environment and pooled environment with y% derived from the pooled environmental component.
A particular simulation regime consisted of a distinct combination of pooled/fixed environmental ratios and environmental factors. A total of ten replicates were run under each simulation regime. The plots of γ-diversity were inspected, and simulations were halted when these stabilized. The number of generations for each simulation varied between 104 and 106 generations.
After each generation was simulated, the diversities of microbiomes are measured by scaled Shannon-Wiener index (α-diversity and γ-diversity) or Bray-Curtis dissimilarity index (β-diversity). The scaled Shannon-Wiener index is calculated as
−∑i=1Rpiln(pi)lnR
, where R represents the total number of taxa and pi represents the relative abundance of ith taxon within the community. The calculation of the Bray-Curtis index is given by the formula:
2(n−1)n*∑i=2n∑j=1i−1∑k=1m|pik−pjk|2
, where n represents the total number of hosts in the population, m represents the total number of microbial taxa within the host population, and pik and pjk represents the relative abundance of kth taxon within the community of host i and host j.
Empirical data from the human microbiome were obtained from the website of the NIH Human Microbiome Project (http://www.hmpdacc.org/HMSMCP/). The original community profiling data is a table of relative abundances for each of 690 samples and 718 taxa of bacteria and archaea (from 2 kingdoms to 397 species). As is described on the HMP website, all the samples were collected from 16 body sites from 103 healthy humans and processed with Whole Genome Shotgun sequencing. Specific information of each sample is available on the website (http://www.hmpdacc.org/HMIWGS/all/). We selected data of microbial communities associated with anterior nares, buccal mucosa, supragingival plaque, stool, tongue dorsum and vaginal posterior fornix because of their large numbers of samples, and removed the replicated samples from the same human subject for each body site (see S1 Dataset). The HMP data provides relative abundances for different taxonomic levels. We calculated diversities for all taxonomic levels ranging from species to kingdom, and found that the values of diversities for genus, family and order were similar. Consequently, we chose to use genus-level diversities. The relative abundances of genera for each site and all samples is given in S1 Dataset.
Fitting simulated results into log-normal and Dirichlet-multinomial distributions was performed in R with methods “fitdistr” of package MASS and “dirmult” of package dirmult.
All simulations were carried out using Python scripts and Java programs, available from https://github.com/qz28/microbiosima.git
Java code: https://github.com/qz28/microbiosima/tree/master/java
Python scripts: https://github.com/qz28/microbiosima/tree/master/python
|
10.1371/journal.pntd.0007000 | Toll-like receptor-2 regulates macrophage polarization induced by excretory-secretory antigens from Schistosoma japonicum eggs and promotes liver pathology in murine schistosomiasis | Schistosomiasis is endemic to many regions of the world and affects approximately 200 million people. Conventional adaptive T cell responses are considered to be the primary contributors to the pathogenesis of Schistosoma japonicum infection, leading to liver granuloma and fibrosis. However, the functional polarization of macrophages and the associated underlying molecular mechanisms during the pathogenesis of schistosomiasis remains unknown. In the present study, we found that excretory-secretory (ES) antigens derived from S. japonicum eggs can activate macrophages, which exhibit an M2b polarization. Furthermore, ES antigen-induced M2b polarization was found to be dependent on enhanced NF-κB signaling mediated by the MyD88/MAPK pathway in a TLR2-dependent manner. In addition, the cytokine profile of the liver macrophages from wild-type-infected mice are quite distinct from those found in TLR2 knockout-infected mice by quantitative PCR analysis. More importantly, the size of granuloma and the severity of the fibrosis in the livers of TLR2-/- mice were significantly reduced compared to that in WT mice. Our findings reveal a novel role for M2b polarization in the pathogenesis of schistosome infection.
| Schistosomiasis is a global health concern that affects primarily tropical and subtropical areas. During a schistosome infection, the eggs are trapped in the host liver and products derived from eggs induce a polarized Th2 response, resulting in granuloma formation and eventually fibrosis. Thus, it is important to elucidate the mechanism of granuloma formation and fibrosis development. Here, we show that activated macrophages play a novel role in the promotion of hepatic granuloma formation and liver fibrosis in a Schistosoma japonicum-infected mouse model. In addition, M2b polarization induced by egg products was dependent on enhanced NF-κB signaling mediated by the MyD88/MAPK pathway in a TLR2-dependent manner. Our findings reveal a novel role and mechanism of M2b polarization in the liver pathogenesis in S. japonicum-infected mice.
| Schistosomiasis is one of the most important health problems in developing countries[1], and can be used as a chronic disease model for investigating the interplay between the immune response and parasite pathogenicity in the host[2]. Following a schistosome infection, the host immune response gradually switches from a predominant Th1 response to a Th2-dominated response following egg deposition [3]. The resulting Th2 cytokine secretion contributes to the development of hepatic fibrosis and portal hypertension[4, 5]. A lipid fraction from Schistosoma mansoni eggs containing lysophosphatidylserine (lyso-PS) has been shown to induce dendritic cell (DC) activation that promotes Th2 and regulatory T-cell development via a Toll-like receptor (TLR)2-dependent mechanism[6]. Moreover, soluble S. japonicum egg antigens (SjEA) can upregulate programmed death ligand 2 (PD-L2) expression on bone marrow-derived dendritic cells (BMDCs) in a TLR2-dependent manner to help inhibit T cell responses in S. japonicum infected mice[2]. More importantly, these data indicate that interactions between host TLRs and pathogen-associated molecular patterns (PAMPs) from schistosome eggs can initiate a Th2-biased immune response and contribute to the egg-induced immunopathology observed in schistosomiasis.
Specialized pattern recognition receptors (PRRs) that recognize PAMPs, and the activation of such PRRs leads to an immediate innate immune response to infection and can profoundly influence the development of an adaptive immune response[7]. Among these PRRs, TLRs are type-1 transmembrane glycoproteins that can identify particular PAMPs and danger associated molecular patterns (DAMPs)[8]. TLRs are well-known to defend against pathogen invasion by triggering innate immune responses and subsequently priming adaptive immunity against infections, including Gram-positive and negative bacteria, as well as fungi, viruses, and parasites[9]. However, some TLRs can trigger a suppressive immune response through the binding of various ligands, which can help avoid excessive inflammation and develop chronic course of the disease, especially in helminth infections[10]. TLRs are expressed on various immune cells, including T cells, B cells, dendritic cells (DCs), and macrophages[11]. TLR engagement results in the activation of the mitogen-activated protein kinases (MAPKs), which, together with the NF-κB pathway, induce extracellular signaling to initiate specific cellular responses[12].
Macrophages, the most abundant mononuclear phagocytes in the human body, are heterogeneous, versatile cells that can undergo dynamic switches in phenotype or function in response to microenvironmental signals[13]. Functional macrophage polarization represents different extremes of a continuum ranging from M1, M2a, and M2b to M2c [14], which can cause different cell populations to display differential gene expression and distinct functions [15]. M1 polarization, driven by IFN-γ and LPS, typically acquires fortified cytotoxic and antitumoral properties, whereas M2 polarization generally obtains immunoregulatory activities, tissue repair, and remodeling[16]. In particular, M2a polarization is induced by IL-4 and IL-13, whereas M2b polarization, induced by immune complexes and TLR or IL-1R agonists, is characterized by an IL-10high and IL-12low phenotype, exerts immunoregulatory functions, and drives Th2 responses. In contrast, M2c polarization, induced by IL-10 and glucocorticoid hormones, results in immunosuppression and tissue-remodeling activities[14]. Although the critical role of macrophage activation in the pathogenesis of schistosomiasis has been validated[17], the precise phenotype and mechanism associated with functional macrophage polarization in schistosomiasis remains unclear.
In this study, we identified a novel role for macrophages in liver pathogenesis using a S. japonicum-infected mouse model and present TLR2 signaling as a novel potential therapeutic target for schistosomiasis.
All animal experiments were performed in strict accordance with the Regulations for the Administration of Affairs Concerning Experimental Animals (approved by the State Council of People’s Republic of China), and efforts were made to minimize suffering. All procedures performed on animals in this study were approved by the Laboratory Animal Welfare & Ethics Committee (LAWEC) of National Institute of Parasitic Diseases (Permit Number: IPD-2016-7).
Female C57BL/6 mice (6- to 8-weeks-old) were purchased from the SLAC laboratory (Shanghai, China). TLR2-/- mice[18] were provided by Dr. Xiao-Ping Chen from the School of Medicine, Tongji University. All mice were maintained under specific pathogen-free conditions and fed with standard laboratory food and water. Gender and age-matched mice were infected percutaneously with 20 ± 1 cercariae of S. japonicum, which were shed from infected Oncomelania hupensis snails provided by the National Institute of Parasitic Diseases in Shanghai, China.
S. japonicum eggs were isolated from the livers of female rabbits 6 weeks following infection with 800 − 1000 cercariae via abdominal skin penetration. ES antigens were prepared as described previously with modifications[19]. The collected eggs were washed twice in serum-free DMEM supplemented with 100 U/mL penicillin and 100 μg/mL streptomycin. The eggs were then resuspended in 24 mL DMEM, and 2 mL aliquots were placed in six-well culture plates (Corning, USA). The culture medium was harvested after 48 h and centrifuged for 10 min at 200 × g to remove eggs, and 10000 × g to remove any debris. The protein concentration of ES antigens was determined using a Bradford assay. The endotoxin level of ES antigens was <0.03 EU/mL as determined by a Limulus amoebocyte lysate assay (Genscript, China) according to the manufacturer’s instructions. ES antigens were stored at -80°C until further use.
To destroy the lipid structures, ES antigens were digested with phospholipase C (Sigma, USA) at 37°C for 12 h, followed by heat inactivation of the enzymes at 100°C for 10 min. To digest proteins, ES was treated with proteinase K (Sigma, USA) at 56°C overnight, followed by heat inactivation of the enzymes at 100°C for 10 min. Mock-treated ES was also performed by heat at 100°C for 10 min without the addition of enzymes. Protein disruption was regularly checked by SDS-PAGE and viewed by silver staining.
BMDMs were prepared as previously described with modifications[20]. Briefly, bone marrow cells were isolated from the leg bones of wild-type and TLR2-/- mice and cultured in DMEM (Gibco, USA) supplemented with 10% FBS (Gibco, USA) and 50 ng/mL macrophage colony-stimulating factor (M-CSF) (Peprotech, USA) and maintained in a 5% CO2 incubator at 37°C. Six days after the initial BM cell culture, the medium was changed, and the purity of F4/80+ cells was > 99%, as determined by flow cytometry.
In some experiments, BMDM cells (5 × 105 cells/mL) were pretreated with one of the following inhibitors: 10 μM BAY 11–7082 (NF-κB inhibitor, Beyotime biotechnoogy, China), 10 μM SP 600125 (JNK MAPK inhibitor, Beyotime biotechnoogy, China), 1 μM SB 203580 (p38 MAPK inhibitor, Beyotime biotechnology, China), 10 μM PD 98059 (ERK MAPK inhibitor, Beyotime biotechnology, China) and 10 μM ST 2825 (MyD88 homodimerization inhibitor, MedChem Express, USA). Furthermore, LPS from Escherichia coli serotype O111:B4 (Sigma, USA) and synthetic lipoprotein Pam3CSK4 (InvivoGen, USA) were used in some experiments.
The total RNA was extracted from macrophages using Trizol reagent (Invitrogen, USA) and reversed transcribed using a cDNA reverse transcription kit (Takara, Japan). The reverse-transcribed cDNA was used as a template in qPCR reactions containing SYBR Green Real-time PCR Master Mix (Takara, Japan) and 0.4 μM forward and reverse primers. Relative mRNA expression was calculated using the 2-ΔΔCt method and normalized to glyceraldehyde-3-phosphate dehydrogenase (GAPDH). The primer sequences were prepared as previously described[16, 21] and are listed in S1 Table.
BMDMs were stimulated with different antigens for 24 h. The levels of IL-1β, IL-12p70, IL-6, MCP-1 (CCL2), TNF-α, and IL-10 in the supernatants were detected by ELISA (eBioscience, USA). Quantification of IFN-γ, IL-4 and IL-13 in the serum sample was also determined by an ELISA in accordance with the manufacturer’s instructions (eBioscience, USA) and expressed as pg per mL.
Macrophages (5 × 105 cells/mL) were stimulated with various doses of ES (0.1 μg/mL, 1 μg/mL, and 10 μg/mL), LPS (100 ng/mL), and Pam3CSK4 (4 μg/mL) for different time points. The treated cells were washed twice with PBS and then lysed for 30 min on ice in a RIPA solution containing a protease inhibitor cocktail and phosphatase inhibitors (Sigma, USA). The expression of proteins in the cell lysates were examined using anti-NF-κB phospho-p65, anti-phospho-JNK, anti-phospho-p38, and anti-phospho-ERK1/2 antibodies (Cell signaling technology, USA). Anti-GAPDH (Sungene Biotech, China) was used as an internal control. Statistical analysis was performed for band intensities and evaluated using image J (NIH, USA).
The cellular suspension of liver leukocytes was prepared using the traditional method according to previous reported methods with modifications [22]. In brief, following a perfusion of 3 mL PBS via the portal vein, mouse liver fragments were pressed through a 70-μm cell strainer (BD, USA). The total liver cells were then resuspended in a 40% Percoll solution (GE Healthcare, USA), and centrifuged for 20 min at 800 × g. The leukocytes were resuspended in an erythrocyte-lysing buffer. The cells were washed and resuspended in a MACS separation buffer (Miltenyi Biotec, Germany), and anti-F4/80 microbeads (Miltenyi Biotec, Germany) were used to isolate macrophages from leukocytes. The purity of the isolated cells was confirmed at > 95%. In some experiments, liver macrophages with a purity of approximately 95% were used as the starting material for ES antigen stimulation. To assess the expression of activation and other biological markers on macrophages, flow cytometry was performed with FITC labeled anti-F4/80, APC labeled anti-Gr-1, Brilliant Violet 421 labeled anti-CD11b, PE labeled anti-MHC class II, PE-Cy7-labeled anti-CD40, PE-Cy7-labeled anti-CD80, APC-labeled anti-CD86, PE-labeled anti-CD16/32 and APC-labeled anti-mannose receptor (CD206) (Biolegend, USA). All flow cytometry data was acquired on an LSRFortessa X-20 (BD Biosciences, USA) and analyzed with FlowJo software (Tree star, USA).
Fresh liver tissues were fixed in 4% formaldehyde overnight and routinely paraffin embedded. Paraffin sections (5 μm) were prepared from each liver tissue sample. H&E staining of liver tissue sections were performed according to the manufacturer’s instructions and assessed by a pathologist blinded to the treatment group. The liver tissue sections were stained with Masson’s trichrome staining to evaluate collagen content and distribution. The collagen fibers were stained blue, the cell nuclei were stained black, and the background was stained red. Each stained section was examined by optical microscopy with 100 × magnification and identical settings. Thirty pictures were taken of granulomas around single eggs from three sections in each tissue. Every picture was evaluated in a double-blind fashion by two independent investigators. The area featuring granulomas and fibrosis surrounding single eggs was evaluated using image J (NIH, USA).
Data represented as the mean ± SEM were analyzed by a two-tailed Student’s t-test, or a one-way or two-way ANOVA using GraphPad Prism version 5.0 (GraphPad Software, USA). Significant differences were accepted when the p-value was less than 0.05.
To confirm that ES stimulation could induce macrophage polarization, BMDMs were stimulated with different concentrations of ES in vitro. Enhanced CD86 expression was observed in the BMDMs that were treated with 0.1 μg/mL or 1 μg/mL of ES compared with that of the control group (Fig 1A). However, CD206 (mannose receptor) and CD16/32 expression did not increase after ES stimulation (Fig 1A). Moreover, compared with 0.1 μg/mL ES, stimulation with 1 μg/mL ES was found to upregulate MHC class II (I-A/I-E) and CD86 expression but downregulate CD80 expression. However, compared with 1 μg/mL of ES stimulation, 10 μg/mL of ES stimulation was found to downregulate MHC class II, CD16/32, and CD86 expression (Fig 1A). RT-qPCR analysis showed that BMDMs stimulated with ES antigens exhibited enhanced IL-6, IL-10, and Arg-1 mRNA levels in a dose-dependent manner (Fig 1B). Furthermore, 1 μg/mL of ES antigen stimulation increased the levels of IL-6, TNF-α, MCP-1, IL-10, and Arg-1 mRNA expression but decreased the levels of iNOS and Ym1 mRNA compared with 0.1 μg/mL ES antigen stimulation. However, RT-qPCR revealed that the levels of TNF-α and IL-12 mRNA expression were downregulated in BMDMs stimulated with 10 μg/mL of ES antigens compared to BMDMs stimulated with 1 μg/mL of ES antigens (Fig 1B). Similar to these results, significantly higher levels of TNF-α, IL-12p70, MCP-1 (CCL2), and IL-1β were observed in the supernatants of BMDMs stimulated with 0.1 μg/mL ES compared with that of the control macrophages (Fig 1C). Moreover, stimulation with 1 μg/mL ES antigens upregulated IL-6, IL-1β, and IL-10 expression but downregulated IL-12p70 expression compared with 0.1 μg/mL ES antigen stimulation. However, compared with BMDMs stimulated with 1 μg/mL of ES antigens, the levels of IL-1β, TNF-α, MCP-1, and IL-6 were downregulated in the supernatants of BMDMs stimulated with 10 μg/mL ES (Fig 1C). Surprisingly, unlike LPS, ES stimulation did not increase iNOS protein expression but enhanced the production of Arg-1 protein (Fig 1D). According to the reported secretory products and biological surface markers associated with macrophage polarization[16], these data suggest that 1 μg/mL of ES antigen stimulation promotes TNF-α, IL-1β, IL-6, and IL-10, as well as promotes MHC class II and CD86 expression, whereas there are low levels of IL-12 production. These findings indicate that macrophages treated with 1 μg/mL ES display an M2b-polarized phenotype in vitro.
Previous reports indicate that MAPK-NF-κB signaling contributes to macrophage activation[23, 24]. To determine the expression pattern of MAPKs and NF-κB on activated macrophages, a Western blot was performed to analyze the levels of phospho-p38, phospho-p65, phospho-ERK, and phospho-JNK. As shown in Fig 2A, BMDMs treated with ES exhibited increased levels of phospho-p38, phospho-p65, phospho-ERK, and phospho-JNK expression in activated macrophages in a dose-dependent manner in vitro. As shown in Fig 2B, the RT-qPCR analysis revealed that BMDMs stimulated with ES antigens exhibited significantly decreased levels of IL-6, TNF-α, IL-10, MCP-1, and iNOS mRNA expression compared with the control BMDMs upon treatment with PD 98059 (ERK1/2 MAPK inhibitor), SP 600125 (JNK MAPK inhibitor), or BAY 11–7082 (NF-κB inhibitor). However, ST 2825 treatment markedly increased the levels of IL-6, RELMa, and Ym1 mRNA but decreased the levels of TNF-α, MCP-1, and Arg-1 mRNA compared with control group. Furthermore, SB 203580 (p38 MAPK inhibitor) treatment significantly decreased the levels of IL-10, MCP-1, and Arg-1 mRNA but increased the levels of RELMa mRNA compared with control BMDMs. Similar to these results, an ELISA was performed to assess the production of inflammatory markers in the supernatants of BMDMs exhibited remarkably decreased production of IL-6, MCP-1, TNF-α, and IL-10 upon treatment with PD 98059 (ERK1/2 MAPK inhibitor), SP 600125 (JNK MAPK inhibitor), or BAY 11–7082 (NF-κB inhibitor) (Fig 2C). However, BMDMs treated with ST 2825 exhibited significantly decreased levels of IL-10 and MCP-1. Furthermore, BMDMs treatment with SB 203580 (p38 MAPK inhibitor) displayed increased levels of IL-6, MCP-1, and TNF-α expression but remarkably decreased levels of IL-10 expression. These data demonstrate that the MyD88/MAPK/NF-κB signaling pathway facilitates ES-induced M2b polarization.
Multiple previous studies have reported TLR2 to be an important PPR for soluble egg antigens (SEA) [2,25]. To further evaluate the role of the TLR2 receptor on macrophage M2b polarization induced by ES, BMDMs derived from wild-type and TLR2 knockout (KO) mice were stimulated with ES in vitro. As shown in Fig 3A, compared with the BMDMs from wild-type mice, BMDMs derived from TLR2 KO mice stimulated with ES did not exhibit increased protein levels of phospho-p38, phospho-p65, phospho-ERK, and phospho-JNK in a dose-dependent manner. However, BMDMs derived from TLR2 KO mice stimulated with a high dose of ES could increase the levels of phospho-ERK protein expression. Treatment BMDMs derived from TLR2 KO mice with ES failed to induce Arg-1 production (Fig 3B). Compared with that of BMDMs from wild-type mice, the ELISA analysis for inflammatory marker production in the supernatants of BMDMs from TLR2 KO mice showed that the levels of IL-6, MCP-1, TNF-α, and IL-10 markedly decreased (Fig 3C). More importantly, RT-qPCR analysis for inflammatory gene expression showed that liver macrophages purified from wild-type mice stimulated with 1 μg/mL of ES exhibited enhanced IL-10, TNF-a, IL-1β, IL-6, IL-12, MCP-1, Arg-1, and iNOS but reduced the RELMa and Ym-1 mRNA levels compared with liver macrophages purified from TLR2 KO mice (Fig 3D). This indicates that similar to BMDMs (S1 Fig), liver macrophages can be activated by ES antigens in a TLR2-dependent manner. Our results show that TLR2 is the pivotal receptor for M2b polarization following ES antigen stimulation.
Previous studies have reported schistosome-specific lysophosphatidylcholine (lyso-PS) in SEA activated TLR2[6, 26]. To clarify whether the lipids in ES contribute to the production of pro-inflammatory cytokines by macrophages, RT-qPCR was used to assess the levels of IL-6, TNF-α, IL-10, MCP-1, and iNOS mRNA, which were found to be decreased in BMDMs stimulated with both proteinase K-treated ES and phospholipase C-treated ES, compared with mock-treated ES (Fig 4A). However, phospholipase C-treated ES treatment increased the levels of Arg-1 mRNA but decreased Ym1 mRNA in BMDMs compared with mock-treated ES. Furthermore, proteinase K-treated ES enhanced the levels of RELMa and Ym1 mRNA but reduced Arg-1 mRNA compared with mock-treated ES. Similar to these data, an ELISA was used to further assess the levels of TNF-α, IL-6, IL-10, and MCP-1 in the supernatants of BMDMs. As shown in Fig 4B, compared to the mock-treated ES, the proteinase K-treated ES as well as phospholipase C-treated ES failed to induce BMDMs to produce high levels of TNF-α, IL-6, IL-10, and MCP-1. Thus, lipids or lipid conjugates contribute to M2b polarization.
Although previous studies on schistosomiasis often focus on T and B lymphocytes, APCs (e.g., macrophages) may play vital roles in the pathogenesis of the disease [27]. To evaluate the role of macrophages during the pathogenesis of liver granuloma formation and fibrosis in a murine model of schistosomiasis, liver leukocytes were isolated and analyzed for the presence of CD11b+F4/80+ cells. The absolute number and percentage of eosinophils (SSChighCD11b+F4/80+ cells) and infiltrated macrophages (SSClowCD11b+F4/80+ cells) were substantially increased in the liver issues of the infected mice (Fig 5A). Four weeks post-infection, a real-time PCR analysis of inflammatory gene expression revealed that purified macrophages from infected WT mice exhibited enhanced IL-10, Arg-1, MCP-1, RELMa, and IL-6 but reduced TNF-α, IL-12, and iNOS mRNA levels, which exhibited an M2 dominant- polarized phenotype (Fig 5B). However, compared with WT mice, the purified macrophages from infected TLR2-/- mice exhibited higher levels of TNF-α, IL-12, and iNOS mRNA, as well as lower levels of IL-10, Arg-1, and RELMa mRNA, which represents a dominant M1-polarized phenotype (Fig 5B). Moreover, a significant decrease in the percentage of infiltrated macrophages and neutrophils (CD11b+Gr-1+ cells) recruitment was observed in infected TLR2-/- mice, compared with that of wild-type mice at 6 weeks post-infection (Fig 5A and Fig 5C). More importantly, the area of granuloma formation and fibrosis surrounding single eggs in the livers of TLR2-/- mice were significantly lower compared with that of WT mice at 6 weeks post-infection (Fig 5D).
Due to their inherent plasticity and heterogeneity, the ability of macrophages to functionally switch from killing pathogens to the promotion of tissue repair is likely critical for the host, especially when the host cannot eradicate a persistent infection but must limit tissue damage (e.g., chronic helminth infection)[17]. The differential activation status of macrophages has the capability to promote, restrict, or resolve inflammation and fibrosis in schistosomiasis [17, 27, 28]. In the present study, we aimed to investigate the phenotypical and functional plasticity of various macrophage subtypes following infection with S. japonicum (e.g., the presence of ES antigens derived from S. japonicum eggs). BMDMs stimulated with ES exhibit an activation status characterized by the production of multiple pro-inflammatory cytokines and abundant anti-inflammatory IL-10 production in vitro, following the activation of several MAPKs downstream of both TLR2 and MyD88.
Several helminth antigens are known to be associated with the stimulation of various TLRs[29, 30], whereas schistosoma antigens have been reported to interact with specific TLRs present on mononuclear phagocytes[25, 31–33]. However, few of these studies describe the signaling pathways triggered in host cells or link them to the production of specific cytokines[25, 33, 34]. In the current study, we demonstrated that ES-stimulated BMDMs produce pro-inflammatory cytokines and the anti-inflammatory cytokine IL-10 in a dose-dependent manner via TLR2. Moreover, we found that lipids or lipid conjugates participate in macrophage polarization, since the levels of cytokines decreased significantly in the supernatants of BMDMs induced by phospholipase C-treated ES. More importantly, our data confirmed that BMDMs exposed to the ES of eggs were dependent on the phosphorylation of three MAPK cascades (p38, ERK1/2, and JNK1/2), an event often reported downstream of TLR activation[35]. Notably, these kinase cascades occur with identical activation profiles, leading to the phosphorylation of p65 reported to be downstream of p38, ERK1/2, and JNK1/2. However, the full details of these signaling pathways has not previously been reported in the context of ES antigens. Moreover, in the present study, we provide the complete molecular mechanism of specific cytokine production by macrophages in response to ES antigens.
M2 macrophages can express Arg1, an enzyme adapted from the urea cycle which converts L-arginine to L-ornithine, enabling L-ornithine amino-transferase (OAT) to supply proline for collagen synthesis[36]. Although alternative macrophage activation is induced by IL-4/13 production, Arg1-expressing macrophages can downregulate inflammation, suppress Th2 cytokine production and reduce tissue scarring and pathology[37]. In addition, the anti-inflammatory cytokine, IL-10, plays a central regulatory role in the pathogenesis of schistosomiasis, and the maintenance of IL-10 expression during acute and chronic schistosome infection is critical for host survival[38]. In this study, TLR2 was found to be the key PRR associated with the production of pro-inflammatory cytokines, IL-10, and Arg-1 expression from BMDMs stimulated with ES in vitro. Surprisingly, we observed a significantly higher mRNA levels of IL-6, MCP-1 (CCL2), IL-10, and Arg-1 in the liver macrophages of infected wild-type mice compared to that of TLR2-/- mice, which was similar to M2b polarization in vitro. This finding suggests that ES antigens could induce the polarization of liver macrophages in vivo. However, the levels of IL-6, MCP-1 (CCL2), IL-10, and Arg-1 mRNA were found to be downregulated at 6 weeks post-infection. Thus, since similar serum levels of IL-4 and IL-13 were enhanced in both WT and TLR2 KO mice (S2 Fig), type II immunity may help to regulate levels of inflammatory cytokine mRNA in this study. Furthermore, the mRNA levels of RELMa were decreased in liver macrophages from TLR2 KO mice compared with that in WT mice at 4 weeks post-infection, which seemed different from that on stimulation with ES in vitro. Higher serum levels of IFN-γ that promoted M1 polarization and inhibited M2 polarization were observed in TLR2 KO mice compared with WT mice at 4 weeks post-infection (S2 Fig), thus type I immunity may help to downregulate the expression of RELMa. More importantly, IL-6 can recruit both neutrophils and monocytes during inflammation [39]. Moreover, MCP-1 accelerates liver fibrosis by promoting Ly-6C+ macrophage infiltration [40]. Our data indicate that TLR2 can increase the population of M2-type macrophages and neutrophil infiltration in the liver, perhaps because higher levels of MCP-1 (CCL2) and IL-6 are produced by ES-stimulated macrophages. Taken together, these data help explain the increased granuloma size and greater collagen deposition in the liver of wild-type mice. However, the mechanisms by which ES antigen-stimulated macrophages promote liver pathology require further investigation.
In summary, we have presented a detailed study of the molecular events that occur in BMDMs following exposure to ES antigens derived from eggs, which led to the production of pro-inflammatory cytokines and IL-10. This mechanism involves the activation of p65 downstream of TLR2 via the phosphorylation of p38, ERK1/2, and JNK1/2. Additionally, we showed that this mechanism is also responsible for the immunoregulatory activity in schistosomiasis, which provides an early in vitro demonstration of the role of ES Ags in modulating immunopathology downstream of TLR2. Finally, we observed that the production of IL-6, IL-10, MCP-1, and Arg-1 by macrophages in response to ES extends beyond an in vitro phenomenon and is also evident in the liver macrophages of WT-infected mice, which rapidly produce high levels of IL-6, IL-10, MCP-1, and Arg-1 mRNA in vivo following egg deposition. This early and rapid release of IL-6, IL-10, MCP-1, and Arg-1 has the potential to greatly modulate the immune response to limit inflammation and tissue damage in the liver by conditioning the microenvironment. More importantly, M2 polarization dependent on enhanced TLR2 signaling would reduce the level of liver damage and promote fibrosis. While our data suggest that targeted TLR2 signaling inhibitors may have therapeutic effect during the acute phases of schistosomiasis, further study is still required to address the role of TLR2 signaling to better understand the potential benefits of its inhibitors in treatment of chronic disease and delineate novel insight into the immune interplay underlying schistosomiasis. For this reason, the results generated from this study might provide evidence supporting the necessity to include TLR2 signaling as a novel therapeutic target for schistosomiasis.
|
10.1371/journal.pntd.0000749 | Health Education through Analogies: Preparation of a Community for Clinical Trials of a Vaccine against Hookworm in an Endemic Area of Brazil | Obtaining informed consent for clinical trials is especially challenging when working in rural, resource-limited areas, where there are often high levels of illiteracy and lack of experience with clinical research. Such an area, a remote field site in the northeastern part of the state of Minas Gerais, Brazil, is currently being prepared for clinical trials of experimental hookworm vaccines. This study was conducted to assess whether special educational tools can be developed to increase the knowledge and comprehension of potential clinical trial participants and thereby enable them to make truly informed decisions to participate in such research.
An informational video was produced to explain the work of the research team and the first planned hookworm vaccine trial, using a pedagogical method based on analogies. Seventy-two adults living in a rural community of Minas Gerais were administered a structured questionnaire that assessed their knowledge of hookworm, of research and of the planned hookworm vaccine trial, as well as their attitudes and perceptions about the researchers and participation in future vaccine trials. The questionnaire was administered before being shown the educational video and two months after and the results compared. After viewing the video, significant improvements in knowledge related to hookworm infection and its health impact were observed: using a composite score combining related questions for which correct answers were assigned a value of 1 and incorrect answers a value of 0, participants had a mean score of 0.76 post-video compared to 0.68 pre-video (p = 0.0001). Similar improvements were seen in understanding the purpose of vaccination and the possible adverse effects of an experimental vaccine. Although 100% of participants expressed a positive opinion of the researchers even before viewing the film and over 90% said that they would participate in a hookworm vaccine trial, an increase in the number who expressed fear of being vaccinated with a novel vaccine was seen after viewing the video (51.4% post-video versus 29.2% pre-video). Increases were also seen in the proportion who thought that participation in a vaccine trial would be inconvenient or disrupt their daily activities.
Even in rural, resource-limited populations, educational tools can be specially designed that significantly improve understanding and therefore the likelihood of obtaining truly informed consent for participation in clinical research. The observed changes in the knowledge and perceptions of the research participants about hookworm infection and the experimental hookworm vaccine demonstrate that the video intervention was successful in increasing understanding and that the subjects acquired knowledge pertinent to the planned research.
| Conducting clinical trials of new vaccines in rural, resource-limited areas can be challenging since the people living in these areas often have high levels of illiteracy, little experience with clinical research, and limited access to routine health care. Especially difficult is obtaining informed consent for participation in this type of research and ensuring that potential participants adequately understand the potential risks and benefits of participation. The researchers have been preparing a remote field site in the northeastern part of the state of Minas Gerais, Brazil, for clinical trials of experimental hookworm vaccines. A special educational video was designed based on the method of analogies to introduce new scientific concepts related to the researchers' work and to improve knowledge of hookworm, a disease that is highly prevalent in their community. A questionnaire was administered both before and after the video was shown to a group of adults at the field site, which demonstrated the effectiveness of the video in disseminating knowledge about hookworm infection and about the vaccine being developed. Therefore, even in a rural, resource-limited area, educational tools can be specially designed that significantly improve understanding and therefore the likelihood of obtaining truly informed consent for participation in clinical research.
| In research involving human subjects, the ethical relationship that must be established and maintained between investigators and research subjects is essential to successfully conduct investigational clinical trials of experimental drugs or vaccines, especially ones in which the participants are drawn from vulnerable populations [1], [2]. In such studies, investigators must attempt to mitigate any possible manipulation of the research population during the recruitment process, and especially to ensure that the risks and benefits to which volunteers are going to be exposed are well understood [3], [4]. The informed consent document is the traditional instrument utilized for this aim; by signing it, it is assumed that the volunteer has freely exercised their will, has formed their own evaluation and critique of the proposed research and has arrived at a truly informed decision about participation [5]–[8].
The process of obtaining informed consent becomes especially challenging when working in rural, resource-limited settings. Although there is no consensual definition of vulnerability, age, socioeconomic status, access to basic services such as health and sanitation, ethnic group, religion, cultural affiliation, and educational level are all characteristics that have been cited as indicating vulnerability, and which may therefore influence an individual's ability to consent to participation in clinical trials both in terms of the ability to exercise autonomy, but also to comprehend the proposed research [3], [9], [10]. With respect to the latter challenge, the informed consent form must frequently convey complex technical information and scientific concepts that are often not easily understandable, especially in populations with low levels of literacy [11]–[13]. To improve the quality of the information transmitted to potential study participants, researchers have, in general, increased the amount of information in informed consent documents [11]. However, a document that contains extensive and complex information may not convey a satisfactory understanding of the study procedures, or of the potential risks and benefits of participation in the research [14]. As an example, after evaluating the understanding of the information in an informed consent document for a research project in San Francisco, California, researchers found that despite using a form that had been simplified using language appropriate for a primary school level, the majority of individuals required more than one explanation of the study before satisfactorily comprehending it [12]. In that study, low literacy level and socioeconomic status were associated with an increased need for interventions that gave an improved comprehension of the information contained in the document.
In addition to the issue of comprehension, recruiting volunteers into clinical trials being conducted in resource-limited settings is further complicated by the limited access to medical care that is frequently found in such areas. Frequently, potential research subjects may feel an obligation to participate in order to receive medical attention for themselves or their family members [15]. Such motivation could influence individuals to participate and undertake risks that they otherwise would not accept.
Despite these very real issues related to obtaining informed consent from vulnerable populations living in resource-limited settings, it is often necessary to conduct clinical trials in such areas, particularly when the product being developed is meant to treat or prevent a disease that affects such a population. As one example, hookworm infection is one of the most prevalent chronic infections of humans, with an estimated 740 million cases worldwide, mostly in rural poor rural areas of the tropics and subtropics [16]. The two hookworms that infect humans are Necator americanus and Ancylostoma duodenale, with infection being transmitted through skin contact with soil contaminated with infective larvae. The major clinical manifestations result primarily from the loss of blood caused by adult worms that attach onto the intestinal wall, resulting in anemia which subsequently can lead to delays in cognitive development in children and reduction of well-being and productivity in adults [17].
Although effective chemotherapy exists to treat hookworm, current anthelminthics have important limitations, not least of which is that re-infection often occurs within a short time after treatment due to ongoing exposure [18], [19]. To develop an alternative control tool, the Human Hookworm Vaccine Initiative (HHVI) is developing a vaccine to prevent the morbidity due to this parasitic infection. Since hookworm does not occur in the developed world, testing the safety and efficacy of vaccines targeting this parasite must be conducted in the rural, resource-limited areas where the disease is endemic, among populations which are frequently referred to as being “vulnerable.”
The HHVI has been preparing a trial site for studies of its investigational hookworm vaccines that is based in the town of Americaninhas, in the northeast part of the Brazilian state of Minas Gerais [20], [21]. The first experimental vaccine being developed by the HHVI – the Na-ASP-2 Hookworm Vaccine – was tested there in a phase 1 clinical trial in 2007. For this study, healthy adult volunteers were recruited from communities surrounding Americaninhas.
In advance of this trial, studies were performed to assess the baseline knowledge of potential study participants in order to design appropriate educational interventions that could be used in the consenting process. Unfortunately, little is known about which are the more appropriate pedagogic models for informing populations involved in clinical trials, especially vulnerable populations that are economically or educationally disadvantaged. In the field of health education, purely informative pedagogic models do not usually result in modification of positions or attitudes, since behaviors are manifestations of firmly-held values and beliefs [22]. The ideal nature of an educational intervention that leads to acquisition of the knowledge necessary to make a conscious decision about participating in a research study is a matter of debate. Several authors have proposed a pedagogic model that makes use of the analogy [11], [23]. Analogies are useful tools for forming mental constructs that simplify or render familiar what an individual is attempting to understand [24]. The use of analogies can introduce new scientific concepts or alter previously held ideas [25], and can overcome barriers to learning by allowing an individual to make creative connections between pre-existing concepts and those related to the new knowledge being presented [23]. With this in mind, the current study aimed to develop an educational intervention based on analogies for a population resident in a hookworm endemic area of Brazil, and to evaluate its effectiveness in disseminating knowledge about the disease caused by hookworm, the experimental Na-ASP-2 vaccine that was about to be tested in their community, and about attitudes related to their willingness to participate in clinical trials of the vaccine.
This research was conducted as part of a larger epidemiological study whose purpose was to establish the prevalence of hookworm infection in various communities in the study area, in advance of planned hookworm vaccine trials.
The study was conducted in 2007 in the communities of Furtado, Beija-Flor and Jamir, all of which are rural areas endemic for hookworm located in the region surrounding the town of Americaninhas in the municipality of Novo Oriente de Minas, 500 kilometers northeast of Belo Horizonte, the capital of the Brazilian state of Minas Gerais. Americaninhas is located in a mountainous region with a tropical climate [20], [26]. The population largely exists on subsistence farming of cassava, sugar cane, coffee and beans. They typically live in simple, hand-made dwellings of packed earth or adobe with roofs of corrugated iron.
The population of Americaninhas consists of around 1000 people living in the urban center, with another 4000 living in the surrounding rural areas in smaller hamlets. Approximately 500 people live in the communities of Furtado, Beija-Flor and Jamir. This region was chosen for conducting clinical trials of experimental hookworm vaccines in view of the elevated prevalences of helminth infections that have been found during previous studies performed by the research team in the area: 68% for the hookworm N. americanus, 45% for Schistosoma mansoni, and 49% for Ascaris lumbricoides [20]. Such high prevalences are the result of socioeconomic and environmental conditions that favor the transmission and development of hookworm, such as a warm and humid climate, a lack of basic sanitation, and the low socioeconomic status of the population.
Individuals were included in the study if they were between the ages of 16 and 50 years and had been resident in the study area for the previous 24 months; were infected with hookworm as determined during the course of the larger epidemiological study; and, had completed a course of treatment for hookworm with albendazole. All volunteers consented to participate in the research, as evidenced by their signature on the informed consent form approved by the ethical review committees of the Centro de Pesquisas Rene Rachou (part of the Fundação Oswaldo Cruz – FIOCRUZ) and the George Washington University Medical Center, and completion of a true/false questionnaire that assessed their understanding of the informed consent document. Volunteers had to respond correctly to all questionnaire questions prior to being considered for inclusion into the study. In the cases of volunteers unable to read, the consent form was read aloud to them in the presence of an independent witness who also signed the form after the volunteer affixed their thumbprint.
The intervention was an educational video containing details about the proposed Na-ASP-2 vaccine trial, as well as explanations regarding the nature of research, the work of researchers, the reasons the area was chosen to test a vaccine against hookworm, and concepts of hookworm disease, vaccination, and the use of placebos in vaccine trials (Videos S1, S2, S3, S4, S5, S6, S7, S8, and S9). The video was filmed in the communities of Jamir and Beija Flor, and was produced using a pedagogical approach based on the use of analogies. In it, the daily experiences of local inhabitants, such as the cultivation of cassava, and the making of flour, sweets and cheese, are compared to the manufacturing of vaccines and to the experiments of researchers working in the laboratory. The characters featured in the film are actual inhabitants of the community who are representative of those from the rural interior of the state of Minas Gerais, thus facilitating identification of the viewer with the individuals on screen and enabling the learning process.
The film opens with scenes showing typical day in the town of Americaninhas, with people enjoying themselves in the town square or observing the main street from their windows (Video S1). After these initial images, the film transitions to describing the production of a traditional regional sweet. Each step in its production is shown, starting with cultivation of the sugar cane from which the basic ingredient is derived, followed by extraction of cane juice using a machine, preparation of the other ingredients, and combining these in specific quantities to create a final, high-quality product. Interspersed with these images are those of FIOCRUZ researchers working in the laboratory, using machines and instruments to assist them in discovering ideal components that, when combined in the correct amounts, may produce an effective vaccine. An analogy is thereby constructed between typical experiences of the region such as the production of sweets and the manufacturing of a vaccine against hookworm.
The video then shows people being interviewed about their knowledge of hookworm using language that is unique to the local population (Videos S2, S3, and S4). The individuals discuss the illness known locally as “amarelão” or the “illness of Jeca-tatu” (after a popular cartoon character from the early 20th century), its mode of transmission, and the associated symptoms. Their perceptions of the researchers working in the area and what they're doing in the region, as well as the hookworm vaccine program and the possible adverse effects of such a vaccine, are also presented.
In the final part of the video, members of the HHVI team speak on camera about the hookworm vaccine project, to clarify details of the planned clinical trial (Videos S5, S6, S7, S8, and S9). The presenters explain why hookworm is endemic in their community, the criteria for inclusion in the forthcoming vaccine trial, the adverse effects that the vaccine might cause, as well as how the vaccine is made.
The video was shown to prospective vaccine trial participants in group sessions in their own communities. After presentation of the film, conversation was encouraged to discuss and debate what was viewed so that the presented knowledge could be consolidated and learned.
Data were collected by means of a structured survey consisting of 45 questions, which was designed to assess knowledge about hookworm, vaccines (in general and the Na-ASP-2 vaccine in particular), and the researchers, as well as attitudes related to their willingness to participate in clinical trials of hookworm vaccines (Text S1). The questionnaire consisted of a combination of true or false questions, multiple choice questions, and subjective questions answered according to the 5-point Likert scale (ranging from “strongly disagree” to “strongly agree”). Survey questions were divided into three categories: a) those assessing knowledge about hookworm (Group 1); b) those assessing knowledge about the hookworm vaccine and upcoming clinical trial (Group 2); and, c) those assessing the attitudes and feelings of individuals about illness due to hookworm, vaccines, and participation in vaccine trials (Group 3).
The survey was first pilot tested on a group of 20 adults. After conducting this pilot test, modifications were made to improve the understanding of the tool by the general public. The survey was administered by specially trained interviewers at two distinct times: once immediately before viewing the educational video described above and then approximately two months later.
Data were tabulated and analyzed using the SPSS software program (version 15). Of the 45 questions on the questionnaire, only questions 2 through 40 were included in the analysis (the first question was for informational purposes whereas the final three concerned only those participants who had children). For questions from Groups 1 and 2, the frequencies of each answer were summarized. Responses to these questions were dichotomized such that each correct answer was assigned a value of “1” and each incorrect answer a value of “0”; a response of “don't know” was also assigned a value of “0” as it was considered a lack of knowledge. Subsequently, the mean of the responses was calculated for both the pre- and post-film administrations of the questionnaire. Additionally, composite scores for knowledge about hookworm and knowledge about the experimental hookworm vaccine were obtained by calculating the mean of responses to questions that fell into these broad categories. For the knowledge about hookworm composite score, responses to 14 similar questions were combined (#2–14, #26) whereas for knowledge about the hookworm vaccine 6 questions (#15, #16, #18–20, #30) were combined.
Questions in Group 3 were divided into two subgroups which were analyzed using different methods: for the first subgroup, answers were dichotomized such that an affirmative answer was assigned a value of “1” and a negative answer a value of “0” whereas for the second subgroup, the Likert scale consisting of five categories was used: −2 (strongly disagree), −1 (disagree), 0 (neither agree nor disagree), 1 (agree), and 2 (strongly agree). As for the dichotomous responses, means were calculated for each Likert scale question and compared pre- and post-film; to do this, the “strongly disagree” and “disagree” categories were combined and assigned a value of “0”, and the “strongly agree” and “agree” categories were assigned a value of “1”. As for the first two groups of questions, a composite score for the attitudes and feelings of study participants was created by combining individual responses to 7 different questions from the questionnaire (#21–23, #27–29, #31).
Student's paired t-test was used to compare pre-film and post-film means for both individual questions and the composite scores for Groups 1 and 2. The chi-square test was used to compare proportions for questions with several possible responses. For all tests, a p value less than 0.05 was considered significant.
Individuals were randomly selected from a list of participants who had been enrolled in a large epidemiological study and were invited to participate in the educational intervention. The initial sample consisted of 127 subjects; however, 55 of these did not undergo the second survey administration after viewing the educational video. The final study sample therefore consisted of 72 volunteers who were included in the data analysis. The average age of these participants was 30.1 years (range, 17 to 46 years), with 33 (45.8%) being female and 39 (54.2%) male.
When assessed as a composite score, knowledge about hookworm improved significantly after viewing the informational video from a mean score of 0.68 before the video to 0.76 after the video (p<0.0001), demonstrating that significant understanding was acquired by the participants through the targeted educational intervention. When assessing knowledge about vaccines and the proposed clinical trials of the experimental Na-ASP-2 Hookworm Vaccine, a small improvement was also seen after viewing the video, with the mean of correct answers on the post-video questionnaire being significantly higher than the mean on the pre-video test (0.58 vs. 0.65, p = 0.03) (Table 1).
Table 2 describes in more detail the specific knowledge acquired about hookworm after viewing the film. The illness, recognized by the popular name “amarelão” (“the big yellow”) by 88.9% of participants prior to viewing the film, was identified by the more scientific name “ancilostomídeo” by 91.7% of the study subjects after seeing the film (compared to only 79.2% before, p = 0.01). The mode of transmission of hookworm, which in Brazil is sometimes confused with other worm infections such as A. lumbricoides and S. mansoni, was already correctly identified as being through contact of skin with contaminated soil by 95.8% of subjects before the film, which increased to 100% after viewing the video (p = 0.08). However, the common misconception that hookworm infections are acquired through contact with contaminated water and unwashed fruit or vegetables, was abandoned by a quarter of the subjects participating in the research after viewing the educational film (88.9% pre-film vs. 63.9% post-film, p<0.001).
After the educational intervention, 79.2% of participants recognized that individuals can be infected with hookworm burt be asymptomatic, compared to 66.7% who held this view prior to seeing the film (p = 0.1). The acquisition of this knowledge may be crucial because health surveillance is intimately associated with the recognition of the gravity of the disease and the acknowledgement that it can be a “silent” illness. When asked about the negative health consequences associated with hookworm infection, even before viewing the educational film, 90.3% of participants correctly identified anemia as an important result of hookworm infection, compared to 93.1% after the film (p = 0.5).
Differences were found before and after the educational intervention when assessing the level of subjects' comprehension about the impact of illness due to hookworm in their community. Although even before viewing the video 84.7% of the population believed that hookworm was an important illness in their community, which increased slightly to 91.7% of participants afterward (p = 0.2), a significant change was seen in the more subtle question that asked whether participants thought that hookworm is not a serious illness because it can be easily treated: initially, 72.2% responded in the affirmative to this question whereas after viewing the educational film and hearing more about the long-term consequences of asymptomatic infection, only 45.8% held this viewpoint (p<0.001). Furthermore, after viewing the educational video more people understood that despite the existence of effective drug therapy for hookworm infection, there are major limitations with this treatment due to the probability of becoming re-infected, which in many cases can occur rapidly following treatment, although this increase in knowledge wasn't statistically significant (30.6% pre-video vs. 38.9% post-video, p = 0.2).
When evaluating the participants' knowledge about vaccines in general and the upcoming hookworm vaccine trial in particular, improvements in understanding were acquired after viewing the educational film (Table 3). Regarding the purpose of vaccination, the participants' knowledge before and after the educational intervention is shown in Figure 1. Results from the pre-film survey demonstrate that a majority of participants (56.9%) believed that the purpose of a vaccine is to treat an established illness. Although following the film, this association was still made by almost half of responders (45.8%), a significant increase did occur in those associating a vaccine with illness prevention (41.7% post-video vs. 20.8% pre-video, p = 0.005).
To assess knowledge about the possible adverse effects caused by vaccination, participants were asked what they thought could happen if an experimental vaccine were administered to them. The potential side effects of vaccination that they chose are presented in Figure 2. Among the responses, the possibility of experiencing an allergic reaction upon being vaccinated, which was recognized by none of the interviewed participants prior to viewing the educational video, was cited frequently during the post-film test (43.1%), as was the possibility of experiencing arm redness (at the site of injection) following vaccination (15.3% post-video compared to 1.4% before, p = 0.003). Of note, the educational intervention was associated with a significant reduction in the proportion of individuals who answered “other” (18.1% post-film vs. 54.2% pre-film, p<0.001) and a non-significant reduction in those who chose “don't know” as their response to this question (9.7% post-film vs. 18.1% pre-film, p = 0.15).
The participants' perceptions about the nature of the work being done by the FIOCRUZ research team in the study area are shown in Figure 3. The level of comprehension regarding the work of the researchers underwent significant change following the presentation of the interventional video. Although the concept that the work of a researcher is to take care of people's health and treat disease remained the perception of 30.6% and 33.3% of the interviewed subjects, respectively, compared to 36.1% and 31.9% who listed these roles prior to viewing the film (p = 0.5 and 0.9, respectively), there was a significant increase in the proportion of participants who listed the function of the researcher as consisting of conducting studies on a new vaccine (27.8% post-film vs. 4.2% pre-film, p<0.001).
Regarding survey questions pertaining to the attitudes and feelings of the participants towards the researchers and participation in a hookworm vaccine trial, there was no significant change in the composite mean of responses before compared to after viewing the film (0.86 vs. 0.83, p = 0.07) (Table 4). Similarly, no significant difference was found between the attitudes and feelings expressed pre-film and post-film when those questions evaluated using the 5-point Likert scale were combined as shown in Figure 4. However, for both of these measures, the attitudes of the study participants were somewhat less favorable towards participating in future vaccine trials after viewing the informational video, even if these differences weren't statistically significant.
Even before the educational intervention took place, more than 95% of the participants interviewed already expressed a favorable attitude toward the planned vaccine study, displayed confidence in the work of the researchers and expressed interest in learning more about hookworm and the experimental hookworm vaccine (Table 5). In fact, 100% of respondents agreed that the researchers are doing good work in their community, a percentage that remained unchanged after viewing the film. After the educational intervention, although no significant changes were recorded in the attitudes and feelings regarding aspects of the hookworm vaccine project, there were reductions in the number of individuals who were interested in participating in a future vaccine trial (95.8% before compared to 88.9% after) and in those who said that their family would approve of their participation in a vaccine trial (86.1% before compared to 80.1% after).
Closely related to these attitudes was the initial perception that participating in a vaccine study would result in focused attention on and treatment of health problems of study participants (95.8%), an aspect that certainly implies improvement of the health and the quality of life of each participating individual and indeed, even of others in the community (91.7%) (Table 5).
Regarding the attitudes and feelings expressed by those surveyed toward the researchers and the hookworm vaccine project, it is important to note that the proportion of people expressing the opinion that being a volunteer in a research study may complicate daily activities or prove inconvenient increased slightly after viewing the video: 15.3% and 23.6% held these views prior to watching the film compared to 23.6% and 27.8% after (p = 0.2 and 0.07, respectively). Furthermore, the percentage of participants who expressed fear of becoming ill or experiencing an adverse event also increased significantly, from 29.2% to 51.4%, respectively (p<0.01).
Our study has demonstrated that despite conducting research in a rural, resource-limited population that has limited access to routine health care and education, specially-designed educational activities can significantly improve understanding and therefore the likelihood of obtaining truly informed consent for clinical research. The observed changes in the knowledge and perceptions of the research participants about hookworm infection and the experimental hookworm vaccine clearly demonstrate that the targeted educational intervention was successful in increasing understanding and that the subjects acquired knowledge pertinent to the planned research. Importantly, the increase in knowledge appeared to be sustained, as the post-video questionnaire was conducted two months following the viewing of the film and not immediately afterward.
The conceptual evolution following an educational intervention can be attributed to the non-cognitive educational methodology that was utilized. The video that was developed and evaluated as part of this study relied on the use of analogies – for example, comparing the production of a new vaccine to the making of a local sweet – to convey new scientific concepts related to hookworm and research. The analogy is a comparison based on similarities between the structures of two different fields of knowledge [27]. Reasoning through analogy is, therefore, a subjective internal process that is effectuated by the interaction between two mental fields. The results of the questionnaire demonstrate that this pedagogical methodology is effective in the population that was studied.
The educational approach chosen for this study differs from conventional educational methods, in that it considers the lifestyle of people, their ideas, beliefs and values, and the specific cultural context. This leads to enhanced self-esteem, increased community participation, thus promoting the values of citizenship [28]. As outlined by Rice and Candeias, the traditional educational model in which information is simply provided to and individual or community has only temporary effects in terms of behavior change [29]. When the educational stimulus ceases, so too are its effects. The main criticism of the traditional cognitive approach is that it does not take into consideration the psychosocial and cultural determinants of health behaviors [30].
In a previous study that we conducted in a rural area of Minas Gerais where schistosomiasis is endemic, different health education approaches were assessed regarding their effectiveness in increasing the knowledge of schoolchildren with respect to the transmission and prevention of this parasitic disease [31]. This study demonstrated that in a group of schoolchildren whose education was based upon a model using analogies or social representations, levels of knowledge about schistosomiasis increased significantly, compared to those in which a cognitive model based on the simple presentation of information, or a control group that received no specific information about the disease.
Regarding the specific changes in knowledge that were observed in the current study after the educational intervention, an increased appreciation of the illness caused by hookworm was observed, such that after watching the film, it was identified as an important affliction in the study area due to its endemnicity and its effect on people's lives. The understanding that hookworm infection is not a major health problem because it can be easily treated with anthelminthic medications was reduced following the film, serving as evidence of an improved understanding of the re-infection process wherein individuals are continuously at risk of infection despite anthelminthic treatment due to the environment in which they live.
In analyzing the attitudes and perceptions of the participants, it appears that there were no significant changes in the subjects' notions in relation to the researcher and the benefits of participating in clinical trials. Capturing the participants' perspective before and after the educational intervention, no change was observed. Initially, an overall favorable opinion toward the researchers and the project was observed, with a full 100% agreeing that the investigators are doing good work in their community. This good-will and trust towards the researchers remained unchanged after viewing the educational film, and it translated into a high level of interest in participating in future vaccine trials, although this willingness decreased slightly after viewing the film, perhaps as a result of an increased understanding of the risks involved, as discussed below.
According to the vast majority of the individuals who were studied, participating in a vaccine trial might not only benefit themselves but also bring benefit to others who are not participants, resulting in betterment of the community as a whole. Surprisingly, although some participants acknowledged that participating in a vaccine trial might interfere with their daily activities, the majority of those interviewed said that it would not be inconvenient to them. The time commitment involved, which for these individuals whose livelihoods depend on long hours of hard manual labor, could be significant but would apparently be outweighed by the perceived potential benefits of participation such as improvement of their own health, greater attention to treatment of health problems, and potential identification of health problems that are otherwise unrelated to the study or vaccination with an experimental product.
A broad increase in understanding of the work being conducted by the research team after viewing the educational film was observed among those surveyed, although the idea that the project's purpose is to treat illness and take care of people's health – instead of to conduct research – remained prevalent despite having viewed the video. Regarding this specific point, it is important to highlight that understanding the research process is a challenge for investigators who are involved in the preparation of communities in advance of conducting clinical trials and beginning the individual informed consent process, especially in resource-limited communities with low levels of literacy and limited access to routine health care [32].
Since the Na-ASP-2 Hookworm Vaccine is an experimental vaccine, individuals participating in the planned study of this investigational product could potentially experience adverse reactions of variable degrees of severity, although in the first clinical trial of this vaccine administered to healthy, hookworm-naïve volunteers living in the United States, observed reactions consisted mostly of mild to moderate intensity injection site reactions such as pain, swelling and erythema [33]. After viewing the film in which the potential risks of participating in the proposed vaccine trial were described, an increase in apprehension related to participation was observed. This should not necessarily be seen as a negative outcome of the educational video, as it may in fact reflect a superior understanding of the risks involved when participating as a volunteer in a clinical vaccine trial – something that should be welcomed. Even though the educational intervention may have resulted in fewer people who would be willing to participate (88.9% after watching the video compared to 95.8% before), if those individuals who remained interested in volunteering were better informed, the intervention was successful.
In contrast to an increased appreciation of the risks of participation in a vaccine trial, no significant changes were observed after the educational intervention in relation to the perceived benefits that might come from participating in the research. Among the perceived benefits of participating in a vaccine trial were improvements in health, treatment of illness, and a better quality of life.
As demonstrated by our findings, the educational intervention utilized in this study was not uniformly successful. In several instances, erroneous perceptions of the study participants – such as the belief that the purpose of a vaccine is to treat disease, or that the role of the research is to take care of the health problems of the population – persisted in a significant proportion of those interviewed even after viewing the educational video. Obviously, although these misconceptions were reduced by the use of the educational video, further work remains to better inform potential research participants of the nature of research and the purpose of interventions that might be tested in future research trials. As with any ongoing research project, obtaining informed consent from volunteers is an on-going process in which individuals are continuously engaged in educational activities.
In summary, not only is it important to assess the current level of understanding of potential vaccine trial participants prior to conducting clinical studies, it is also useful to design specially-tailored educational interventions to develop a more informed community that is able and willing to participate in such research. This process should be continuous, with frequent re-assessment of the understanding of individuals that results in revision of the educational materials. It is the ethical imperative of the investigators and research team to ensure that potential study participants have an adequate understanding of the research to be undertaken. In resource-limited areas such as our study site, this often requires more than simply reading the informed consent document, and may include targeted educational activities similar to the one tested at our study site.
|
10.1371/journal.ppat.1007220 | The Sec1/Munc18 (SM) protein Vps45 is involved in iron uptake, mitochondrial function and virulence in the pathogenic fungus Cryptococcus neoformans | The battle for iron between invading microorganisms and mammalian hosts is a pivotal determinant of the outcome of infection. The pathogenic fungus, Cryptococcus neoformans, employs multiple mechanisms to compete for iron during cryptococcosis, a disease primarily of immunocompromised hosts. In this study, we examined the role of endocytic trafficking in iron uptake by characterizing a mutant defective in the Sec1/Munc18 (SM) protein Vps45. This protein is known to regulate the machinery for vesicle trafficking and fusion via interactions with SNARE proteins. As expected, a vps45 deletion mutant was impaired in endocytosis and showed sensitivity to trafficking inhibitors. The mutant also showed poor growth on iron-limited media and a defect in transporting the Cfo1 ferroxidase of the high-affinity iron uptake system from the plasma membrane to the vacuole. Remarkably, we made the novel observation that Vps45 also contributes to mitochondrial function in that a Vps45-Gfp fusion protein associated with mitotracker, and a vps45 mutant showed enhanced sensitivity to inhibitors of electron transport complexes as well as changes in mitochondrial membrane potential. Consistent with mitochondrial function, the vps45 mutant was impaired in calcium homeostasis. To assess the relevance of these defects for virulence, we examined cell surface properties of the vps45 mutant and found increased sensitivity to agents that challenge cell wall integrity and to antifungal drugs. A change in cell wall properties was consistent with our observation of altered capsule polysaccharide attachment, and with attenuated virulence in a mouse model of cryptococcosis. Overall, our studies reveal a novel role for Vps45-mediated trafficking for iron uptake, mitochondrial function and virulence.
| Cryptococcus neoformans is a causative agent of cryptococcal meningitis, a disease that is estimated to cause ~ 15% of AIDS-related deaths. In this context, cryptococosis is one of the most common causes of mortality in people with HIV/AIDS, closely behind tuberculosis. Unfortunately, very few antifungal drugs are available to treat this disease. However, understanding mechanisms involved in the pathogenesis of C. neoformans can lead to new therapeutic avenues. In this study, we discovered a new role for a regulatory protein involved in vesicle transport. Specifically, we found that the Vps45 protein, which regulates vesicle fusion, participates in the trafficking of iron into fungal cells, supports mitochondria function, mediates antifungal resistance and is required for virulence. These discoveries shed light on the molecular mechanisms underlying the uptake and use of iron as an essential nutrient for the virulence of C. neoformans. Further investigations could lead to the development of drugs that target Vps45-mediated processes.
| The pathogenic fungus Cryptococcus neoformans attacks immunocompromised people to cause cryptococcosis, a particularly devastating disease in HIV/AIDS sufferers [1]. Adaptations of the fungus to cause disease in mammalian hosts include the ability to grow at 37°C, to deliver key virulence components to the external milieu, and to acquire nutrients for proliferation [2,3]. In the latter case, iron plays a key role in the virulence of C. neoformans as a cofactor in essential biochemical reactions and as a regulator of the elaboration of the polysaccharide capsule, a major virulence factor [4,5]. As with other pathogens, C. neoformans must compete against host nutritional immunity to obtain iron during infection. Iron withholding by the host is achieved by iron-binding proteins such as transferrin, lactoferrin, and ferritin that maintain available iron at extremely low levels [6,7]. On the other hand, iron overload exacerbates cryptococcal disease in a mouse model of cryptococcosis [8]. Because of the iron-limited nature in the host, C. neoformans has developed multiple strategies to acquire iron including the use of a high-affinity iron uptake system composed of the cell surface iron permease Cft1 and the ferroxidase Cfo1 [9,10], the secreted mannoprotein Cig1 for iron uptake from heme [11] and the requirement of the endosomal sorting complex required for transport (ESCRT) pathway for endocytosis and intracellular trafficking of exogenous heme [12,13]. Furthermore, these systems are known to participate in the virulence of C. neoformans in a murine inhalation model of cryptococcosis [9–13].
The discovery that the ESCRT pathway is critical for iron acquisition from exogenous heme implies that intracellular transport occurs via endocytosis. Endocytic pathways internalize extracellular components, fluids and membrane-bound factors, including lipids and proteins, into cytoplasmic vesicles. These cargo-loaded vesicles fuse with the early endosome (EE) where sorting by ESCRT machinery and maturation occurs into late endosomes (LE) or multivesicular bodies (MVBs). Depending on its nature, cargo is either sent to the vacuole for degradation, to other organelles, or it may be recycled back to the trans-Golgi network (TGN) where re-secretion to the plasma membrane may occur [14–16]. The hypothesis that endocytic processes are involved in iron uptake in fungi was first inferred from studies in Saccharomyces cerevisiae and Candida albicans. For example, the ferrichrome siderophore receptor Arn1p in S. cerevisiae is found in endosome-like vesicles and is sorted directly from the Golgi to the endosomal compartment in the absence of its substrate. However, when cells are exposed to ferrichrome, Arn1p is relocalized to the plasma membrane where the siderophore-loaded receptor complex rapidly undergoes endocytosis [17]. In C. albicans, the plasma membrane proteins Rbt5 and Rbt51 are required for binding and iron uptake from heme and hemoglobin. A screen for mutants defective in iron utilisation from hemoglobin in S. cerevisiae expressing heterologously Rbt51 revealed mutants impaired in vacuolar functions, including vacuolar ATPase, and components of the ESCRT pathway and the HOPS complex. Subsequent analysis confirmed the role of ESCRT-I complex VPS23 and VPS28 in iron uptake from hemoglobin, but not from ferrichrome, suggesting that the processes of iron uptake from hemoglobin and ferrichrome are distinct [18]. Likewise, ESCRT mutants in C. neoformans show defects in iron acquisition from heme and hemoglobin, but not from ferrichrome [12,13]. In addition, there was an extended lag phase for ESCRT mutants when grown in media containing iron chloride as a sole iron source suggesting that ESCRT pathway plays a role in recycling the high-affinity iron uptake system Cft1-Cfo1 to the plasma membrane [9,10]. This idea is consistent with the findings in S. cerevisiae that levels of iron influence cellular localization of the iron permease/ferroxidase proteins Ftr1 and Fet3 (i.e., recycling back to the plasma membrane in low levels of iron or targeting to the vacuole for degradation in high levels of iron); these endocytic sorting processes also require the ESCRT pathway [19].
Vesicle fusion events in endocytic compartments are a dynamic process that is regulated by specific mediators such as tethering factors, SNAREs, Rab GTPases, guanine-nucleotide exchange factors (GEFs) and Sec1/Munc18 (SM) proteins. SNARE (soluble N-ethylmaleimide-sensitive attachment protein receptor) proteins form tight four-helix bundles (core complexes) that bring specific membranes together [20]. Prior to fusion, transport vesicles are loaded with tethering factors that promote the assembly of specific and functional SNARE complexes (also called SNAREpins) by interacting with SM proteins, Rab GTPases and GEFs [21]. SM proteins are evolutionarily conserved proteins that are considered to be key elements of the fusion machinery because they strongly promote SNARE-mediated fusion and specificity [22,23]. Four SM proteins were identified in S. cerevisiae and they function at distinct intracellular transport steps depending on their cognate partners. For example, the SM protein Sec1 functions at the plasma membrane during exocytosis while Sly1 operates at the endoplasmic reticulum (ER)-Golgi membrane [21]. The SM protein Vps33 is required for delivery of endocytosed cargo into the vacuole. Vps33 is a core member of the endosomal Class C core vacuole/endosome transport (CORVET) complex and the vacuolar homotypic fusion and vacuole protein sorting (HOPS) complex that play a role in the endosome-to-vacuole pathway [24–26]. The SM protein Vps45 interacts with and stabilizes the SNARE Tlg2 and acts on the TGN-early endosome pathway [27–30]. Vps45 also operates at the late endosome in collaboration with the SNARE Pep12p; this process has also been reported in mammalian cells and in Aspergillus nidulans [27,30–32]. In this study, we characterized the SM protein homologue Vps45 in C. neoformans and discovered roles for the protein in the uptake of exogenous iron, intracellular sorting of an iron uptake protein, mitochondrial function and virulence.
The C. neoformans gene locus CNAG_03628.2 encoding a candidate Vps45 protein was first identified by a reciprocal BLASTp search using the amino acid sequence of S. cerevisiae Vps45. The predicted polypeptide (686 aa) from C. neoformans displayed 36% identity and 56% similarity to the S. cerevisiae protein and 47% identity and 63% similarity to the corresponding protein from Aspergillus nidulans. A phylogenetic analysis for Vps45 sequences is presented in Fig 1. Considering the strong relatedness between orthologs of VPS45, we proposed to change the annotation of CNAG_03628.2 to Vps45. To examine the role of Vps45 in C. neoformans, we constructed two independent targeted deletion mutants and corresponding strains in which the mutation was complemented with the wild-type (WT) gene. The genotypes of the strains were confirmed by PCR and genome hybridization (S1 Fig).
As mentioned above, endomembrane trafficking is conducted by specific membrane fusion events that occur via the formation of SNARE bundles, with regulation by SM proteins. To determine whether Vps45 plays a role in endocytosis in C. neoformans, cells were grown in low iron conditions, transferred to high iron media, stained with the endocytic tracker dye FM4-64, incubated at 30°C and 37°C, and observed for 90 min. At 30°C, accumulation of the dye on endosomes and at the vacuolar membrane was detected in the WT and complemented strains within 15 min. However, this accumulation at the vacuole was delayed in vps45 mutants by 15 min. An increase of disorganized cytoplasmic membrane material was also observed over time in mutant cells as opposed to a more organized formation of endosomes observed in WT and complemented strains (Fig 2A). Similar results were observed when the assay was performed at 37°C, with a general delay in vacuolar staining. Endosomes were observed in WT and complemented strains only after 45 to 90 min while none were detected in mutant strains (S2A Fig). Moreover, endocytosis assays with FM4-64 were also performed using the VPS45-GFP complemented mutant grown at 30°C and 37°C (S2B Fig). At 30°C, Vps45-GFP fusion protein was first localized at the plasma membrane at 0 min and then to the vacuolar membrane within 15 min (Fig 2B). Vacuolar co-localization of Vps45-GFP was found after 30 min of incubation at 37°C (S2B Fig). Colocalization analysis revealed a positive correlation between Vps45-GFP and endocytic membranes as determined by Pearson’s R value (0.27–0.40) and Costes P-value (0.99–1.00) [33]. Notably, introduction of the VPS45-GFP gene fusion into the vps45 mutant complemented growth defects for a number of conditions at 30°C and 37°C (S3 Fig).
Intracellular acidification due to endosome trafficking is the hallmark of endocytosis [14]. Intracellular acidification was assessed by flow cytometry using the acidic pH sensor carboxy-DCFDA dye. WT, mutants and complemented cells did not display any fluorescence differences when grown in YNB at 30°C and 37°C. However, when grown in iron chelated-conditions, vps45 mutants display increased and reduced fluorescence at 30°C and 37°C, respectively, in comparison to WT and complemented strains (Fig 2C and 2D). This suggests that an increase in intracellular pH occurs under iron-depleted conditions at 30°C in vps45 mutants compared to WT and the complemented strains. However, an increase of intracellular pH was only observed in WT and complemented strains when grown in iron-chelated conditions at 37°C. These results indicate that endocytic processes mediated by Vps45 operate differently depending on the temperature, a finding that is also consistent with the reduce rate of endocytosis and endosome formation previously observed at 37°C. In addition, the mutants, WT and complemented strains were tested for their susceptibility to trafficking inhibitors (i.e., N-Ethylmaleimide (NEM), brefeldin A (BFA) and monensin) and the glycosylation inhibitor tunicamycin (TNC) since many secreted or membrane-bound proteins are N-glycosylated. Brefeldin A and monensin did not impair the growth of any of the strains, but the vps45 mutants showed increased sensitivity to the trafficking inhibitor NEM and the glycosylation inhibitor tunicamycin (Fig 2E). Taken together, these results are consistent with a role for Vps45 in endocytic events.
We have previously demonstrated that iron acquisition from exogenous heme requires the ESCRT machinery, and this observation implies that intracellular transport of heme occurs by endocytosis [12,13]. ESCRT complexes recognize ubiquitinated cargo proteins at endocytic vesicle membranes and are sequentially recruited to deliver endosomes containing extracellular material to the vacuole [34]. Therefore, we investigated the role of Vps45 in the ability of C. neoformans to use different iron sources. The WT strain, deletion mutants, and the complemented strains were examined for growth in iron-chelated medium (YNB-BPS) (Fig 3A). After 48h of iron starvation in a pre-culture at 30°C, all strains were subsequently unable to grow in YNB-BPS, but robust growth in YNB with iron was observed, indicating that iron starvation did not kill the cells (Fig 3A). Furthermore, all of the strains grew at 30°C in YNB-BPS supplemented with 10 μM and 100 μM FeCl3 (Fig 3A). However, when incubated at 37°C, the vps45 mutants exhibited a delay in growth on YNB-BPS with 25 μM or 100 μM FeCl3 (Fig 3B). This result indicates that Vps45 is required for the use of FeCl3 at the host temperature (37°C), but not at lower temperature. We speculate that another system may contribute to iron uptake at 30°C but may be insufficient at 37°C. Similar results were obtained with spot assays on solid media (Fig 3C). However, when incubated with organic iron sources (i.e. heme, hemoglobin, transferrin and sheep blood), the vps45 mutants exhibited a delay in growth at both 30°C and 37°C, as compared with WT and complemented strains (Fig 3A and 3B). Also, growth was inhibited for the vps45 mutants when incubated with hemoglobin and blood at 37°C. Again, similar results were obtained with spot assays on solid media (Fig 3C). These results suggest that Vps45 is required for the use of organic iron sources found in mammalian hosts. Interestingly, when intracellular iron content was measured by inductively coupled plasma mass spectrometry (ICP-MS) for cells grown in high iron conditions at 37°C, the vps45 mutants displayed a greater amount of iron compared to the WT and complemented strains (Fig 3D). We hypothesize that the ability to properly traffic the components of iron uptake systems may be required at 37°C. That is, loss of Vps45 may impede the delivery of iron to the vacuole (and sensing of iron repletion) to create an imbalance in iron homeostasis leading to enhanced iron uptake and the increase in intracellular iron content in vps45 mutants. Altogether, the results indicate that Vps45 is required for intracellular iron transport in a temperature-dependent manner.
In C. neoformans, iron from FeCl3 or transferrin is transported by the high-affinity iron uptake system composed of the iron permease Cft1 and the ferroxidase Cfo1 [9,10]. Indeed, deletion of either CFT1 or CFO1 results in severe growth defects with FeCl3 or transferrin as sole iron sources, but not with heme or hemoglobin, which suggests that there are distinct uptake pathways for these iron sources [9,10]. The Cft1-Cfo1 system is homologous to the iron permease/ferroxidase Ftr1-Fet3 in S. cerevisiae, which was demonstrated to be endocytosed in the presence of iron [19]. Given that the Cfo1-GFP fusion protein was found at the plasma membrane and that the vps45 mutants showed a growth defect in FeCl3, we next tested whether the trafficking of Cfo1-GFP was impaired in absence of VPS45. As shown in Fig 4 and S4 Fig, Cfo1-GFP started to accumulate at the plasma membrane within an hour of incubation at 30°C (Fig 4A) and 37°C (S4A Fig) when cells were transferred from a rich medium (YPD) to the iron-chelated medium (YNB-BPS). These results are consistent with our previous analysis of the location of the Cfo1-GFP protein in the background of a cfo1Δ mutant [9]. Accumulation of Cfo1-GFP at the plasma membrane was not affected by deletion of VPS45 suggesting that Vps45 does not play a role in the transport of de novo synthesized Cfo1 to the plasma membrane. Furthermore, Cfo1-GFP was located in vesicular bodies that resemble endosomes in the WT and complemented strains, but not in the vps45 mutant. After 2h of incubation, Cfo1-GFP started to accumulate inside the vacuole of the WT and complemented strains at 30°C (Fig 4A). Vacuolar localization was maintained after 24h of incubation in low iron media (Fig 4B). In contrast, Cfo1-GFP was retained at the plasma membrane and dispersed in the cytoplasm in the absence of Vps45, but was never located at the vacuole at the time (3h) and conditions (30°C and 37°C) tested (Fig 4 and S4A Fig).
Upon addition of FeCl3, the Cfo1-GFP signal was reduced in the WT and complemented strains, and was found mostly in vesicular bodies after 24h of incubation at 30°C (Fig 4B). The signal was also found in vesicular bodies in cells from the YNB medium at 30°C and, as mentioned above, the signal was mainly in the vacuole in cells from YNB-BPS (iron-chelated) medium. With incubation at 37°C in YNB-BPS+FeCl3, a strong signal was detected for Cfo1-GFP, and the fusion protein was found at the plasma membrane, in punctate structures and at the vacuole in the WT and complemented strains (S4B Fig). In these strains, the signal was also found in the vacuole and vesicular bodies in cells grown in YNB medium at 37°C, and the signal was mainly in the vacuole in cells grown in YNB-BPS medium. Again, Cfo1-GFP was located mainly at the plasma membrane in vps45 mutants when grown in YNB, YNB+BPS and YNB-BPS+FeCl3 at 37°C for 24h (S4B Fig). We established that full length Cfo1-GFP protein was present in the WT, mutant and complemented cells incubated for 3h in YNB-BPS at 30°C and 37°C, as visualized by Western blot (S4C and S4D Fig). Taken together, our findings suggest that the intracellular trafficking and expression levels of Cfo1 may be regulated by iron availability and temperature, and that Vps45 plays a role in intracellular transport of the protein from the plasma membrane to the vacuole.
A vps45 mutant in S. cerevisiae shows increased sensitivity to the antiarrhythmic drug aminodarone that mediates disruption of calcium homeostasis [35]. Additionally, mitochondria play key roles in metabolism including iron and calcium homeostasis, and interact with other organelles such as the ER to carry out their functions through communication via diffusion, vesicular transport and direct contact [36,37]. These findings, and the involvement of Vps45 in iron homeostasis at least partially through uptake and through endocytosis of Cfo1, prompted us to investigate the potential participation of Vps45 in mitochondrial function. We first determined whether Vps45 associates with mitochondria by examining the strain expressing Vps45-GFP after growth in YNB under iron-depleted conditions at 30°C and 37°C. Mitochondrial staining with mitotracker revealed that the Vps45-GFP protein was co-localized with a subset of mitochondria under every condition tested (Fig 5A and 5B). Colocalization analyses were performed on images of whole cells using the ImageJ coloc2 test. Positive correlations between Vps45-GFP and subcellular region were determined by assessing Pearson’s R value and Costes P-value [33]. Positive associations occurred in 47% of cells grown in YNB and 58% in YNB+BPS at 30°C, with mean R values of 0.274 ± 0.019 and 0.278 ± 0.018 respectively. This association was also observed at 37°C. Indeed, 58% of cells grown in YNB and 75% of cells grown in YNB-BPS at 37°C were found to show Vps45 in association with mitochondria with mean R values of 0.284 ± 0.020 and 0.273 ± 0.018, respectively (S5A Fig). Moreover, observations of mitochondria morphology did not reveal any differences between WT, mutants and complemented strains under iron-depleted conditions at 30°C and 37°C (S5B and S5C Fig).
In parallel, the WT, mutant and complemented strains were tested for their susceptibility to inhibitors of the mitochondrial electron transport chain (ETC) complexes. Growth was not impaired for any strains in the presence of inhibitors of ETC complexes I and II (i.e., rotenone, malonic acid, and oligomycin A) (S5D Fig). However, the vps45 mutants showed increased sensitivity to the alternative oxidase inhibitor salicylhydroxamic acid (SHAM), the complex III inhibitor antimycin A, and the complex IV inhibitor potassium cyanide (KCN), when compared to the WT and complemented strains (Fig 5C). Increased sensitivity to hydrogen peroxide (H2O2) was also observed for the vps45 mutants, but no growth defect was noted when the mutants were exposed to other agents that caused oxidative stress such as paraquat, plumbagin, diphenyleneiodonium chloride (dpi) and menadione (Fig 5C and S5D Fig).
Mitochondrial membrane potential (MMP) was also measured by flow cytometry in cells stained with JC-1. The JC-1 dye exhibits potential-dependent accumulation in mitochondria as indicated by a fluorescence emission shift from green (~ 529 nm) to red (~ 590 nm). Consequently, mitochondrial depolarization is indicated by a shift from red to green fluorescence (S5E Fig). The WT, mutant and complemented strains were grown in YNB under iron-depleted and replete conditions at 30°C and 37°C, and stained with JC-1 for 30 min. Mitochondria membrane potential was assessed for each strain by determining the % of the cells in the population with polarized (only red signal), depolarized (only green signal) and mixed polarization (red and green signal). When grown in YNB at both 30°C and 37°C, all cells displayed a mixture of polarized and depolarized mitochondria membrane potential. However, upon iron depletion at 30°C, a shift from mixed polarization to a fully depolarization population was observed and restoration of the mixed polarization population was noted under iron-replete conditions (Fig 5D). Conversely, under the iron-depleted condition at 37°C, vps45 mutants maintained a mixture of polarized and depolarized mitochondrial membrane potential while WT and complemented strains showed an increase in the depolarized population. Addition of iron restored mixed polarization in the WT and complemented strains, while a shift towards depolarization in vps45 mutants was detected (Fig 5E). These results suggest that iron homeostasis impacted mitochondria membrane potential in the vps45 mutants at 37°C. Since iron uptake, trafficking and sensing are impaired in the absence of Vps45, especially at high temperature where cells displayed a greater amount of intracellular iron, we speculate that iron accumulation in these mutants contributes to their mitochondria membrane depolarization. Paradoxically, chelation of iron by the addition of BPS may alleviate the potential dysfunction in iron homeostasis, and this may translate into maintenance of mitochondrial membrane potential.
Along with the ER, mitochondria also play important roles in calcium homeostasis and signaling [38]. Given the connections between Vps45 and mitochondrial function established above and known connections with calcium homeostasis in yeast [35], we next tested our strains for their susceptibility to calcium inhibitors. Growth was not impaired for the WT and complemented strains in the presence of aminodarone, cyclosporine A, tacrolimus (FK506), and the extracellular calcium chelator EGTA, but the vps45 mutants showed increased sensitivity to these compounds (Fig 6A). Moreover, chelation of calcium with EGTA resulted in growth inhibition for all strains in liquid media (Fig 6B). Addition of calcium restored the growth of the strains when incubated at 30°C, but the growth of the vps45 mutants remained hindered at 37°C (Fig 6B). This result suggests that Vps45 is required for calcium use at host temperature. Mitochondrial morphology was assessed in cells grown in calcium-chelated conditions at 30°C and 37°C (Fig 6C and 6D). A mixture of puncta and rod shaped mitochondria were observed when cells were grown in YNB at 30°C or 37°C, and no marked change in mitochondrial morphology was observed when cells were exposed to the calcium chelator EGTA at either temperature.
In light of our finding that Vps45 participates in mitochondrial functions, we next determined whether Vps45 associated with mitochondria under calcium-limited conditions. The strain expressing the VPS45-GFP construct was grown in YNB under calcium-depleted conditions at 30°C or 37°C. Mitochondrial staining with mitotracker showed that Vps45-GFP protein was found associated with a subset of mitochondria upon calcium restriction in 61% of cells grown at 30°C and 87% of cells grown at 37°C with mean R values of 0.347 ± 0.039 and 0.299 ± 0.031, respectively (Fig 6E and 6F and S5A Fig). We also measured the mitochondrial membrane potential (MMP) by flow cytometry in cells stained with JC1-1 (S5D Fig). When the WT and complemented strains were grown under calcium-starved conditions (YNB-EGTA) at 30°C, a shift from mixed polarized to depolarized mitochondria was observed. A similar shift was also obtained for vps45 mutants but to a lesser extent. Interestingly an increased in temperature restored mitochondrial membrane potential population to levels similar to those found when cells were grown in YNB at 37°C (Fig 6G). We went on to measure intracellular calcium content by inductively coupled plasma mass spectrometry (ICP-MS), and found that only the vps45 mutants displayed a greater amount of calcium compared to WT and complemented strains in high iron conditions at 37°C (Fig 6H). Taken together, these results suggest that Vps45 influences calcium homeostasis and mitochondrial membrane potential in a manner related to iron availability.
Our analysis of the role of Vps45 in iron and calcium homeostasis indicated that mutant phenotypes were generally exacerbated at elevated temperature. Given that the transcription factor Crz1 is an effector of calcineurin, a key mediator of calcium signaling that is activated by temperature stress [39], we further examined phenotypes related to calcineurin and Crz1 function. In particular, genes involved in cell wall remodelling are regulated by Crz1 in C. neoformans and we therefore investigated whether vps45 mutants were sensitive to agents that challenge cell wall integrity [39,40]. Specifically, we exposed our strains to calcofluor white, SDS, caffeine or NaCl, and found that the vps45 mutants displayed growth defects, especially at 37°C, when compared to the WT and complemented strains (Fig 7A). Furthermore, deletion of CRZ1 in a vps45 mutant abolished growth on calcofluor white and caffeine at 37°C (Fig 7B). In addition, the double knockout mutants displayed a growth defect on YPD at 37°C, and when grown on the calcium chelator EGTA or the drug aminodarone that disrupts calcium homeostasis. These results suggest a role for Vps45 in calcium signaling that may overlap with functions controlled by the calcineurin and Crz1 pathway. To further investigate cell wall remodelling in the absence of VPS45, we stained the cells with the fluorescent probe pontamine, a non-specific cell wall dye, and with lectin-binding probes, such as eosin Y and concanavalin A, which respectively bind chitosan and mannans present in the cell wall. By measuring differential fluorescence intensity by flow cytometry, we found that vps45 mutants showed increased fluorescence when stained with pontamine and concanavalin A, but not with eosin Y (Fig 7C). The observed differences were greater when the mutant cells were grown at 37°C and compared to the WT and complemented strains. These observations suggest that defects in cell wall structure result from the absence of VPS45.
Changes in cell wall integrity may also influence the exposure of surface molecules that are important for virulence, such as capsular polysaccharides. We therefore analyzed the impact of the loss of VPS45 on capsule elaboration (Fig 7D–7G). First, we assessed capsule size by India ink staining of cells cultured in defined limited-iron media for 48h at 30°C (D) or 37°C (F). We observed a significant change in capsule size after deletion of VPS45 and the change was temperature dependent. That is, capsule and cell size increased when cells of the vps45 mutants were grown at 30°C (Fig 7D). However, a reduction in capsule size only was observed at 37°C (Fig 7F). Glucuronoxylomannan (GXM) is the major polysaccharide of the capsule that is both attached to the cell wall and shed into culture medium [41,42]. Given that cell wall remodelling defects in the vps45 mutants might influence capsule attachment, we determined whether vps45 mutants shed different amounts of GXM versus WT cells by performing a capsule blot assay [41,43]. The cells were incubated in defined low-iron media for 48h at 30°C or 37°C and the relative abundance of shed polysaccharide was analyzed by reactivity with an anti-GXM antibody (mAb18B7). As shown in Fig 7E, the vps45 mutants shed very little GXM into the supernatant compared to WT and complemented strains when grown at 30°C. This result suggests that capsular material is fully attached to the cell wall rather than being secreted into the extracellular space, a result consistent with the finding that cells and capsule are larger in vps45 mutants when grown at 30°C. Conversely, capsular material was barely detected in the supernatants of cultures from vps45 mutants when incubated at 37°C (Fig 7G). Given the reduced capsule size of these cells, this result suggests that capsular material is synthesized at a reduced level and/or that a defect in secretion may occur under these conditions. However, further investigation revealed that mutant cells are unable to survive and proliferate under the conditions of minimal media and high temperature of the experiment, and this could explain the reduced capsular material on the cell surface and in the supernatant (S6 Fig). Overall, these observations revealed that deletion of VPS45 caused defects in the capsule formation and cell survival under austere conditions.
Chloroquine and quinacrine are antimalarial drugs that accumulate inside lysosomes (vacuoles) of the parasite Plasmodium and interfere with hemoglobin digestion. Briefly, ingested hemoglobin is deposited inside the digestive vacuole where degradation occurs and heme is transformed into an inert crystalline polymer called hemozoin. Chloroquine and derivatives accumulate in the vacuole and bind to hematin, a toxic product of hemoglobin degradation, preventing its incorporation into hemozoin, thus intoxicating the parasite [44]. On the other hand, azoles, including imidazoles (e.g. miconazole) and triazoles (e.g. fluconazole) interfere with ergosterol biosynthesis by inhibiting the enzyme lanosterol 14-α-demethylase. Because ergosterol is a major constituent of the fungal membranes, its depletion results in growth inhibition [45]. Since Vps45 was found to be associated with plasma and vacuolar membranes, and mutants displayed defects in cell wall integrity, we tested the sensitivity of vps45 mutants to cell wall, antimalarial, and azole drugs. As shown in Fig 8A, vps45 mutants display an increased sensitivity to caspofungin, quinacrine, chloroquine and fluconazole when compared to the WT and complemented strains, whereas all strains were sensitive to miconazole. This marked sensitivity is noticeable at both 30°C and 37°C. The drug sensitivity of the vps45 mutant could be explained by the fact that perturbation of membrane trafficking, fusion events and potentially cargo delivery impairs intrinsic resistance mechanisms. Addition of exogenous heme has been demonstrated to reverse sensitivity to fluconazole in a cfo1mutant [9]. Therefore, we tested if addition of iron and heme had an impact on drug susceptibility and we found that supplementation of iron but not hemin improved the growth of vps45 mutants on fluconazole when incubated at 30°C, but not at 37°C (Fig 8B). This result is in agreement with the previous results that Vps45 is required for the use of FeCl3 at the host temperature (37°C), but not at lower temperature. Furthermore, addition of iron and heme to miconazole aided the growth of WT and complemented strains, but not the vps45 mutants (Fig 8B). Overall, these results suggest that Vps45 is required for resistance to drugs that target the cell wall, membranes and the vacuole.
Given the contributions of Vps45 to capsule formation and the influence of host temperature on mutant phenotypes, we next tested the importance of VPS45 for the development of cryptococcosis. We first examined the survival and proliferation of fungal cells upon interaction with murine macrophage-like J774A.1 cells. As shown in Fig 9A, CFU numbers of vps45 mutants recovered at 24h were significantly lower than for the WT and complemented strains suggesting a reduced ability to survive and proliferate inside macrophages. This result prompted a further assessment of virulence in mice. Groups of 10 female BALB/c mice were inoculated intranasally with WT, a vps45 mutant or the corresponding complemented strain and monitored daily for disease development. The WT and complemented strains caused a lethal infection in all mice by days 18 and 23, respectively. However, mice inoculated with the vps45 mutant showed no sign of disease and survived until the end of the experiment (50 days) (Fig 9B). This avirulent phenotype was confirmed by examination of the fungal loads in organs harvested from the infected mice. As opposed to the mice infected with the WT or complemented strains, no cells of the vps45 mutant were retrieved from the bloodstream or any of the organs (i.e., lungs, brain, liver, spleen or kidney) (Fig 9C). Overall, these results indicate that VPS45 is required for the survival inside macrophages as well as virulence and dissemination to systemic organs in mice.
The SM protein Vps45 plays an important role in vesicle trafficking by conferring specificity on SNARE proteins that mediate docking and fusion events [27]. As initially demonstrated in S. cerevisiae, a vps45 mutant is defective in vacuolar biogenesis due to impaired fusion of Golgi-derived vesicles with the prevacuolar compartment [28,46]. In this study, we investigated the role of an ortholog of Vps45 in iron trafficking in C. neoformans, and found that the protein is needed for robust growth of the pathogen on both inorganic and organic iron sources. Notably, the iron-related phenotypes were more severe at 37°C. Interestingly, loss of Vps45 changed the intracellular distribution of a fusion protein of GFP with the ferroxidase Cfo1 that mediates high affinity iron uptake in C. neoformans. Specifically, the WT strain accumulated Cfo1-GFP at the plasma membrane and vacuole under low iron conditions while the protein was found in the PM and internal punctate structures (presumably endosomes) in a vps45 mutant. That is, loss of Vps45 prevented vacuolar accumulation. Under high iron conditions, the protein accumulated in internal punctate structures in the WT strain, and in the PM and internal dispersed structures in the mutant. These results suggest that proper trafficking of Cfo1, including localization at the vacuole, contributes to robust growth on FeCl3. The differences in localization and growth were most marked at 37°C, and the mutant hyper-accumulated iron at this temperature. We speculate that proper trafficking of Cfo1 to the vacuole contributes to robust iron uptake, correct intracellular distribution, and overall homeostasis (Fig 10).
Our analysis of the Vps45-GFP fusion also revealed an unexpected co-localization with mitochondria, and we subsequently found that vps45 mutants were sensitive to specific inhibitors of electron transport and to reactive oxygen species. Additionally, the vps45 mutants appeared to be impaired in calcium homeostasis, and mitochondria are known to be important organelles for calcium storage and signaling. Given that mitochondria–endomembrane system connections are emerging as important features of intracellular communication [37,47], it is tempting to speculate that Vps45 functions in these connections. That is, Vps45 may directly or indirectly contribute to establishing or maintaining membrane contact sites between mitochondria and the vacuole or other endomembranes. Interestingly, contacts contribute to iron and calcium distribution in the cell including trafficking to mitochondria because it has been demonstrated in erythrocytes that iron loaded transferrin is directly transported from the endosomes to mitochondria [48]. The observed association of the Vps45-GFP fusion with mitochondria supports a possible direct contribution to membrane contact sites. However, alternative or additional possibilities are that loss of Vps45 might impair the delivery of proteins needed for contact site formation and/or vacuolar function to indirectly influence mitochondrial activities via changes in metabolism, response to oxidative stress and mitophagy [47,49]. These ideas are supported by genetic screens in yeast that identified many VPS genes as contributing to mitochondrial morphology [49,50].
The participation of Vps45 in other endosomal trafficking connections and/or impaired vacuolar function may also explain the additional phenotypes of the vps45 mutant. For example, impaired delivery of proteins and materials to the PM and cell wall may explain increased sensitivity to CFW, Caffeine, SDS and NaCl. Similarly, participation of Vps45 in vesicle transport to the ER or establishment of connections between the ER and the PM, vacuole or mitochondria might explain the increased sensitivities to tunicamycin, chloroquine, quinacrine, the azole antifungal drugs and capsofungin.
The vps45 mutant was also attenuated for virulence in a mouse model of cryptococcosis and for survival in a macrophage-like cell line. Adaptation to the host environment requires proper nutrient acquisition, and virulence factor elaboration. In this regard, defects in iron acquisition are known to attenuate the virulence of C. neoformans [3,9,12]. Importantly, the observed cell wall changes and altered capsule elaboration with reduced shedding also likely reduced virulence given the established importance of the capsule in fungal cell protection and immuno-modulatory properties. Capsular polysaccharide shedding (mainly composed of GXM) tends to accumulate in patient serum and cerebrospinal fluid and supress patient immune responses by limiting immune cell infiltration leading to devastating consequences. The release of GXM is regulated by environmental cues such pH and nutrient availability [51]. Furthermore, many of the phenotypes of the vps45 mutants are exacerbated at 37°C and this would further preclude robust proliferation in mammalian hosts. Additional phenotypes for signaling (calcium), and mitochondrial function (respiration and ROS maintenance) could also potentially reduce growth in the host as mitochondria metabolism and ROS accumulation are linked to capsule formation and enlargement [52].
Our results support the need for additional work on Vps45 and other SM proteins for C. neoformans. In particular, the identification of interacting partners and a more detailed analysis of the location vis-à-vis mitochondria in different growth conditions is needed to more fully understand the contribution of Vps45. Vps45 and Vps33 were recently characterized in the fungus Aspergillus nidulans as Sec1/Munc-18 proteins involved in the regulation of the syntaxin PepAPep12 which contributes to membrane compartment identity. The authors demonstrated that Vps45 plays role in regulating PepAPep12 in the early endosomes formation whereas Vps33 contributes to the regulation of PepAPep12 in late endosomes/vacuoles [31]. Bioinformatics analysis of H99 genome indicated that a homologue of Vps33 does exist (CNAG_02373), but further investigation is needed to characterize the role of VPS33 in endosome sorting in C. neoformans. A homologue of Sly1 was also found in H99 (CNAG_05933) which is annotated as a hypothetical protein and identical at 36% to ScSly1 from S. cerevisiae. Likewise, a homologue of ScSec1, the syntaxin-binding protein 1 (CNAG_07800), was found with 28% identity in H99. The roles of CNAG_05933 and CNAG_07800 have yet to be characterized.
Overall, this study provided new insights into the machinery that mediates essential nutrient uptake, trafficking of material between organelles and the relevance of endocytosis to the pathogenicity of C. neoformans. Many questions remain regarding the roles of additional trafficking components and the mechanisms of delivery to target organelles, particularly in the context of iron and heme delivery to mitochondria.
The serotype A strain H99 of Cryptococcus neoformans var. grubii was used for all experiments and was maintained on YPD medium (1% yeast extract, 2% peptone, 2% dextrose, 2% agar). The Cfo1-GFP strain was constructed as described previously [9]. Growth under low iron conditions was performed in yeast nitrogen base (YNB) without amino acids media plus 2% dextrose at pH 7.0 with the addition of the iron chelator bathophenanthroline disulfonate (100–150μM BPS) (YNB-BPS). Defined low-iron media (LIM) was prepared as described [53] with the addition of 20 mM HEPES and 22mN NaHCO2. Mammalian iron sources such as human hemoglobin (2μg mL-1), bovine heme (10–100μM), human holo-transferrin (50μg mL-1), and sheep blood (0.05%), as well as ferric chloride (FeCl3) (10–100μM) were added to cultures. Different compounds were added at the following concentrations: 500nM N-Ethylmaleimide (NEM), 20μg mL-1 brefeldin A (BFA), 500μg mL-1 monensin, 150ng mL-1 tunicamycin (TNC), 25μM aminodarone (AMD), 100μg mL-1 cyclosporine A (CSA), 500ng mL-1 tacrolimus (FK506), 5mM or 50mM ethylene glycol-bis(2-aminoethylether)-N,N,N′,N′-tetraacetic acid (EGTA), 50μg mL-1 rotenone, 1mM malonic acid, 10 mM salicylhydroxamic acid (SHAM). 2μg mL-1 antimycin A, 5mM potassium cyanide (KCN), 0.01% hydrogen peroxide (H2O2), 50μM plumbagin, 5μg mL-1 menadione, 1 mg mL-1 calcofluor white (CFW), 0.01% sodium dodecyl sulfate (SDS), 0.5mg mL-1 caffeine, 1.5M sodium chloride (NaCl), 10 μg mL-1 caspofungin, 1.6mM quinacrine, 6mM chloroquine, 10μg mL-1 fluconazole and 50ng mL-1 miconazole. All chemicals were obtained from Sigma-Aldrich.
All deletion mutants were constructed by homologous recombination using gene specific knockout cassettes, which were amplified by three-step overlapping PCR [54] with the primers listed in S1 Table. The resistance marker for nourseothricin (NAT) was amplified by PCR using the primers Cassette F and Cassette R and the plasmid pCH233 as a template. The gene-specific knockout primers 1 and 2, and 3 and 4 were used to amplify the flanking sequences of their respective genes; and primers 5 and 6 were used to amplify the gene-specific deletion construct containing the resistance marker. All constructs for deletions were introduced into the H99 wild-type (WT) and Cfo1-GFP strains by biolistic transformation, as described previously [55]. Complementation of the vps45Δ mutants was performed by cloning a PCR product of VPS45 into the integrative Safe Haven vector pSDMA58. The plasmid construct was linearized with PacI and introduced into the vps45 and Cfo1-GFP strains by biolistic transformation. Multiplex PCR was performed on genomic DNA of hygromycin-resistant colonies using primers UQ1768, UQ2962, UQ2963 and UQ3348, as described [56]. The GFP coding sequence was added to the pSDMA58-VPS45 plasmid using fast cloning techniques [57]. The GFP sequence was amplified from the pHD58 plasmid whereas the GFP sequence flanked by the GAL7 terminator was introduced into pSDMA58. The plasmid construct was linearized with PacI and introduced into the vps45 strain by biolistic transformation. PCR for deletion constructs was performed using ExTAQ polymerase (TaKaRa Bio Inc) and the Phusion High Fidelity DNA Polymerase (New England Biolabs, USA) was used for complementation and fast cloning.
To assess growth on solid media, 10-fold serial dilutions of cells were spotted on YPD agar plates containing the compounds as indicated in the text. Plates were incubated at 30°C and 37°C for 2–5 days before being photographed. For growth assays in liquid and solid YNB-BPS media, cells were pre-grown in YNB + 150 μM BPS for 48h at 30°C, washed 3 times with iron-free water (chelex-treated water) and counted. 1X105 cells mL-1 were inoculated into YNB-BPS + iron sources for liquid assays and 10-fold serial dilutions (started at 1X106 cells mL-1) were spotted onto plates of solid media. Flasks and plates were incubated at 30°C and 37°C and optical density at 600 nm was measured after 4h and every 24h using a spectrophotometer (BECKMAN DU 530) for liquid assays, whereas plates were photographed after 2–4 days.
Staining of cells with the lipophilic dye FM4-64 [N-(3-triethylammoniumpropyl)-4-(6-(4-(diethylamino) phenyl) hexatrienyl) pyridinium dibromide] (T-3166; Invitrogen, Ontario, Canada) was performed to observe internalized vesicle trafficking and vacuolar membrane. The cells were stained with 5 μM for FM4-64 in PBS. The cells were incubated for an additional 90 min at 30°C and 37°C and pictures were taken every 15 min using an Olympus Fluoview FV1000 laser scanning confocal system. Colocalization analyses were performed on images of whole cells using the ImageJ coloc2 test. Positive correlations between Vps45-GFP and subcellular region were determined by Pearson’s R value (0.10–0.65) and Costes P-value (0.95–1.00) [33]. A total of 15–45 cells were analysed per condition.
Cells were grown overnight at 30°C and 37°C in YPD, washed in PBS, and resuspended at 107 mL-1 for staining as follows (all manipulations at RT). For Eosin Y (Sigma), cells were resuspended in McIlvaines buffer, pH 6.0, and stained with 250μg mL-1 of the dye. For Pontamine (Pontamine fast scarlet 4B, Bayer Corp., Robinson, PA), cells were stained in PBS with 100μg mL-1. For Concanavalin A-FITC (Sigma-Aldrich), cells were stained with 30μg mL-1 in Hepes-buffered saline, pH 7.0, containing 10mM each MgCl2 and CaCl2. Cells were stained for 15 min at room temperature, and then washed three times in appropriate buffers. Cells were then resuspended in PBS with 10mM NaN3 and analyzed by flow cytometry on a BD FACSCalibur. The acidic pH sensor carboxy-DCFDA (5-(and-6-)-Carboxy-2’7’-dichlorofluorescein diacetate) (ThermoFisher Scientific) was used to assess intracellular acidification. Overnight cultures were stained with 50μM carboxy-DCFDA and incubated for 30 min at 30°C and 37°C. Cells were spun down and resuspended in 10mM NaN3 in PBS. 50, 000 cells were analyzed by flow cytometry on a BD LSR II-561. JC-1 (ThermoFisher Scientific was used to stain mitochondria and assess their membrane potential (MMP). Overnight cultures were stained with 5μM JC-1 for 30 min at 30°C and 37°C, washed 3 times, resuspended in media, and analyzed by flow cytometry on a BD LSR II-561. Flow analysis was performed using FlowJo software.
Capsule formation was examined by differential interference contrast microscopy on an Axio imager M.2 microscope (Zeiss) with magnification X1000 after incubation for 48 h at 30°C and 37°C in defined LIM and staining with India ink. Capsule measurements were performed using ImageJ. Capsule shedding from cells was examined with a blot assay using the anti-GXM 18B7 antibody as described [43].
Cellular iron and calcium contents were measured by inductively coupled plasma mass spectrometry (ICP-MS). Briefly, cells were pre-grown iron starved cells inoculated for 48h in YNB + 150 μM BPS ± 100μM FeCl3 at 30°C or 37°C, washed 3 times with metal-chelated PBS, and lyophilized. The lyophilized cells were digested with 5 mL of HNO3 and 3 mL of H2O2 using a microwave digestion system (START D). Total iron and calcium contents were analyzed using an OPTIMA 5300 DV (PerkinElmer) system.
Macrophage infections were performed as described previously [58]. Briefly, macrophage-like J774.A1 cells were grown to 80% confluence in DMEM supplemented with 10% fetal bovine serum and 2mM L-glutamine at 37°C with 5% CO2. Macrophages were stimulated 2 h prior to infection with 150ng mL-1 phorbol myristate acetate (PMA). Fungal cells were grown in YPD overnight and PBS-washed cells were opsonized in DMEM with 0.5 μg mL-1 of the monoclonal antibody 18B7 for 30 min at 37°C. Stimulated macrophages were infected with 2x105 opsonized fungal cells (MOI 1:1) for 2 and 24 h at 37°C with 5% CO2. Macrophages containing internalized cryptococci cells were washed thoroughly 3 times with PBS and then lysed with sterile water for 30 min. Lysate dilutions were plated on YPD agar and incubated at 30°C for 48 hrs, at which time the resulting CFUs were counted. Statistical significance of intracellular survival was determined by two-tailed unpaired t-tests (GraphPad Prism 7 for Windows, GraphPad Software, San Diego, CA).
The virulence of the WT strain H99, the vps45Δ-34 mutant and the complemented strain vps45Δ::VPS45 [53] was assessed using female BALB/c (4 to 6 weeks old) from Charles River Laboratories (Ontario, Canada). A mice survival assay with assessment of fungal burden at the humane endpoint was executed. Fungal cells were grown in 5mL of YPD at 30°C and washed twice with PBS (Invitrogen). Mice were anesthetized intraperitoneally with ketamine (80 mg kg-1) and xylazine (5.5 mg kg-1) and suspended on a silk thread by the superior incisors. A suspension of 2X105 cells in 50μL was slowly inoculated into the nares of the mice. The health status of the mice was monitored daily post-inoculation and mice reaching the humane endpoint were euthanized by CO2 anoxia. Fungal burden of organs (lungs, brain, liver, spleen, and kidney) and cardiac blood was assessed. The organs and blood were aseptically removed. Blood was retrieved from the heart using sterile syringes pre-rinsed with 500 units of heparin. Organs were weighed, and homogenized using a Retsch MM301 mixer mill. The samples were serial diluted, plated on YPD containing chloramphenicol (30μg mL-1) and incubated at 30°C for 2 days; CFUs were then counted. Statistical analyses of survival differences in mice were performed with the log rank test and a two-tailed unpaired Mann-Whitney test was used to assess the fungal load (GraphPad Prism 7 for Windows, GraphPad Software, San Diego, CA).
This study was carried out in strict accordance with the guidelines of the Canadian Council on Animal Care. The protocol for the virulence assays employing mice (protocol A17-0117) was approved by the University of British Columbia Committee on Animal Care.
|
10.1371/journal.ppat.1007016 | MoYvh1 subverts rice defense through functions of ribosomal protein MoMrt4 in Magnaporthe oryzae | The accumulation of the reactive oxygen species (ROS) in rice is important in its interaction with the rice blast fungus Magnaporthe oryzae during which the pathogen scavenges ROS through the production of extracellular enzymes that promote blast. We previously characterized the MoYvh1 protein phosphatase from M. oryzae that plays a role in scavenging of ROS. To understand the underlying mechanism, we found that MoYvh1 is translocated into the nucleus following oxidative stress and that this translocation is dependent on MoSsb1 and MoSsz1 that are homologous to heat-shock protein 70 (Hsp70) proteins. In addition, we established a link between MoYvh1 and MoMrt4, a ribosome maturation factor homolog whose function also involves shuttling between the cytoplasm and the nucleus. Moreover, we found that MoYvh1 regulates the production of extracellular proteins that modulate rice-immunity. Taking together, our evidence suggests that functions of MoYvh1 in regulating ROS scavenging require its nucleocytoplasmic shuttling and the partner proteins MoSsb1 and MoSsz1, as well as MoMrt4. Our findings provide novel insights into the mechanism by which M. oryzae responds to and subverts host immunity through the regulation of ribosome biogenesis and protein biosynthesis.
| ROS accumulation is important for the interaction between the blast fungus M. oryzae and its rice host. The protein phosphatase MoYvh1 affects the scavenging of host-derived ROS that promotes M. oryzae infection. We found that MoYvh1 is translocated to the nucleus under oxidative stress by a mechanism that is dependent on its interactions with MoSsb1 and MoSsz1. MoYvh1 triggers the release of MoMrt4 from the ribosome in the nucleus that contributes to ribosome maturation. Importantly, we have provided evidence to demonstrate that MoYvh1 is important for the synthesis of extracellular proteins that are involved in ROS scavenging. Our findings provide insight into the mechanism by which M. oryzae responds to host immunity through MoYvh1 that regulates ribosome function to evade the host defense response.
| Magnaporthe oryzae is the causal agent of rice blast and also an established model organism to study plant-pathogen interactions [1,2]. In a previous study, we have characterized MoYvh1 as a homolog of the budding yeast Saccharomyces cerevisiae protein phosphatase Yvh1 that regulates growth, sporulation, and glycogen accumulation [3]. We found that MoYvh1 not only plays a similar important role in vegetative growth and conidia formation but also regulates virulence [4]. In addition, we found that deletion of MoYVH1 results in an increased accumulation of the host-derived reactive oxygen species (ROS) [4]. ROS levels are known to govern the pathogen and host interaction, how MoYvh1 regulated growth and virulence is linked to its role in affecting ROS levels remains an interesting but unresolved research subject.
S. cerevisiae Yvh1 is known to also have a role in ribosome maturation and function [5]. In eukaryotic cells, mature ribosomes are composed of five different proteins that include Rpp0 and two copies of each of proteins P1 and P2 [6–8]. Rpp0 interacts directly with the 60S ribosome subunit to form the base of the stalk for binding to P1 and P2 proteins [9,10]. Also in S. cerevisiae, the ribosome assembly factor and the nucleolar protein Mrt4 are closely related to Rpp0, based on the conserved N-terminal ribosome binding domain they shared with [11,12].
Eukaryotic cells respond to environmental stresses, including elevated temperatures, via a family of well-characterized heat-shock proteins (Hsp) [13]. As ubiquitous molecular chaperones that function in a wide variety of cellular processes, Hsp70s act by reversibly binding and releasing the short hydrophobic stretches of amino acids in a nucleotide-dependent fashion [14,15]. Hsp70 heat shock proteins are known to affect ribosomal function and protein biosynthesis [16]. For example, the ribosomal L31 protein binds to chaperone Zuo1 that in turn anchors Hsp70 Ssz1 and Hsc70 proteins to regulate polypeptide translocation [17–21].
Given the multifaceted role of MoYvh1 previously established [4], further addressing of MoYvh1 functional mechanisms would promote our understanding of the rice blast mechanisms. We here showed that MoYvh1 is translocated to the nucleus under the oxidative stress condition and that MoYvh1 functions through interactions with Hsp70 protein homologs MoSsb1 and MoSsz1. In addition, we showed that MoYvh1 is required for proper translocation of the ribosomal maturation factor homolog MoMrt4, since the loss of MoYvh1 caused MoMrt4 mislocalization to the cytoplasm resulting in virulence defects.
We have identified MoYvh1 as a homolog sharing amino acid sequence conservation with S. cerevisiae Yvh1 that in turn shares homology with the dual-specificity phosphatase from vaccinia virus [22]. We found that MoYvh1 has a multiple role in the growth and virulence of M. oryzae and that deletion of MoYVH1 results in an accumulation of ROS surrounding the infection sites [4]. To address whether MoYvh1 exhibits a cytoplasmic-nuclear shuttling ability, similar to S. cerevisiae Yvh1, we constructed the strains expressing MoYvh1-GFP, in which the expression of the C-terminal GFP fusion protein is under the control of the native MoYVH1 promoter. Notably, MoYvh1 was present in both the cytoplasm and the nucleus in conidia, which is the expected default steady-state distribution pattern. Treated with 5 mM H2O2 for 2 hours (h), an enhanced nuclear localization was observed in conidia (68.42 ± 7.31%) (Fig 1A). In the aerial hyphae, however, no changes were seen under the same stress condition (Fig 1B). Plants generate a vast array of oxidative agent in response to pathogen invasion including superoxide radical and hydroxyl radical. To further understand the changes observed in the localization pattern of MoYvh1 was specific to H2O2 or general to other oxidative stress, KO2 and hydroxyl radical were used to treat the ΔMoyvh1/MoYVH1-GFP strain. The results showed that both KO2 and hydroxyl radical induced an accumulation of MoYvh1 in the nucleus in conidia but not the aerial hyphae (S1 Fig). These data suggested that oxidative stress could promote cytoplasmic MoYvh1 nuclear localization in conidia.
To understand MoYvh1 functions associated with its nuclear translocation, we identified MoSsb1 and MoSsz1 that are heat-shock 70 (Hsp70) protein homologs following screening a yeast two-hybrid cDNA library constructed with RNA pooled from various stages including conidia and infections (0, 2, 4, 8, 12 and 24 h) (Fig 2A). We then validated these interactions by co-introducing the MoYVH1-FLAG and MoHSP70s-GFP fusion constructs into the protoplasts of the wild type strain Guy11. Total proteins were extracted from conidia of the putative transformants, and MoYvh1, MoSsa1, MoSsb1, and MoSsz1 were detected using the anti-FLAG and anti-GFP antibodies. In proteins eluted from MoSsb1 and MoSsz1 anti-GFP beads, MoYvh1 was detected (Fig 2B). The interactions were further confirmed by the bimolecular fluorescence complementation (BiFC) assay. The MoYVH1-CYFP and MoSSB1-NYFP, MoSSZ1-NYFP fusion constructs were introduced into the protoplasts of Guy11, with the empty vectors used as negative controls. The recombined YFP fluorescence signal was detected in the cytoplasm containing corresponding protein pairs (Fig 2C and S2 Fig). Interestingly, the interactions were also observed under the oxidative stress, with YFP fluorescence being transferred to the nucleus following treatment with 5 mM H2O2 (Fig 2C). Previously, we demonstrated that the MoYvh1 C-terminal zinc-binding domain is required for growth and virulence of the fungus. Our evidence here showed that the same C-terminal is also responsible for binding with MoSsb1 and MoSsz1 (S3 Fig).
An interaction between MoYvh1 and MoSsa1 could not be established, suggesting that MoYvh1 interactions with MoSsb1 and MoSsz1 are specific (Fig 2A, 2B and 2C). We also generated MoSSB1 and MoSSZ1 deletion mutants and assessed their effects on MoYvh1 distribution. The MoYVH1-GFP fusion construct was introduced into ΔMossb1, ΔMossz1, and ΔMoyvh1 mutants. In the resulting transformants, GFP signal was observed in both the cytosol and the nucleus. However, the GFP signal was predominantly observed in the cytoplasm of ΔMossb1 and ΔMossz1 upon oxidative stress, in contrast to complemented strains (72.16 ± 5.77%) where GFP was predominantly seen in the nuclei (Fig 2D). To further evaluate the nuclear translocation of MoYvh1 in these strains, we separated nuclear proteins from cytoplasmic ones and performed Western blotting analysis. MoYvh1-GFP was significantly enriched in the nucleus in the complement strain when treated with 5 mM H2O2. However, MoYvh1-GFP was uniformly distributed in the conidia of the ΔMossb1 and ΔMossz1 mutants (S4 Fig). These results suggested that MoSsb1 and MoSsz1 could facilitate nuclear translocation of MoYvh1 under oxidative stress through direct interactions.
To further understand MoYvh1 nuclear translocation upon stress and associated functions, we searched for additional proteins that interact with MoYvh1 and identified MoMrt4 (MGG_08908) that shares sequence homolog with S. cerevisiae nucleolar protein Mrt4. The yeast Mrt4 contains a Gly residue at position 68 whose substitution with Asp or Glu could suppress the growth defect of the Δyvh1 strain [5,23,24]. To investigate whether MoYvh1 shares functional conservation with S. cerevisiae Mrt4, we constructed strains expressing MoMRT4G69D-GFP and MoMRT4G69E-GFP, respectively. We found that MoMrt4G69D and MoMrt4G69E, but not MoMrt4, were able to rescue the defect on growth and virulence of the ΔMoyvh1 strain (Fig 3A, 3B and 3C and S5A Fig). Because MoYvh1 functions upstream of MoPdeH to regulate the cAMP levels and pathogenicity [4], we also found that MoMrt4G69D and MoMrt4G69E suppress the defects in cAMP levels of the ΔMoyvh1 mutant (S5B Fig).
As MoYvh1 plays a role in scavenging host-derived ROS, we examined ROS levels by staining host cells with 3, 3’-diaminobenzidine (DAB) at 36 h after inoculation. The primary infected rice cells with infectious hyphae of the ΔMoyvh1 and ΔMoyvh1/MoMRT4 strains were stained intensely by DAB, with reddish-brown precipitate around the infected cells, while the ΔMoyvh1/MoMRT4G69D and ΔMoyvh1/MoMRT4G69E strains exhibited weak staining, a phenotype similar to Guy11 (Fig 3D).
MoMrt4 is normally accumulated in the nucleus of the wild-type strain (Fig 4A). To study how MoMrt4G69D and MoMrt4G69E suppress the defects of the ΔMoyvh1 mutant, we assessed the effect of MoYvh1 on the subcellular localization of MoMrt4. As expected, MoMrt4G69D and MoMrt4G69E were predominantly nuclear localized, while MoMrt4 was mostly cytoplasmic, in the ΔMoyvh1 mutant (Fig 4A). As MoMrt4G69D and MoMrt4G69E mutation showed similar roles in the ΔMoyvh1 mutant, we used the ΔMoyvh1/MoMrt4G69E strain to determine whether its affinity for the ribosome was compromised. We found that binding of MoMrt4G69E to the ribosome was more sensitive to 100 and 500 mM NaCl than MoMrt4 that was largely unaffected. 500 mM NaCl caused the majority of MoMrt4G69E to be dissociated from the ribosome (Fig 4B). Therefore, MoMrt4G69E showed weaker affinity for ribosomes than MoMrt4, implying easier separation from the ribosome. We further speculated that the affinity for ribosomes between MoYvh1 and MoMrt4 is important for the normal function of M. oryzae. To test this hypothesis, we assessed whether MoMrt4 competes with MoYvh1 in ribosome binding. Western blotting analysis showed that MoYvh1 bound to the ribosome in both the wild-type and the ΔMomrt4 mutant. However, the MoMrt4 recruitment to ribosomes in the presence of MoYvh1 was significantly reduced in the wild-type strain (Fig 4C), suggesting that MoYvh1 and MoMrt4 indeed compete for binding to the ribosome.
The Rpp0 protein is one of the five conserved components of mature ribosomes [7,10]. We have cloned the MoRpp0 homolog and generated the ΔMoyvh1/MoRpp0-FLAG-MoYvh1-GFP, ΔMomrt4/MoRpp0-FLAG-MoYvh1-GFP, and Guy11/MoRpp0-FLAG-MoYvh1-GFP strains to investigate whether deletion of MoYvh1 or MoMrt4 causes any defects in ribosome maturity. Ribosome proteins were extracted. In the ΔMomrt4 mutant, MoRpp0 was bound to the ribosome similar to that in the wild-type strain. However, MoRpp0 remained in the suspension in the ΔMoyvh1 mutant (Fig 4D), suggesting that MoYvh1 has a role in ribosome maturity.
Since MoMrt4 is important for MoYvh1 function, we characterized its function in growth and pathogenesis. The ΔMomrt4 mutant displayed significantly attenuated growth on CM, minimal medium (MM), straw decoction and corn agar (SDC), and oatmeal medium (OM) plates (Fig 5A and S5C Fig). Conidia formation was drastically reduced in the ΔMomrt4 mutants by ~70% when compared with the wild-type strain (Fig 5D and 5E). To determine whether MoMrt4 plays a role in pathogenicity, susceptible CO-39 rice seedlings were respectively sprayed with the conidia of the wild-type, ΔMomrt4 mutant, and complemented strains. The production of fewer, small lesions by the ΔMomrt4 mutant at 7-day post-inoculation (dpi) (Fig 5B and 5C) indicated that MoMrt4 is required for full virulence.
Our previous study showed that deletion of MoYVH1 results in an increase in the accumulation of ROS but reduced virulence [4]. To test that the reduced virulence was due to a lack of ROS scavenging, we examined host-derived ROS levels by DAB staining. At 30 h after inoculation, no staining was observed in the primary rice cells infected by the ΔMomrt4 mutant (S6 Fig). We also evaluated binding of MoMrt4 to ribosomes in these strains and found that MoRpp0 remained in the suspension of the ΔMoyvh1 mutant. However, MoRpp0 bound to the ribosome in the ΔMomrt4 mutant which was similar to that in Guy11, indicating that deletion of MoMRT4 was not involved in the ribosome maturity, in contrast to MoYvh1 (Fig 5F). These results revealed that MoMrt4 is required for vegetative growth, conidiation, and full virulence, but these functions are independent of ribosome maturity.
As the nuclear localization of MoYvh1 is enhanced in conidia upon oxidative stress, we hypothesized that MoYvh1 is also translocated to the nucleus during host-imposed stress during infection. To test this, we screened rice cultivars resistant to Guy11 and the ΔMoyvh1/MoYVH1 strains. We found that the wild type strains caused only the restricted lesions on the rice cultivar K23 that contains the resistant gene Pi12 [25] (Fig 6A). As the ΔMoyvh1/MoYVH1 strains showed restricted lesions on the K23, DAB was further used to evaluated the host-derived ROS accumulated around the infection sites in K23. Cells with ΔMoyvh1/MoYVH1-GFP infectious hyphae on rice cultivar K23 were stained by DAB, with the reddish-brown precipitate around the infected cells, indicating that the ΔMoyvh1/MoYVH1-GFP strain fails to scavenge H2O2 on K23 (Fig 6B). To assess nuclear translocations of MoYvh1, we extracted nuclear proteins and performed Western blotting analysis. In K23, MoYvh1-GFP was significantly enriched in the nucleus in comparison with LTH (Fig 6C). As it is difficult to stain the nucleus by DAPI during infection, we used the histone H1 fused to red fluorescent protein (RFP) to mark the nucleus of infectious hyphae. We found an enrichment of MoYvh1 in the nucleus when co-localization with H1-RFP in the infection hyphae of K23, in comparison with LTH cultivar (Fig 6D). However, MoYvh1 was not translocated into the nucleus in the ΔMossb1 and ΔMossz1 mutants (S7 Fig). These results indicated that MoYvh1 is indeed nuclear enhanced during infection.
At the early stages of infection, M. oryzae secretes numerous effector proteins to suppress plant defense responses and modulate host cellular processes that promote infections [26]. Since MoYvh1 has a role in ribosome maturity (Fig 5D) and the ribosome is important for the synthesis of proteins, we tested whether the production of extracellular proteins was compromised in the ΔMoyvh1 mutant. The extracellular fluid (EF) was prepared as described by Patkar and colleagues [27]. Conidia from Guy11 and the ΔMoyvh1 strains were inoculated on a hydrophobic glass sheet and EF was harvested following 24 h incubation. We first detected the localization of MoYvh1 under this condition and the results showed that MoYvh1 was present in both nucleus and cytoplasm, indicating MoYvh1 functions in the nucleus during the appressorium formation (S8 Fig). EF extracts from wild type were subsequently added to the rice leaf sheaths following infection by the ΔMoyvh1 mutant. We found that native EF, but not that denatured by boiling, rescued the defects in host cell invasion and ROS scavenging at the infected sites (Fig 7A and 7B). Our previous study showed that deletion of MoYVH1 resulted in reduced peroxidase and laccase activities [4], so we further assayed the peroxidase and laccase activities in both EF harvested from the wild type and ΔMoyvh1 mutant strains. The enzyme activity assay was performed as described by Chi and associates [28] by using the EF from both Guy11 and ΔMoyvh1 mutant. We observed very low levels of laccase and peroxidase activities in the ΔMoyvh1 mutant when compared with Guy11 (Fig 7C). We further performed the spray and drop assays on rice leaves. Conidia of the ΔMoyvh1 mutant were collected with 5 ml of the EF or boiled EF of Guy11. The conidial suspensions of each treatment were sprayed onto rice leaves. After inoculation for 7 days, the results showed that the EF of Guy11 could suppress the defects of the ΔMoyvh1 mutant in pathogenicity (Fig 7D and 7E). The conidial suspensions of each treatment were also drop inoculated onto detached rice leaves and the results revealed that the EF of the wild type strain partially rescues the defect in pathogenicity on the detached rice leaves (S9 Fig). These results indicated that the EF of Guy11 contains candidate proteins that are important for infection.
To identify candidate proteins in EF regulated by MoYvh1, we compared the EF production through SDS-PAGE analysis and found that the amount of EF proteins in ΔMoyvh1 EF was significantly less than that of wild type Guy11 (Fig 7F). In addition, mass spectrometry (MS) analysis revealed the presence of over 70 proteins with signal peptides in EF of the wild type strain but not of the ΔMoyvh1 mutant. To address whether the absence of these proteins is caused by the defect in ribosomal biogenesis, we randomly chosen 30 of them to evaluate the transcriptional difference between Guy11 and the ΔMoyvh1 mutant in the conidia after 24 h incubation on a hydrophobic glass sheet. Among these genes, only three were significantly reduced in transcription (p < 0.01) (S10 Fig). In 70 identified proteins, 13 were associated with oxidoreducation (Fig 7G and S2 Table). Thus, the defect in scavenging host-derived ROS of the ΔMoyvh1 mutant was associated with the defect in the production of extracellular proteins.
Virulence in the rice blast fungus M. oryzae is a multifaceted trait contributed by not only the complex circuitry in the pathogen side but also that of the host. In dissecting molecular events leading to virulence, we have previously characterized the dual specificity phosphatase MoYvh1 that shares sequence conservation and functional mechanisms to certain degree with S. cerevisiae Yvh1. Importantly, we found that MoYvh1 plays a role in not only growth, conidia formation, but also virulence in M. oryzae [4]. Here, we provided mechanistic evidence to show that MoYvh1 undergoes cytoplasmic to nuclear translocation in response to oxidative stress and that MoYvh1 affects ribosome maturation. Our findings reveal a novel link between ribosome biogenesis and fungal virulence that is mediated by MoYvh1.
In S. cerevisiae, Yvh1 is a shuttling protein that could remain in the nucleus if fused with a nuclear localization sequence [23,29]. Previous studies also found that Yvh1 binds with the pre-60S ribosome subunit to export it to the cytoplasm. Once arrives there, Yvh1 is released from the pre-ribosome following mature ribosomal protein P0 binding to the ribosomal stalk [23,24,30,31]. How Yvh1 is translocated into the nucleus remains unclear. Through studies of MoYvh1 here, we provided evidence that MoYvh1 exhibits similar nucleo-cytoplasmic shuttling ability and that it functions through interactions with MoSsb1 and MoSsz1.
An unexpected finding is that the interaction between MoYvh1 and MoSsb1 and MoSsz1 differs from aerial hyphae to conidia and the infection stage (Fig 2, S2 and S11 Figs). Similar differentiated interactions were seen before. A BiFC assay showed that Pth11 and Rgs1 interacted in vivo during early pathogenesis but not during vegetative growth [32]. The interaction between MoCap1 and MoMac1 was weak during vegetative growth but was enhanced during appressorium formation [33]. In addition, Liu and colleagues showed that MoAtg4 interacts with MoAtg8 only under the nitrogen starvation condition [34]. In view of these findings, we hypothesized that 1) the interaction occurs only under oxidative stress, and 2) MoYvh1 and Hsp70s interactions are developmental stage specific. In agreement with these hypotheses, the YFP fluorescence signal was transferred to the nucleus only following treatment with 5 mM H2O2 (Fig 2C). Therefore, we concluded that MoSsb1 and MoSsz1 recruitments to MoYvh1 to facilitate its nuclear translocation upon oxidative stress during specific growth stages in M. oryzae. Further evidence showed that MoYvh1 is not accumulated in the nucleus when MoSsb1 and MoSsz1 are not interacted with MoYvh1 during the aerial hyphae (Fig 1B and S11 Fig). And also, deletion of MoSSB1 or MoSSZ1 which interdicted the interaction caused cytoplasmic-MoYvh1 not translocated into the nucleus even in the conidia and infection stages (Fig 2D, S4 and S7 Figs). These results further confirmed that the accumulation of MoYvh1 in the nucleus under oxidative stress is dependent on the interaction between MoSsb1 and MoSsz1. Upon the exposure to oxidative stress, MoYvh1 nuclear translocation is accelerated by its interaction with MoSsb1 or MoSsz1 during the conidial and infection stages. Since MoYvh1 still could be detected in the nucleus in the ΔMossb1 and ΔMossz1 mutants (S4 Fig), we postulated that additional translocation facilitators of MoYvh1 that are independent of oxidation stress may also exist.
We found that MoYvh1 and MoMrt4 bind with ribosomes in a competitive manner. To further examine the relationship between MoYvh1 and MoMrt4, we characterized the function of MoMrt4 by generating a ΔMomrt4 mutant, which is significantly attenuated in growth, conidia production, and pathogenicity. However, the lesions produced by the ΔMomrt4 mutant on rice leaves were fewer and smaller than those of the ΔMoyvh1 mutant. A DAB assay suggested that deletion of MoMRT4 did not affect the scavenging of ROS accumulated around the infection sites. Our analysis further suggested that MoYvh1 binds to pre-ribosomes and thereby helps to release MoMrt4. Thus, ribosome immaturity resulted in pathogenicity defects in the ΔMoyvh1 mutant but not the ΔMomrt4 mutant. The ribosome extract assay confirmed that cells continue to synthesize ribosomes in the absence of MoMrt4. This finding is consistent with studies in S. cerevisiae in which the deletion of MRT4 causes defects in growth but no blocking in ribosome synthesis [5,11,23,35].
M. oryzae secretes a wide array of factors into the host to facilitate invasion [26,36,37]. However, host plants have evolved to recognize these effectors and counteract by activating defense responses to limit pathogen spreading [38–41]. Small-molecule phytohormones, such as jasmonates, salicylic acid and brassinosteroids, play key roles in regulating this defense response [42–44]. Our recent findings showed that the scavenging of host-derived ROS at the infection site is important for virulence of the ΔMoyvh1 mutant, as deletion of MoYVH1 causes virulence defect [4]. Consistent with the findings, the EF of the wild-type scavenges ROS accumulated around the sites of infection, in contrast to the ΔMoyvh1 mutant, where the restricted invasion is the result of ROS accumulation due to lack of extracellular proteins in EF. When treated with ROS, MoYvh1 in the cytoplasm is translocated into the nucleus causing an enhanced nuclear localization in both the conidia and invasion hyphae (Figs 2, 7C and 7D). In the mycelium, however, the cytoplasmic location pattern of MoYvh1 remained unchanged. These results suggested that the differential localization patterns of MoYvh1 might be developmentally regulated and it may be relevant to virulence. Since H2O2 stress blocks the formation of appressorium, the glass surface was not subjected to H2O2 stress allowing conidia to germinate, and under this condition MoYvh1 was present in both nucleus and cytoplasm (S8 Fig). Here, we found that the EF of non-induced Guy11 rescued the defects in invasion and ROS scavenging around the infected sites (Fig 7A and 7B), suggesting that original MoYvh1 in the nucleus (S8 Fig) without treatment regulated the ribosome maturation that provides abundant extracellular proteins to inhibit host-derived ROS. Under ROS stress, MoYvh1 in the cytoplasm translocates into the nucleus and accelerates ribosome synthesis to produce more extracellular proteins, which inhibits host-derived defense and promote infection.
Why does the wild-type EF suppress the defects of the ΔMoyvh1 mutant and does the EF contain necessary ribosome or other proteins? In S. cerevisiae, Rpp0 is loaded onto the 60S ribosome subunit to assemble the mature stalk [23,24]. In this study, deletion of MoYVH1 led to the separation of MoRpp0 from the ribosome, suggesting a ribosomal maturation defect in the ΔMoyvh1 mutant which would impair the production of proteins (Fig 5D). During the early stages of infection, M. oryzae secretes various extracellular proteins to suppress plant immunity for promoting colonization. Blockage of secreted protein synthesis due to immature ribosome in the ΔMoyvh1 mutant likely results in the defect in virulence. Indeed, we showed that EF from the wild type strain restored pathogenicity of the ΔMoyvh1 mutant when added to the conidia suspension (Fig 7A and 7C and S9 Fig). These findings indicated that MoYvh1 has a role in the production of EFs that inhibit rice immunity.
We therefore present a model for how MoYvh1 functions in growth, virulence, and host immune avoidance in M. oryzae (Fig 8). Our findings demonstrate that during M. oryzae infection, rice produces an ROS burst to suppress pathogen invasion. Under this stress, MoSsb1 and MoSsz1, together with MoYvh1, are translocated to the nucleus, where MoYvh1 has a role in ribosome maturation through the release of MoMrt4 from the pre-ribosome. Mature ribosomes promote EF synthesis and secretion to scavenge ROS and modulate the rice defense response. Our model reveals an important pathway by which M. oryzae recruits a nucleocytoplasmic shuttling phosphatase, MoYvh1, in response to host immune response. Further studies of MoYvh1-mediated response and the identification of EFs regulated by MoYvh1 would promote the understanding of M. oryzae pathogenesis mechanisms.
M. oryzae Guy11 strain was used as the wild type in this study. All strains were cultured on complete medium (CM) agar plates for 3–15 days at 28°C [45]. Mycelia were harvested from liquid CM and used for DNA and RNA extractions. Protoplasts were prepared and transformed as described previously [46]. Transformants were selected on TB3 medium (3 g of yeast extract, 3 g of casamino acids, 200 g of sucrose, and 7.5 g of agar in 1 l of distilled water) with 300 μg/ml hygromycin B (Roche) or 200 μg/ml zeocin (Invitrogen).
Ribosomal proteins were extracted from mycelia as previously described [23]. Briefly, fungal strains were cultured on solid CM medium for 7 days at 28°C, and approximately 1 x 1 mm square of agar containing the culture was inoculated in liquid CM and grown for another 2 days. Mycelia were filtered through Miracloth, blotted dry, and ground into powder in liquid nitrogen with a mortar and a pestle. 5 g mycelium was mixed with 15 ml 1st extraction buffer I (0.1 M natrium aceticum, 10 mM Tris-HCl, 10 mM MgCl2 with 0.07% β-mercaptoethanol) and incubated at 4°C for 2 h. Samples were centrifuged at 5000 g for 30 min at 4°C and repeated once before discarding the pellets. The upper phase was centrifuged at 96000 g at 4°C for 2 h and the pellet were dissolved in 2nd extraction buffer (20 mM Tris, pH 7.5, 6 mM MgCl2, 10% glycerol, 0.1% NP-40, 1 mM PMSF, 1 μM leupeptin, and 1 μM pepstatin A). 2.5 ml of protein extracts were overlaid on 7.5 ml 1 M sucrose in 20 mM Tris, 8 mM MgCl2, and 100 mM KCl in 10 ml ultracentrifuge tubes (Beckman Coulter). Samples were centrifuged again at 96000 g at 4°C for 2 h. Finally, the pellets were dissolved in dissolution buffer (5 M Urea, 2 M Thiourea, 2% CHAPS, 2% SB3-10, 40 mM Tris, 5mM Mercaptoethanol).
To generate the MoMRT4 gene replacement vector pCX62, approximately 1 kb upstream and downstream fragments were amplified with primer pairs (S1 Table). The resulting PCR products were ligated to the hygromycin resistance cassette released from pCX62, as previously described [25]. Putative mutants were screened by PCR and confirmed by Southern blotting analysis. To complement the ΔMomrt4 mutant, the DNA fragment containing the putative promoter and the coding sequence was amplified and inserted into pYF11 (bleomycin resistance) by homologous recombination in S. cerevisiae. Plasmids were extracted and introduced into Escherichia coli competent cells, and then the plasmids with correct inserts were introduced into protoplasts, as previously described [25].
cDNA of MoYVH1, MoYVH1ΔC (the N-terminus), MoYVH1ΔN (the C-terminus), MoSSB1, MoSSZ1 and MoSSA1 was respectively amplified with Super Fidelity DNA Polymerase (Vazyme, Nanjing). Amplified products were cloned into pGBKT7 and pGADT7 vectors (BD Biosciences, Oxford, UK), respectively. After sequence verification, they were introduced into yeast AH109 strain. Transformants grown on synthetic medium lacking leucine and tryptophan (SD–Leu–Trp) were transferred to synthetic medium lacking leucine, tryptophan and histidine (SD–Leu–Trp–His).
For BiFC assay, the MoYVH1-CYFP fusion construct was generated by cloning MoYVH1 into pHZ68 [47]. Similarly, MoSSB1-NYFP and MoSSZ1-NYFP fusion constructs were generated by cloning MoSSB1 and MoSSZ1 into pHZ65, respectively. Construct pairs of MoYVH1-CYFP, MoSSB1-NYFP and MoSSZ1-NYFP were introduced into the protoplasts of Guy11, respectively. Transformants resistant to both hygromycin and zeocin were isolated and confirmed by PCR.
To generate MoYVH1G69D and MoYVH1G69E constructs, the 2.7 kb upstream fragment including the MoYVH1 native promoter, the 1.1 kb fragment from the start codon of the coding sequence (containing the Gly 69), and the 0.5 kb downstream fragment including the rest of the gene coding sequence (containing both of the Gly 69) were co-introduced with XhoI digested pYF11 into yeast strain XK1-25 [47,48]. Plasmid pYF11::MoYVH1G69D and pYF11::MoYVH1G69E were rescued from the resulting Trp+ yeast transformants.
Conidial germination and appressorium formation were measured on a hydrophobic surface as previously described [49]. Appressorium induction and formation rates were obtained also as described previously [50,51].
For infection, conidia were harvested from 10-day-old SDC agar cultures, filtered, and resuspended to a concentration of 5 × 104 spores /ml in a 0.2% (w/v) gelatin solution. For the leaf assay, leaves from two-week-old seedlings of rice (Oryza sativa cv. CO39) and 7-day-old seedlings of barley were used for spray inoculation. For rice leaves, 5 ml of a conidial suspension of each treatment was sprayed. Inoculated plants were kept in a growth chamber at 25°C with 90% humidity and in the dark for the first 24 h, followed by a 12 h /12 h light /dark cycle. Lesion formation was observed daily and recorded by photography 7 days after inoculation [52,53].
Mycelia were harvested and ground into powder in liquid nitrogen. 1 mg mycelium was mixed with 20 μl 6% TCA solution, centrifuged (1700 g, 15 min), and top layers were collected. After washing twice with five volumes of anhydrous ether, pellets were collected and subjected to HPLC analysis using a programmable Agilent Technology zorbax 1200 series liquid chromatography. The solvent system consisted of methanol (90%): water (10%), at a flow rate of 1 ml /min. 0.1 mg/ml cAMP solution was eluted through the column (SB-C18, 5 μl, 4.6 × 250 mm) and detected at 259 nm UV. Samples were loaded through the column in turns.
Conidia of indicated strains were harvested from 10-day-old SDC agar cultures, filtered, and resuspended to a concentration of 1 × 105 spores /ml in a 0.2% (w/v) gelatin solution. 4 ml of the suspension was sprayed onto rice leaves and harvested 24 hpi. 5 g of Leaves were ground into powder in liquid nitrogen. The powder was transferred to a 50 ml tube and mixed with 20 ml M1 buffer (10 mM Tris-HCl (pH = 8.0),10 mM MgCl2, 0.1 mM PMSF, 1 M NaCl, 0.07% β-mercaptoethanol and 0.4 M Sucrose) using a chilled spoon. After agitated the tube in ice box for 10 min. The suspensions were filtered through Miracloth (Calbiochem) into a 50 ml tube and the supernatants containing cytoplasmic proteins were collected following centrifugation at 1000 x g for 20 min at 4°C. Remove the supernatant for the cytoplasm protein. Five ml of M2 buffer (10 mM Tris-HCl (pH = 8.0), 10 mM MgCl2, 0.1 mM PMSF, 1 M NaCl, 0.07% β-mercaptoethanol, 0.25 M Sucrose and 1% TritonX-100) was added to the pellet portion, re-suspended, and tubes re-centrifuged at 12000 x g for 10 min at 4°C. The supernatant was removed and the step was repeated 3 times. Finally, 300 μl Nuclei Lysis Buffer (P0013B, Beyotime Biotech) was added to the pellet and the suspension (nuclear proteins) was recovered.
After 7 days cultivation, conidia were collected and suspended in 100 ml of distilled water in a concentration of 1 × 105 spores /ml. 50 ml of the conidia were centrifuged at 3600 g for 10 min to extract the protein for equalization of protein amounts. The rest of 50 ml of conidia were divided into 200 μl (1 x 105 spores/ml) and placed onto the hydrophobic glass sheet at 28°C for 24 h. Suspensions harvested from the hydrophobic glass sheet were centrifuged at 3600 g for 10 min and the supernatants were recovered. For HLPC-MS/MS analysis, a 100 μg protein suspension was harvested. The suspension was mixed with 2.5 μg trypsin for digestion at 37°C for 4 h. 2.5 μg trypsin was added again and incubated for another 8 h. The peptides were then dechlorinated by Strata X and separated by a 65 min gradient elution at a flow rate 300 nl/min with the LC-20AD nano-HPLC system (Shimadzu, Japan), which was directly interfaced with Q-Exactive mass spectrometer (Thermo Fisher Scientific, USA). Mobile phase A consists of 0.1% formic acid and 2% acetonitrile, and mobile phase B consists of 0.1% formic acid and 98% acetonitrile. The mass spectrometer was operated in the DDA (data-dependent acquisition) mode and there was a single full-scan mass spectrum in the Orbitrap (350–1600 m/z, 70,000 resolution).
Results were presented as the mean ± standard deviation (SD) of at least three repeats. The significant differences between samples were statistically evaluated by using SDs and one-way analysis of variance (ANOVA) in SPSS 2.0. The data between two different treatments were then compared statistically by ANOVA, followed by the F-test, if the ANOVA result is significant at P< 0.01.
|
10.1371/journal.pntd.0003045 | Evidence for Circulation of the Rift Valley Fever Virus among Livestock in the Union of Comoros | Rift Valley fever virus (RVFV) is an arthropod-borne phlebovirus reported to be circulating in most parts of Africa. Since 2009, RVFV has been suspected of continuously circulating in the Union of Comoros. To estimate the incidence of RVFV antibody acquisition in the Comorian ruminant population, 191 young goats and cattle were selected in six distinct zones and sampled periodically from April 2010 to August 2011. We found an estimated incidence of RVFV antibody acquisition of 17.5% (95% confidence interval (CI): [8.9–26.1]) with a significant difference between islands (8.2% in Grande Comore, 72.3% in Moheli and 5.8% in Anjouan). Simultaneously, a longitudinal entomological survey was conducted and ruminant trade-related information was collected. No RVFV RNA was detected out of the 1,568 blood-sucking caught insects, including three potential vectors of RVFV mosquito species. Our trade survey suggests that there is a continuous flow of live animals from eastern Africa to the Union of Comoros and movements of ruminants between the three Comoro islands. Finally, a cross-sectional study was performed in August 2011 at the end of the follow-up. We found an estimated RVFV antibody prevalence of 19.3% (95% CI: [15.6%–23.0%]). Our findings suggest a complex RVFV epidemiological cycle in the Union of Comoros with probable inter-islands differences in RVFV circulation patterns. Moheli, and potentially Anjouan, appear to be acting as endemic reservoir of infection whereas RVFV persistence in Grande Comore could be correlated with trade in live animals with the eastern coast of Africa. More data are needed to estimate the real impact of the disease on human health and on the national economy.
| Rift Valley fever (RVF) is a viral disease transmitted by mosquitoes to ruminants. The disease may affect humans and has a great impact on the economy of the affected country. RVF occurs mostly in African countries, but epidemics have been reported in Madagascar and in the Arabian Peninsula. In the Union of Comoros, RVF virus (RVFV) has been suspected of continuously circulating since 2009 without any notifications of the typical clinical signs by the Comorian Animal Health Services. From April 2010 to August 2011, we conducted a livestock longitudinal survey in Grande Comore, Moheli and Anjouan. Our study aimed to detect RVFV-specific antibody acquisitions in cattle and goats. Simultaneously, a longitudinal entomological survey was conducted to describe the diversity of mosquitoes in the study zones and ruminant trade-related information was collected. Our investigations showed that Comoros ruminants acquired RVFV-specific antibodies all along the year and particularly in Moheli during the dry season. Our findings suggest a complex RVFV epidemiological cycle in the Union of Comoros with probable inter-islands differences in RVFV circulation patterns. The disease appears to be endemic in Moheli and potentially Anjouan, but the persistence of the disease in Grande Comore could be correlated with trade in live animals with the eastern coast of Africa.
| Rift Valley fever (RVF) is an arthropod-borne zoonotic disease caused by a RVF virus (RVFV), a member of the Phlebovirus genus of the family Bunyaviridae [1]. RVFV causes significant morbidity and mortality among sheep, goats, cattle and also affects humans. In livestock, abortion storms and high mortality observed among the younger animals cause significant economic losses [2], [3]. Humans are usually infected by contact with infectious animal tissues through inhalation or aerosols generated by slaughtering and necropsy [4]. Arthropod vectors play an important role during the onset of epidemic and inter-epidemic periods [5]. In endemic areas, RVFV is maintained in the environment through an enzootic vertebrate-arthropod cycle [6]. RVFV has been isolated from many vectors in the field [7], such as ticks and sand flies which are able to transmit the virus in experimental conditions [8], [9]. However, mosquitoes are the main insects involved in the spread of RVFV during epidemics. RVFV has been isolated from at least 40 species of mosquitoes belonging to 8 genera but only some of them are susceptible and able to transmit RVFV under laboratory conditions [10]. RVF is widely present in Africa and has been spreading to Madagascar and the Arabian Peninsula [11], [12]. In 2007, RVF outbreaks were reported in several eastern and southern African countries [13]. A few weeks later, and for the first time, RVFV was detected in the Comoros archipelago following the hospitalization of a young Grande Comorian boy showing symptoms of severe encephalitis [14]. In addition, during the 2008 and 2009 rainy seasons, outbreaks due to RVFV strains imported from mainland Africa were reported in Madagascar causing 59 confirmed human cases and seven deaths [11], [15].
In Mayotte, the French overseas territory that belongs to the Comoros archipelago, a retrospective study conducted in 2008 confirmed the presence of the disease with 10 human cases infected with RVFV strains genetically closely linked to the 2006–2007 Kenyan isolates [16]. It was also found that the Mayotte livestock has been infected by RVFV prior to 2004 [17]. Regarding the Union of Comoros, in 2011, Roger et al. reported widespread exposure of Comorian livestock with 32.8% of animals shown to be RVFV-seropositive without any notifications of massive abortions or abnormal mortality in the younger animals by the Comorian Animal Health Services. However, the origin of this infection remains unknown [18].
The Union of Comoros is located in the South West Indian Ocean at the northern end of the Mozambique Channel and is considered to be a gateway to islands in the Indian Ocean for various infectious agents imported from mainland Africa. Since 2002, live ruminants are imported from Tanzania and have entered the country without a period of quarantine or a clinical examination [18]. Finally, in the past, animal trade has already affected the country health status on several occasions, with regard to many diseases, like blackleg in 1970 and 1995, the contagious ecthyma in 1999, and the East Coast fever in 2003 and 2004 [19]–[21].
Some of the Culicidae species described in the Comoros archipelago [22] have already been shown to be involved in RVFV transmission. The establishment of RVFV in the Union of Comoros remains unconfirmed and the threat to the Comorian population and neighboring countries needs to be considered.
The trade and resulting movements of ruminants, the composition and abundance of the vector population and many other environmental and anthropological factors determine the nature of the RVF viral cycle. In order to elucidate how RVFV persists in the Union of Comoros, longitudinal and cross-sectional livestock surveys were conducted between April 2010 and August 2011 in six separate geographical zones. Mosquito populations were categorized in parallel over the same period via a longitudinal entomological survey. Additionally trade frequencies were analyzed, providing an estimate of regional ruminant flux and allowing for evaluation of the risk associated with animal importation, and the likelihood of RVFV persistence in the Comoros islands.
The research protocol was implemented with the approval of the Vice-Presidency of Agriculture, Fisheries and Environment of the Union of Comoros. No endangered or protected species were involved in the survey. Farmers in each zone gave their verbal consent to be included in the study. Permissions for the blood sample collection were obtained. The animals were bled without suffering. Regarding the trade survey, no personal data were collected, and only information concerning the number of animals travelling from one island to another was taken into account.
The Comoros islands form an archipelago of volcanic islands located off the southeastern coast of Africa, east of Mozambique and northwest of Madagascar. The archipelago is divided between the sovereign state of the Union of Comoros composed of three islands named Grande Comore, Moheli, and Anjouan, and the French overseas department of Mayotte. The tropical climate of the Comoros islands is characterized by daytime temperatures around 26°C at sea level, with limited variation during the year, and by annual heavy rainfall (2,679 mm) with two seasons: a humid season from November to April, and a dry season from May to October.
Based on the results of a previous RVFV antibody prevalence study in 2009 [18], six zones were selected in the Union of Comoros (Figure 1). Four zones were selected on the island of Grande Comore: zones 1 and 2 located in the center of the island where low RVFV antibody prevalence was found, and zones 3 and 4 located in the south with high RVFV antibody prevalence [18]. Zones 2 and 4 are located along the coast (0–200 m above sea level (asl.)) where ruminants are mostly goats stall reared or ranging free within and outside villages. Zones 1 and 3 are located at a moderate altitude (500–650 m asl.) where ruminants are mostly cattle reared in pastures (zone 1) or raised in stalls in the forest (zone 3).
Zone 5, which was located on the southern coast of the island of Moheli, was selected because of its highest RVFV antibody prevalence observed during the 2009 survey [18]. On this island, cattle are reared in stalls on an old coconut plantation. Finally, in zone 6 located close to the airport on Anjouan island, cattle were raised in stalls in vegetable production areas.
Five ml of whole blood was collected from the jugular vein of goats and cattle in Vacutainer tubes (Becton Dickinson, USA). Samples were allowed to clot at 15°C and serum was separated from whole blood by centrifugation; samples were stored in liquid nitrogen in the field and at −80°C in the laboratory.
The livestock longitudinal survey was conducted in the six separate zones detailed in Figure 1. From 20 to 30 ruminants were individually identified in each zone using ear tags. The number of animals sampled per zone was based on the previous survey, with a RVFV antibody prevalence ranging from 20% to 50% [18] with 70% relative precision [23]. To avoid colostral immunity, animals were selected as follows: cattle were between 10 months and one-year of age, and small ruminants were between three to eight months of age. Animals were sampled monthly from April 2010 to August 2010 and every four months from August 2010 to August 2011. The first series of serological tests determined the RVF serological status of each sampled animal. Only RVFV antibody negative animals were included in the livestock longitudinal survey and continued to be sampled until their IgG RVFV antibody positive status and were then excluded from the study. When possible, new ruminants were included in the study to substitute lost, dead or RVFV IgG positive animals.
The RVFV antibody prevalence based on the different study zones was estimated in August 2011. The sample size was based on the previously estimated prevalence [18] with a relative precision of 20% and a confidence level of 95% giving a required minimum of 385 animals to be collected [23]. Without any particular Comorian livestock census, animals were selected on the farmer's willingness to cooperate during the study.
Blood-sucking insects were sampled every four months from November 2010 to August 2011 along with the longitudinal serological survey using double-net goat baited traps placed from 4:00 pm to 10:00 am. The sampling was carried out for three consecutive days in the study zones numbered 1, 3, 5 and 6 (Figure 1). No sampling was performed in zones 2 and 4 for logistic reasons.
In order to generate hypotheses on potential associations between the estimated RVFV incidence and prevalence with environmental risk factors for RVF infection, we collated climatic variables [24]. Two remotely-sensed MODerate-resolution Imaging Spectroradiometer (MODIS) data sets were sourced from the National Aeronautics and Space Administration (http://modis.gsfc.nasa.gov/), namely the Daytime Land Surface Temperature (DLST) and the Nighttime Land Surface Temperature (NLST), both with spatial and temporal resolution of 1 km and 8 days. In addition, rainfall data were obtained from the Malaria Early Warning System (MEWS) program, freely available in the MEWS repository (http://iridl.ldeo.columbia.edu/expert/SOURCES/.NOAA/.NCEP/.CPC/.FEWS/.Africa/.TEN-DAY/.RFEv2/.est_prcp/), with a spatial and temporal resolution of 11 km and 10 days respectively. DLST, NLST and rainfall values were extracted within a 5-km radius buffer around each farm corresponding to the maximum daily distance for cattle (grazing and watering). For each sampled animal that became RVFV antibody positive, MODIS and MEWS data recorded at the time of the seroconversion in the zone concerned were compared with MODIS and MEWS data recorded at the same time in the other zones.
The aim of the trade survey was to estimate the movement of live animals between continental Africa and the Comoros archipelago and among the islands of the archipelago themselves. To date, only approximate figures are known without any quantitative data available [17]. The number of imported ruminants was collected monthly between November 2010 and August 2011 as follows i) the local veterinary authorities provided records of animal movements through the official ports of Moroni (Grande Comore), Fomboni (Moheli), and Mutsamudu (Anjouan), ii) one interviewer per island had the task of identifying undeclared animal arrivals on the coast, either in the field or from information provided by village chiefs.
All statistics were performed using R.3.0.1 [29]. For both Fisher's exact test and the Student-t test, a value of P<0.05 was considered significant.
A seroconversion was defined as an animal found with either a positive IgM ELISA result or a positive IgG ELISA result or both following a previous negative RVFV ELISA sample result.
A total of 191 ruminants (88 cattle and 103 goats) were included in the livestock longitudinal survey: 135 animals in Grande Comore, 27 in Moheli and 29 in Anjouan. Detection of RVFV antibodies (IgM and IgG) was performed by ELISA for a total of 849 serum samples over the duration of study.
Table 1 presents by date and per zone the number of animals that acquired RVFV antibodies over the duration of the livestock longitudinal survey. A total of 15 animals out of the 191 sampled acquired RVFV antibody during the study. Each of the 13 RVFV IgG ELISA positive samples were confirmed by VNT. Only one RVFV IgM ELISA positive sample was not confirmed by VNT (July 200, Moheli). This animal was confirmed RVFV IgG ELISA positive and VNT positive four months later in November 2010. Out of the 112 RVFV IgG ELISA negative samples randomly chosen, all were found negative by VNT. RVFV IgM antibodies acquisition was detected in three animals and RVFV IgG antibodies acquisition in 12 animals. Only one RVFV IgM ELISA positive animal in Moheli converted to RVFV IgG antibodies. The two others RVFV IgM ELISA positive ruminants were lost or slaughtered before the next sampling session (Table 1). Nine out of the 15, which acquired RVFV antibody, were recorded in Moheli, five in Grande Comore and one in Anjouan. Nine out of those fiftteen occurred during the dry season (six in Moheli, one in Anjouan, two in Grande Comore).
The overall annual incidence of RVFV antibody acquisition for the Union of Comoros was estimated at 17.54% (n[animal risk time] = 91), with a 95% confidence interval (CI) [8.95–26.14]) (Table 2). A significant difference was found when incidence of RVFV antibody acquisition was compared between zones (Fisher exact test, p<0.001) or between islands (Fisher exact test, p<0.001) (Table 2). Zone 5 (Moheli) incidence of RVFV antibody acquisition (72.3% 95% CI [0.255–1.000]) was significantly higher than in others zones (Table 2).
The statistical analysis did not reveal any significant difference in incidence of RVFV antibody acquisition between the rainy (from November to April,) and the dry season (from May to October) either for the Union of Comoros as a whole or per zone (Table 2). DLST, NLST (MODIS data) and cumulative rainfall (MEWS data) were similar in all six zones at the time of fourteen out of fifteen seroconversions occurred. There was one exception when the last RVFV seroconversion was recorded in Grande Comore in May 2011 (zone 3). Between March and May 2011, DLST, NLST and cumulative rainfall recorded in Grande Comore (29°C, 25°C and 730 mm respectively) were higher than those recorded in Moheli and Anjouan at the same time (DLST : 24°C, NLST : 22°C and cumulative rainfall : 420 mm). No RVFV seroconversion was recorded on Moheli and Anjouan during this period.
In August 2011, to determine the RVFV antibody prevalence, a total of 275 ruminant samples (i.e. 163 cattle and 112 goats) were tested for the presence of RVF IgG antibodies. A total of 37 ruminants (20 cattle and 17 goats) came from the longitudinal follow-up study and 238 ruminants (143 cattle and 95 goats) were randomly selected in the six separate study zones.
The overall RVFV antibody prevalence in the Union of Comoros study zones in 2011 was 27.6% (n = 275, 95% CI, [22.3–32.9]). We found a significant difference of RVFV antibody prevalence between islands (Fisher exact test, p = 0.007), with a higher RVFV antibody prevalence in Moheli (45.8%, 95% CI, [33.7–57.9], Table 3).
Twelve trapping days were conducted in each of the four zones under study (zones 1, 3, 5 and 6, see Figure 1). Blood-sucking insects were trapped in five out of the twelve trapping days in central Grande Comore (zone 1), in eight trapping days in southern Grande Comore (zone 3), eleven trapping days in Moheli (zone 5), and in seven trapping days in Anjouan (zone 6) (Table 4). Out of the 1,568 blood sucking insects caught with the double-net goat baited trap, 1,548 were identified as mosquitoes and 20 were identified as Stomoxys niger. A total of 1,133 insects were collected in Moheli (zone 5), 291 in Anjouan (zone 6), 108 in southern Grande Comore (zone 3) and 36 in central Grande Comore (zone 1). Although the number of comparisons was not large, the average number of trapped mosquitoes per trapping day per zone was significantly higher in Moheli (average was 113 insects) and Anjouan (average was 42 insects) compared to Grande Comore zones 1 (average was 7 insects) and zone 3 (average was 14 insects) (Table 5).
The diversity and number of blood-sucking insects caught with the double net goat baited trap per trapping day per zone are presented in Table 4. A total of seven genera and 16 species were caught of which 14 could be morphologically identified. Fifteen species out of the 16 caught were collected in Moheli (zone 5), nine species were collected in Anjouan (zone 6), three and eight in central and southern Grande Comore respectively (zone 1 and zone 3). Eighty-seven percent of the total number of insects caught belonged to three species with 52% belonging to two Eretmapodites species (E. quinquevittatus and E. subsimplicipes,) and 35% to Aedes cartroni.
No RVFV RNA was detected in any of the 442 pools tested.
The study highlighted movements of live ruminants between the three islands of the Union of Comoros, the African mainland, Mayotte and Madagascar (Figure 2). Data recorded by veterinarians and technicians showed movements of live ruminants from i) the east coast of Africa to Union of Comoros and ii) between the three islands of the Union of Comoros (Figure 2A.). Animals were observed being landed on beaches without any controls or in secondary “ports” like Chindini in the south of Grande Comore. Figure 2B represents the dynamics of live animal importations in Union of Comoros from May 2010 to July 2011. We recorded up to ten fold more ruminants imported in Grande Comore than in Moheli or Anjouan.
Rift Valley fever was detected for the first time in Grande Comore in the human population in 2007 [14] and in livestock in 2009 [18]. Our study demonstrates that RVFV is still circulating in the Union of Comoros despite of the absence of apparent clinical signs in livestock.
Fifteen RVFV seroconversions were observed in the Union of Comoros between 2010 and 2011 giving an overall incidence of RVFV antibody acquisition of 17.5%. These results suggest continuous circulation of RVFV on the three islands. However, significant differences in incidence were observed between islands (p<0.001). The incidence of RVFV antibody acquisition was higher in Moheli (72.3%) than in Anjouan (5.8%) and in Grande Comore (8.2%). This is in accordance with differences in RVFV antibody prevalence between the Union of Comoros islands recorded in 2009 and 2011. In 2011, RVFV antibody prevalence in Anjouan was still below the one in Grande Comore, whereas RVFV antibody prevalence remained the highest in Moheli. However, in Grande Comore and Anjouan RVFV antibody prevalence in 2011 appeared to have decreased whereas in Moheli, RVFV antibody prevalence remained similar to the level recorded in 2009, despite herd replacement estimated at 12% (L. Cavalerie, personal communication). These results suggest the existence of island specific RVF circulation patterns.
Seasonality of the incidence of RVFV antibody acquisition needs to be explored. The Comorian livestock farming characteristics (small herd size and small total number of ruminants) as well as the field issues did not allow a sufficient number of young ruminants (nrisk too small) reducing the power of the statistical analysis.
No clinical signs were reported in the Union of Comoros during the period of our study, as reported in Madagascar, Tanzania, and Mozambique in recent years [33]–[35], but the fifteen seroconversions observed suggest that RVFV could be circulating in the Comorian environment thanks to local mosquito-mammalian host cycles even if the numbers of caught mosquitoes were not large nor positive for RVF RNA. Out of the 1,568 blood-sucking insects caught, none were found to be RVFV RNA positive by PCR but in the absence of RVF outbreaks, chances of detecting RVFV in vector populations are known to be very low [36]. In 1978, Bruhnes described 30 mosquito species in the Union of Comoros [22]. Four of them: Ae. aegypti, Ae. fowleri, Ae. circumluteolus and Cx. quinquefasciatus are considered as RVFV potential vectors because the virus has been already isolated in these species in the field and because of their capacity to transmit RVFV under laboratory conditions [37]–[39]. All these species, except Ae. fowleri, have been caught at least on one island during our study, suggesting a role for this mosquito species to be involved in the transmission cycle on each of the islands. Five other mosquito species caught during our study, Er. quinquevittatus, An. arabiensis, M. uniformis, An. coustani and Ae. simpsoni were previously identified as RVFV RNA positive by PCR in the field [40]–[44]. An. coustani and Ae. simpsoni were found RVFV RNA carrier for the first time in the Indian Ocean area: respectively in Madagascar in 2011 and in Mayotte in 2009 [43], [44]. Thus, some of these mosquito species may play a role in RVFV transmission in the Union of Comoros. Geological inaccessibility, sampling design and climatic conditions likely explain the small number of specimens caught and the heterogeneity of entomological findings between islands [45]. These volcanic islands are characterized by a tropical climate with only slight variations in daily temperatures and abundant rainfalls, which theoretically should enable populations of Culicidae species to persist throughout the year. Nevertheless, each island has its own environmental characteristics, as the age of the three islands decreases westward: Moheli is 2.73±0.20 million years old, Anjouan, 1.18±0.03 million years old, and Grande Comore is 0.13±0.02 million years old [46].
On Moheli and Anjouan, the oldest islands, the landscape includes permanent rivers [47] and, as a result, many artificial and natural breeding mosquito sites exist. Moheli has a wide variety of natural and artificial sites in which mosquitoes can breed all year round [47], [48]. The presence of clay, resulting from the decomposition of volcanic soils, ensures the presence of abundant surface water impoundments. It allows the cultivation of irrigated rice hence and favors the development of diversified mosquito populations [47]. A greater number of mosquito species were caught in Moheli (15 species) than in the other islands which is in agreement with Brunhes' inventory in 1978, including two mosquito species known as RVFV potential vectors. Thus, favorable conditions for RVFV persistence being present a better chance for a possible RVFV cycle involving vectors and animals is suggested.
The abundance of mosquitoes trapped in Anjouan (zone 6) was similar to that in Moheli (zone 5) and three mosquito species known as RVFV potential vector have been caught during our study. However, RVFV antibody prevalence in Anjouan was the lowest and appeared to be decreasing. Moreover in 2011, only one ruminant exhibited a RVFV seroconversion. Anjouan shares some similar environmental characteristics with Moheli that could allow mosquitoes to survive all year round but Anjouan has some characteristics that could limit the circulation of RVFV. For example, the landscape is comprised of hill slopes and irrigated field rice is not cultivated on the island. Ruminants are mainly reared in stalls in the highlands in the eastern part of the island. For that reason, the probability of contact between infected vectors and ruminants may be lower in Anjouan than in Moheli and the maintenance of a vector-ruminant cycle may be harder to get. More investigations in other cattle-rearing areas are thus needed to conclude on RVF circulation in Anjouan.
Incidence of RVFV antibody acquisition and the RVFV antibody prevalence in Grande Comore are hard to explain based only on entomological parameters. Presence of steep slopes with decomposed and highly permeable soils characterize Grande Comore, the youngest island of the country [47]. Surface water is rare and only artificial containers (such as tanks and troughs) and some natural breeding sites (such as coconut shells and hollow trees) enable the development of Culicidae. Our results were in accordance with these observations as fewer blood-sucking insects were caught in Grande Comore when compared to Moheli and Anjouan. Thus, mosquito abundance in Grande Comore was likely correlated with the number of breeding sites that appeared after rainy episodes, as observed for the seroconversions we detected in Grande Comore following on from a major increase in cumulative rainfall. Two out of eight mosquito species caught during our study have been described as RVFV potential vectors. Consequently, environmental conditions for a local mosquito-mammalian host cycle could be met after important rainy episodes but a continuous circulation of RVFV in Grande Comore all year round is less likely to happen. However, regular introductions of the virus through the arrival of live animals from Tanzania [49], Anjouan, and Moheli may play a role in the persistence of RVFV in Grande Comore.
Analysis of trade in live animals confirmed observations reported by Cêtre et al., in 2012 in an overview of the movement of live ruminants between east Africa and the Comoros archipelago, as well as within the archipelago. Per year, more than 3000 live ruminants are imported from Tanzania (Chief Veterinary Officer of Comorian Vet services, personal communication), where RVF is endemic [34]. These animals enter the Union, mostly Grande Comore, without any quarantine or clinical examination. The risk of the introduction of new exotic strains of RVFV is consequently quite high and could affect the country in the same way as many other diseases in the past [19]. Tanzanian ruminants are imported for “great weddings” which are usually celebrated in July and August in Grande Comore. During these traditional weddings, villagers sacrifice ruminants without any particular sanitary rules. However, no major cases in humans and no ruminant seroconversions were reported during the “great weddings” period during our study but to date, human and veterinary health surveillance networks remain not very efficient. Occasional imports of Tanzanian ruminants into Moheli and Anjouan have also been reported; so new RVFV strains could have been also introduced on these islands. The regular introduction of live ruminants from Anjouan and Moheli could also contribute to the regular introduction of RVFV in Grande Comore as well.
Rift Valley fever epidemiology in the Union of Comoros is complex and further virological investigations should help to explain the origin of the RVFV strain(s) circulating within the islands. However, based on the results of the present study, RVFV seems hardly to persist on Grande Comore through a local vector cycle only but repeated reintroduction of viruses is possible. The situation regarding Rift Valley fever in Anjouan and Moheli appeared to look like that in Mayotte, Madagascar, Tanzania, and Mozambique [33]–[35], [50] where RVFV seroconversions have also been observed in the dry season without any apparent clinical signs. These findings could identify Moheli and Anjouan as endemic areas for RVFV. Given the incidence of RVFV seroconversions and antibody prevalence, RVFV is more likely to be circulating in Moheli than in Anjouan. However, additional data are needed to firmly conclude on the circulation of RVFV in the Union of Comoros. Wildlife such as bats and lemur species in our zone should be investigated even though no wildlife reservoir has been identified in any other country so far [51], [52].
Rift Valley fever is still a burden for the Union of Comoros as new human cases were diagnosed as RVFV positive in 2011 and in 2012 either by IgM or RVFV RNA detection with clinical signs [53], [54]. The real impact of the disease on human health and on the national economy is still unknown. Human and veterinary health networks need to be strengthened including the establishment of quarantine for imported ruminants.
|
10.1371/journal.pcbi.1003843 | Fast Synchronization of Ultradian Oscillators Controlled by Delta-Notch Signaling with Cis-Inhibition | While it is known that a large fraction of vertebrate genes are under the control of a gene regulatory network (GRN) forming a clock with circadian periodicity, shorter period oscillatory genes like the Hairy-enhancer-of split (Hes) genes are discussed mostly in connection with the embryonic process of somitogenesis. They form the core of the somitogenesis-clock, which orchestrates the periodic separation of somites from the presomitic mesoderm (PSM). The formation of sharp boundaries between the blocks of many cells works only when the oscillators in the cells forming the boundary are synchronized. It has been shown experimentally that Delta-Notch (D/N) signaling is responsible for this synchronization. This process has to happen rather fast as a cell experiences at most five oscillations from its ‘birth’ to its incorporation into a somite. Computer simulations describing synchronized oscillators with classical modes of D/N-interaction have difficulties to achieve synchronization in an appropriate time. One approach to solving this problem of modeling fast synchronization in the PSM was the consideration of cell movements. Here we show that fast synchronization of Hes-type oscillators can be achieved without cell movements by including D/N cis-inhibition, wherein the mutual interaction of DELTA and NOTCH in the same cell leads to a titration of ligand against receptor so that only one sort of molecule prevails. Consequently, the symmetry between sender and receiver is partially broken and one cell becomes preferentially sender or receiver at a given moment, which leads to faster entrainment of oscillators. Although not yet confirmed by experiment, the proposed mechanism of enhanced synchronization of mesenchymal cells in the PSM would be a new distinct developmental mechanism employing D/N cis-inhibition. Consequently, the way in which Delta-Notch signaling was modeled so far should be carefully reconsidered.
| During vertebrate embryonic development, the segmented structure of the axial skeleton is laid down by the process of somitogenesis. Periodically, blocks of cells separate at the anterior end of a mesenchymal tissue (PSM) on either side of the neural tube and develop later into spinal vertebrae. Cellular oscillators operating in each cell of the PSM control this process. Their synchronization is essential, and is effected by direct cell-to-cell signaling of the Delta/Notch (D/N) pathway. To better understand the regulation of the genes involved, we employ computer modeling. In this case, the fast synchronization of the oscillators represents a special challenging and worked so far only by the integration of cell movements. Now, we have succeeded in accelerating the synchronization for the first time without cell movements by the interposition of the novel mechanism of intracellular reciprocal inhibition termed D/N cis-inhibition into our computer simulations.
| Adaption to the day-and-night-cycle on earth provides an evolutionary advantage for organisms that can adjust their gene activity to this 24-hour rhythm. Therefore many metabolic processes show a circadian periodicity because they are all controlled by a GRN forming the so-called circadian clock [1]. Shorter period oscillators are called ultradian [2]. Some play an important role in the embryonic process of somitogenesis, where the vertebrae-precursors, the somites, bud off with a species-specific periodicity at the anterior end from a mesenchymal tissue on both sides of the notochord, the presomitic mesoderm. For mice this period is with two hours much shorter than circadian.
The core of the somitogenesis clock, first simulated in a computer model by Meinhardt [3], is set up in probably all vertebrate species by the Hes/Hairy/her gene families [4], which are negative feedback oscillators. A short decay time for the gene products together with a long enough time delay between gene expression and binding of the protein on its own gene promoter results in oscillatory gene expression. In mice the Hes1, Hes5 and Hes7 genes (and many others connected to them in an intricate network) were found to oscillate in the PSM [5]. Hes1, which also oscillates in neural progenitors [6], could be stimulated to oscillate with a two-hour period in vitro in fibroblasts, neuroblasts, myoblasts and other cell types [7]. In the anterior unsegmented PSM of mice, also called wave zone, Hes7 needs additional activation by D/N signaling to maintain oscillatory gene expression [8]. The D/N pathway works by juxtacrine signaling: Membrane-anchored DELTA or JAGGED ligands of a signal-sending cell bind to NOTCH receptors embedded in the cell membrane of an adjacent cell. This induces a proteolytic cleavage of the NOTCH receptor and releases the intracellular domain of NOTCH (NICD) into the cytoplasm, which then moves into the nucleus where it serves together with various co-factors as transcription regulator and activates, among others, the Hes1/7 genes [9]. These events finally lead to a moving wave of NICD from posterior to anterior in the PSM. We proposed in our 2012 model that this wave is generated by the action of the posterior-to-anterior gradients of FGF8 and WNT3a on decay rates of the core oscillator consisting of D/N and Hes7 [10]. When the NICD wave comes to a halt in the anterior PSM, NICD determines together with TBX6 the expression of Mesp2 that induces the formation of a border between a forming somite and the remaining PSM [11].
Another important function of D/N signaling in somitogenesis is synchronization of the cellular oscillators in the PSM [12], [13], which requires cell-cell contact [14]. Without this synchronization somite formation is severely disturbed [14]. The synchronization of cellular oscillators was also examined theoretically, mostly for the zebrafish her1/7 system. Using delay differential equations, D/N signaling was able to synchronize two cells [15] or a row of cells [16]. However, if this system is expanded to 2-dimensional arrays of cells the short-range interaction of D/N causes different domains to be synchronized to different phases and no domain is able to conquer the others [17], [18]. It was shown for zebrafish and chicken that cell movements in the posterior part of the PSM occur depending on the concentration of FGF8 [19], [20]. Uriu et al. included these movements into simulations of the zebrafish PSM and could thereby demonstrate a much better synchronization of the her oscillators [18]. Later, this theory was extended to find an optimal rate for cell movements and to describe the effect of gradual recovery of intercellular coupling experienced by a cell after movement [21].
All these models assumed direct interaction between DLL1 and NOTCH1 when they are positioned in membranes of adjacent cells. However, Delta-ligand and Notch-receptor molecules can also interact within the endoplasmatic reticulum (ER) or cell membrane of the same cell [22], [23], which would lead to a fast clearance of the intracellular dimer. This mechanism, where Delta and Notch inhibit each other in the same cell, was therefore termed D/N-cis-inhibition. For example, D/N cis-inhibition is able to generate mutually exclusive signaling states in a mammalian cell-culture system [24]. Applied to computer simulations, D/N cis-inhibition leads to sharper and faster boundary formation during development of the Drosophila wing vein system and improves the equidistant distribution of bristle precursor cells by lateral inhibition [25].
Here, we propose another beneficial effect of D/N cis-inhibition: It accelerates in computer simulations the synchronization of D/N coupled ultradian oscillators and extends the parameter range wherein synchronization is possible without taking cell movements into account. Although experimentally not yet confirmed, the proposed mechanism of enhanced synchronization of mesenchymal cells in the PSM would be the third distinct developmental mechanism employing D/N cis-inhibition. Consequently, the way in which Delta-Notch signaling was modeled so far should be carefully reconsidered.
We employ the same cell- and gene-based simulation tool as described in [10]. The GRN incorporated in each virtual cell consisting of Hes7, Delta1, Notch1, is shown in Fig. 1 A and in an advanced version including also Lfng in Fig. S1. Oscillations are generated by a negative feedback of HES7 onto the Hes7 promoter with delay, which is simulated by the transport of proteins and mRNAs between the nucleus and cytoplasm similar to the transport-model by Uriu et al. [18]. Furthermore, the Hes7 oscillators are coupled by D/N signaling and we assume that HES7 acts on the Dll1 promoter as it was shown for HES1 [26]. The DLL1 ligand and NOTCH1 receptor are modeled with two compartments for the proteins (cytoplasm and membrane) and for their mRNAs (cytoplasm and nucleus): Since we assume that Notch1 expression does not oscillate we do not differentiate between nucleus and cytoplasm in this case, because a mathematical description without delay for the mRNA is sufficient. Our model is designed for the simulation of mouse development, therefore the reaction rates are taken from literature or if not available adjusted to reproduce a mouse specific oscillation period of around 2–3 hours. However, our program allows other oscillation periods by simply rescaling all reaction rates in the differential equations – except those in the denominators – via its graphical user interface, which is equivalent to a rescaling of time. In addition to the reaction of DLL1 and NOTCH1 between neighboring cells leading to the release of NICD as transcription co-factor (trans-activation), this work also considers the reaction of NOTCH1 and DLL1 in the membrane and cytoplasm of the same cell (cis-interaction), which leads to their immediate decay – shown graphically in Fig. 1 B. So, the titration of one membrane protein against the other in each cell leads to an excess of either the ligand or the receptor and consequently determines whether the cell acts as a sender or receiver.
Contrary to our previous work [10], where every cell started with the same initial concentration values and received after mitosis the concentration values of its mother cell at their respective oscillation phases, here, all cells start with random initial values. To avoid that the cells start too far away from their limit cycle we add random values between zero and one multiplied to each of the initial concentration values used in [10] and scaled with a percentage value that gives a simple measure for the initial noise. For instance, 200% noise means: to each concentration its doubled value is added multiplied by a random number taken from the interval between zero and one.
Our program allows for real time observation of the simulation, so synchronization can be easily observed by visual inspection. However, to get a quantitative measure for synchronization we introduced a simple correlation function that falls to zero when perfect synchronization is achieved and shows oscillatory behavior otherwise. In the case of anti-synchronization, the values of the correlation function display negative oscillations.Here stands for any concentration value of a gene product in cell k (or i or j) at time t and is the average concentration value. For each cell with index i its concentration is multiplied with the average concentration of its neighboring cells with index j, where N is the number of neighboring cells. A rectangular arrangement of cells results in N = 2 for 1 dimension, N = 4 for 2 dimensions, and N = 6 for 3 dimensions. For cells situated on an edge or corner the number of neighbors is reduced, i.e. we use not periodic boundary conditions in our simulations. So N in the formula above depends on cell index i, but we suppress this dependence to simplify the notation. Furthermore, the user can define an extended neighborhood, which means that e.g. in 2 dimensions the diagonal adjacent cells are counted as neighbors. If all cells are synchronized, ci(t) and cj(t) have the same value, which is equal to the average value. So, the difference in the first formula will become zero. For the evaluation of the correlation function we used Hes7 mRNA concentration in the cytoplasm if not stated otherwise.
Although the correlation function uses only information about neighboring cells, it shows us synchronization by dropping to zero, because if each cell is synchronous to its neighbor, all cells are synchronized. Compared to the R-synchronization measure (defined in the supplementary material Text S1), which goes to one for perfect synchronization, the advantage of the correlation function C(t) is the observation that it becomes negative, if the configuration becomes anti-synchronized, i.e. one observes a salt-and-pepper pattern, which can be oscillating or not. See Fig. 2 first and last row for an example for each case.
In our search for parameter values resulting in fast synchronization we observed in the case without D/N cis-inhibition that parameters that allowed for fast synchronization made the system unstable against anti-synchronization. After a period of almost perfect synchronization with C(t) almost exactly zero the system drifts slowly into an oscillating salt-and-pepper pattern with the difference between neighboring cells becoming ever larger. Unfortunately, the faster the synchronization, the shorter the duration of synchronized behavior before reverting into the anti-synchronized state. Because the correlation function allows us to see this behavior before it becomes visible by eye, it is very useful for interactively searching for parameters providing for fast synchronization.
The effect of D/N cis-inhibition on synchronization of a 7×7×7 cell cube with 100% noise added is shown in Fig. 2, where simulation snapshots are displayed for increasing strengths of D/N cis-inhibition. Clearly, D/N cis-inhibition accelerates synchronization, whereas without (see movie S1) or small cis-inhibition the oscillator-system synchronizes badly and turns after some time into an anti-synchronized state, which was already described for a 2-cell [15] and a 2-dimensional system [17] (see also supplementary movie S2 for the case of rDNcis = 0.01). For intermediate (0.005) values of D/N cis-inhibition one observes incomplete synchronization. Large parts of the cube are synchronous but in different phases to each other so that ‘waves’ of expression moving over the cube volume can be observed. Increasing the D/N cis-inhibition strength leads to complete and ever faster synchronization with the best result achieved for 0.0115. However, increasing D/N cis-inhibition further leads to a progressive damping of the oscillations. This non-oscillating state then turns slowly into a static salt-and-pepper pattern. So in this case we get the classic lateral inhibition case without oscillations.
Simulation snapshots and the time course of our correlation function for systems with different dimensions are shown in Fig. 3. Compared to the 3-dimensional simulation with a 7×7×7 cube of interacting cells, the synchronization of a 2-dimensional array of cells is slower and deviations from perfect synchronization are larger. Only if one reduces the noise amplitude to 60%, the initial deviations in the correlation function are comparable, but synchronization is still slower. A similar effect is observed for a 1-dimensional chain of cells. This can be explained by the nature of our model, where the effect of D/N signaling in the receiving cell is averaged over the number of its neighbors due to practical reasons. This has the advantage that one does not have to change all parameters in the network when dimensionality of the system is changed. Consequently, the noise one cell receives in D/N signaling reduces with the number of its neighbors because fluctuations are cancelled out better in summation with more neighboring cells sending noisy signals. This effect is also demonstrated in Fig. S2, where a 3-dimensional array with 6 neighbors per cell gives comparable results to a 2 dimensional array with 8 neighbors per cell. Likewise, we analyzed the influence of cell number, i.e. the volume of a cell array, on synchronization and compared cubes with a length of 5, 7, 9, 11, and 14 cells (Fig. S3). While at the beginning the correlation functions vary due to the randomly chosen initial values, they decay in the further course of the simulation to very small values with a similar behavior. The same behavior can be observed also for the R-synchronization-measure, which quickly reaches values very near 1, indicating very good synchronization, independently of the size of the cell cube.
To explore the robustness of the system and the speed of D/N-mediated synchronization, with and without cis-inhibition, we performed an extensive scan over all parameters in a simple two-cell system. As expected, D/N cis-inhibition provides for faster synchronization of cells over a wide parameter range, independent from the chosen initial concentration values (for details see supplemental text S1). There are also parameter ranges where synchronization is not achieved with D/N cis-inhibition, if one looks at the R-synchronization measure. However, if one looks at the concentration time course behavior one sees that this downward trend of the R-function results from a progressive damping of the oscillations if one increases the Hes7 mRNA or protein decay rates more than ten percent, for instance.
The influence of the system parameters on the amplitude (minimal and maximal cytoplasmic HES7 expression) of the cellular oscillator is shown in Fig. 3 of the supplemental text S1. One can observe the strong dependence of the oscillator amplitude on Hes7 mRNA and protein decay rate, for instance, and that the cis-inhibition strength rDNcis abolishes the oscillation if it increases beyond 0.014, as already seen in Fig. 2.
To examine the robustness of the system further we generated 40 parameter sets by randomly varying all production, transport, and decay rates within a range of plus-minus ten percent around our standard parameters and tested these parameters sets in a cube with an edge length of 7 cells with 100% initial noise added. 16 of the random parameter sets resulted in damped oscillation and of the 24 undamped oscillating systems 21 showed complete synchronization. Only for three parameter sets synchronization was not complete. Instead, expression waves were generated. Results for all oscillating parameter sets are shown Fig. S4. The input files for running simulations with the different parameter sets are supplied in the supplemental material as file S1 (Config-files.tar.gz).
In our previous work on boundary formation in the PSM of mouse [10] we postulated a positive action of LFNG on D/N signaling. Likewise, we have extended our minimal model by Lfng, which is controlled by HES7 (Fig. S1). Here, the parameters chosen for the relative contributions of unaided D/N signaling and D/N interaction with LFNG-action have to meet two demands: (i) they should allow fast synchronization with D/N cis-inhibition, and (ii) they should reproduce the diminished oscillation amplitude observed experimentally in the mouse PSM when Lfng is non-functional [27]. These demands are fulfilled when we set the ratio of unaided to LFNG-promoted D/N reaction to about 1∶4 (Fig. S5).
So far, all discussions on synchronization of ultradian oscillators by D/N signaling examined the static case, i.e. a non-growing tissue. However, a real test for synchronization would be a growing tissue, for example, the tail bud during somitogenesis (Fig. 1 C). Therefore, we implemented D/N cis-inhibition in one of our models of somitogenesis, which is characterized by a growing tissue and a posterior-to-anterior FGF8 gradient controlling HES7 degradation [10]. When daughter cells inherit the concentration values of their mother cells and a 100 percent noise is added, we observed a clear difference between simulations without (movie S3) and with (movie S4) D/N cis-inhibition (Fig. 4). However, even with cis-inhibition instabilities have arisen after the fourth oscillation. To allow for more realistic noise-affected gene expression, we simulated mitosis by developing a model in which the dividing cells in the growth zone of the PSM shut off transcription, which consequently disturbs Hes7 expression waves after two oscillations even when the cells started synchronized at the beginning of the simulation. Furthermore, we allowed diagonal neighbors to signal via D/N. For a mitosis phase of 20 min, D/N cis-inhibition was able to maintain phase coherence reasonably well (movie S5), whereas without D/N cis-inhibition (movie S6) the initial order was lost after two oscillation periods (Fig. 4).
In summary, our results demonstrate that the inclusion of D/N cis-inhibition in the formulation of the model brings about a decisive improvement in the ability of D/N signaling to synchronize cellular oscillators. This is achieved not only for a specially chosen set of parameters, but a wide range of model parameters.
The aim of our modeling work in somitogenesis is to explain how the various expression waves in the mouse PSM are generated, why they slow down when they are nearing the anterior end of the unsegmented PSM, and how the boundary between the PSM and the next forming somite is formed. In our previous paper [10] we were concerned with the generation of the NICD wave and why it stops, because together with the TBX6 and FGF8 gradients NICD induces Mesp2, which is critically important for boundary formation. Our hypothesis for the generation of the NICD wave was that the WNT3A and/or FGF8 gradients in the PSM influence an intracellular process of the core oscillator consisting of Hes7 and D/N thereby slowing the oscillator down when it gets out of the range of the gradients. Therefore, we modeled the core oscillator as a transport model with the most important cellular compartments (nucleus, cytoplasm, and membrane) and processes like transcription, translation and transport and allowed a possible coupling of each gradient to each cellular process. Furthermore, we included as many measurable parameters and especially promoter information as we could find in the literature (which is unfortunately rather sparse). However, with plausible assumption one can generate at least the qualitative behavior with its characteristic expression pattern rather well. The drawback of our method is that one cannot sample the multidimensional parameter space. However, if new information becomes available, one can feed it directly into our model.
In our 2012 paper [10] we had excluded the synchronization problem. Cells started synchronized and stayed so, because during proliferation daughter cells inherited the oscillatory phase of their mother cells. However, as NICD and D/N-signaling are widely held to be responsible for the maintenance of oscillations and synchronization of wave formation and in creating boundaries in space as the waves come to rest, one should work towards a comprehensive model including synchronization.
In somitogenesis the formation of sharp boundaries between the block of cells forming a new pair of somites and the remaining PSM works only when gene expression in the cells forming the boundary is synchronized. It has been shown experimentally that D/N signaling is responsible for this synchronization. The species-specific periodicity of somitogenesis is controlled by cellular oscillators, in mouse most probably by the negative feedback oscillator Hes7. The synchronization has to happen rather fast as a cell experiences about five oscillations from its birth to its incorporation into a somite [28]. Computer simulations describing oscillators coupled by classical modes of D/N-interaction failed so far to achieve synchronization in an appropriate time approach except by introducing cell movements in simulations. Here we show that fast synchronization of Hes-type oscillators can be achieved without cell movements by including the process of D/N cis-inhibition.
While in conventional models of D/N synchronized oscillations each cell is sender as well as receiver of D/N-signaling because DELTA ligands as well as NOTCH receptors are active in the membrane of the cell, in a system with perfect cis-inhibition i.e. perfect titration of DELTA against NOTCH or vice versa, a cell is either sender or receiver. That means that a cell with DELTA excess – an information sender - can enforce a change in NICD controlled gene expression in a neighboring receiver cell, i.e. with NOTCH excess, as fast as intrinsic NICD processes allow in the receiver cell. If Delta expression is oscillatory – as in our model - the sender cell could go into receiver mode if Delta expression is low. So other cells could influence/synchronize this cell. In this manner, fast synchronization could be achieved despite the fact that the cell-interaction is still local (even if one considers communication by cytonemes as observed in zebrafish [29]). This does not exclude the possibility that for very large volumes the locality of cell-cell-communication leads to domains synchronized to different phases, but for realistic numbers of cells the above acceleration of synchronization could be sufficient i.e. fast enough.
However, for D/N synchronization of Hes7-oscillators the considerations shown above are too simplified, as a cell cannot be only sender, i.e. have any active NOTCH in its membrane. This is so because Hes7 activation relies on NICD and in our model of the core oscillator HES7 suppresses Dll1 expression leading to the oscillatory DELTA expression mentioned above. Consequently, a sender-only cell would have no interesting message to send. So a perfect titration of NOTCH against DELTA is not desirable. There has to be an optimum value of cis-inhibition. If this value is surpassed oscillations are damped and die out. This was shown in Fig. 2.
At least for mouse, there is strong evidence that the Hes7 gene oscillates by negative feedback of its protein on its own promoter, thereby serving as the core oscillator of the somitogenesis clock [30]–[32]. Furthermore, promoter analysis revealed that Hes7 is induced by D/N signaling [33]. The NOTCH modifying gene Lfng is also induced by D/N-signaling and oscillates in the PSM because its expression is inhibited by HES7 [33].
The fact that D/N-signaling is required for the synchronization of ultradian oscillators in the PSM was shown for zebrafish [16], [34] in experiments with single cell resolution. Because it is not easy to separate the induction of oscillation and synchronization in mouse on the cellular level, Okubo et al. used chimeric embryos composed of wild-type and Dll1-null cells to demonstrate that D/N-signaling is responsible for the synchronization of oscillations in the PSM also in mouse [13]. To clarify the role of Lfng in the somitogenesis clock, Okubo et al. also analyzed Lfng chimeric embryos and used Notch signal reporter assays in a co-culture system [13]. As interpretation of the results they proposed a novel, in this form not yet described action of LFNG on DLL1. The knockout of Lfng resulted in an enhanced activity of NICD in the PSM, which indicates that LFNG might affect NOTCH1 and DLL1 negatively. Okubo et al. also demonstrated that the synchronization of cellular oscillators was proportional to the number of Dll1 expressing (wild-type) cells in chimeric embryos, which confirmed that D/N synchronizes Hes7 oscillations in the PSM. Similarly, using Lfng chimeric embryos, they showed that LFNG seems to be required for this synchronization. Interestingly, computer simulations that integrated the proposed effect of LFNG on NOTCH1 and DLL1 showed fast oscillator synchronization and were able to reproduce their experimental findings [13]. In their model the Hes7 oscillator in every cell is coupled to neighboring cells via LFNG, which is itself driven by HES7 oscillations and regulates the intracellular coupling by inhibition of both NOTCH1 and DLL1 activity in the same cell. Thus, LFNG not only represses D/N signaling inside the LFNG expressing cell by modifying NOTCH1 cell-autonomously, but also represses D/N signaling between neighboring cells by also modifying the DLL1 ligand. In short, in their model the output of the Hes7 oscillator is coupled to D/N signaling exclusively by the way of LFNG action.
In contrast, in our model we assume that HES7 inhibits Dll1 expression like Her1/7 inhibits deltaC in zebrafish. We will not repeat the extensive discussion provided in our previous publication [10], but strengthen the main arguments, which are that expression of Dll1 is dynamic in the PSM [35] and that only the orthologs of Hes7 and Dll1 are dynamic in the PSM of all vertebrate systems examined so far [36]. For example, Lfng expression is constant in zebrafish as well as in medaka [36]. Therefore, we argue for an evolutionary mechanism with a zebrafish-like core oscillator in which LFNG acts only in a modulatory role. Consistent with this notion, NICD expression is still dynamic in Lfng deficient mice [37] and Lfng is not required for somite formation in the tail bud phase [38]. In this work, we therefore examined the effect of D/N cis-inhibition primarily in a model without modulation of D/N signaling by LFNG.
Quantitative data regarding cell cycle parameters in mouse embryogenesis are sparse. Power and Tam give a value of ca. 30 min for 7.0-day embryos [39].
When judging about the success or failure of our model with respect to the real facts one should not forget that there may be biological mechanism that are not covered by the model, but could be crucial for the functioning of the synchronization. For example, it was found that Dll1 mRNA is stabilized during mitosis, by Elavl1/HuR in neuroepithelial cells [40]. If similar mechanisms are operative in the growth zone of the PSM, our assumption that mRNA decay rates are constant in time could be too pessimistic. A smaller decay rate during mitosis would very probably diminish the perturbation to oscillations and thereby improve synchronization. Interestingly, a study observing oscillatory expression of a Her1-Venus reporter at single cell resolution in the zebrafish PSM found that her1 oscillations are linked to mitosis [34]. Therefore, it is possible that cell divisions introduce less noise than our model assumes.
In the hypothalamus of the mammalian brain, 20000 nerve cells function as circadian oscillators and have to be synchronized to function as the master circadian clock of the body [41]. Like ultradian oscillators, these circadian oscillators function by a negative transcription-translation feedback loop and are often also modeled by Goodwin-models (see for example [42], [43] and references therein), but also by delay differential equations or very simple toy models [44]. However, compared to the somitogenesis clock, in the circadian clock there are more interlocking feedback loops [41] and the communication between cells works either by secretion of neuropeptides and/or by direct innervation. So, coupling in the circadian clock is not mediated by communication between directly adjacent cells but by non-local interactions, which probably favors tissue-wide synchronization and prevents the phenomenon of cell territories synchronized to different phases ‘fighting’ for dominance. Furthermore, in circadian clock models the synchronization signal acts positively on the transcription of the clock genes. This is also the case in our model of the ultradian oscillator, where NICD acts as an activator on Hes7 transcription. However, HES7 represses Dll1 in the same cell and therefore NICD generation in the adjacent cell. This is the reason why lateral inhibition occurs in the static case or leads to anti-synchrony in the dynamic setting. Another difference concerns the coupling of the synchronization signal to the promoter of the clock feedback loop. In circadian models, this is mostly assumed to be additive, whereas we do not assume an additive but a multiplicative coupling of D/N signaling to the Hes7 promoter because it was shown that in most of the PSM Hes7 ceases to oscillate without D/N input. We disregard in our model the fact, that Hes7 is induced by FGF8 in the tailbud [8], which would be an additive coupling to FGF8. It was found in circadian oscillator models that weak oscillators, which are damped without a synchronization signal, synchronize faster [42], [43], [45]. As our Hes7 oscillator is coupled in ‘AND’ modus to the synchronization signal (NICD), this could possibly be seen as an example of this principle. (It was also found for the circadian clock that the oscillator's radial relaxation time scale and the ratio of synchronization signal to the oscillator amplitude are important for synchronization and oscillator entrainment [44].)
Contrary to Wang et al. [46] who simulate neural fate decisions in the developing nervous system and proposed that D/N cis-inhibition causes asynchrony between adjacent cells, adding D/N cis-inhibition terms to our model of ultradian oscillators of the Hes/Hairy/her type clearly leads to a faster synchronization. Furthermore, the phenomenon of different regions that are synchronized to different oscillation-phase values, and that one region cannot overwhelm another, can be overcome without cell movements, at least for the non-growing case, by introducing D/N cis-inhibition. Since cis-inhibition allows faster reaction of cells on changes in their neighborhood, cell movement may not be required for all situations in which synchronization is mediated by D/N signaling. We also show that D/N cis-inhibition does not interfere with a proposed mechanism for wave generation in the PSM, in which the control of HES7 degradation by the posterior-to-anterior FGF8 gradient slows down the oscillators as they get out of the range of the gradients by the continuous growth of the PSM. That D/N cis-inhibition does not lead to complete synchronization in the whole PSM, which would resist slowing down, is probably caused by the fact that the slowing down gets appreciable only in the last oscillation a cellular oscillator experiences before being incorporated into a somite [10]. However, ultimately, only experiments can clarify whether D/N cis-inhibition [22], [23] is functional also during somitogenesis.
Download information including a mini manual of the program is provided in supplemental Text S2.
We also supply SBML files describing the system for 2 cells without growth (Text S3 (SBML_DeltaNotchModel_2cells_cis.xml) for the model described in Fig. 1 and Text S4 (SBML_DeltaNotchModel_2cells_cis_lfng.xml) for the model described in Fig. S1).
To model gene expressions we use essentially the same methodology as described in [17], i.e. a gene- and cell-based simulation program that numerically solves differential equations describing a gene regulatory network and displays the concentration of a selected gene product by color intensity (virtual in situ staining) in each cell. For showing the consequences of the gene regulatory network (Fig. 1) we use the same cell- and gene based simulation program as in [10] except that cis-inhibitory interaction-terms in the membrane and cytoplasmic compartment were added. Specifically, we use the same formulas and rate constants as in our previous publication, except the addition of the D/N cis-inhibition terms, different values for Hill coefficient and Hill threshold describing the action of NICD at the Hes7 promoter, and the LFNG coupling. Furthermore, it is now possible to enlarge the neighborhood of a cell so that also diagonally adjacent cells are treated as interacting neighboring cells. In addition, we take into account that the Hill-coefficient for the action of the NICD complex on the Hes7 promoter could be higher than 2 because of cooperative effects between the dimer formed of a NICD-Maml1-Rbpj-kappa complex and additional chromatin modifying co-factors. As discussed in [10], we introduce distinct variables for cytoplasmic and nuclear concentrations of proteins and the respective mRNAs. This distinction is made for the oscillatory factors HES1/7, NICD and LFNG, but not for the slow-changing concentrations of protein and mRNA of Notch1. The DLL1 ligand and the NOTCH receptor are modeled with independent variables in the cytoplasm and membrane compartments.
In the somitogenesis model we included only the genes from our previous model [10] that are needed to generate the ‘wave’-pattern i.e. Dll1, Notch1, Hes7, Fgf8, Wnt3a, and Tbx6, because the downstream genes like Mesp2, Ripply2 and Epha4 would function similar as in our 2012 publication [10] except for possible Hill-threshold adjustments.
A schematic view of the GRN used in our simulations is depicted in Fig. 1. Its central element is the negative feedback oscillator Hes7. By binding to the promoter it inhibits its own production. The Hes7 promoter also receives input from D/N signaling while we disregard here the contribution of Fgf signaling in the tailbud [8]. In an extended model, HES7 inhibits Lfng, which is induced by NICD, and in turn modulates D/N interaction. NICD acts as an activator of Hes7. Here, we assume that HES7 inhibits Dll1 expression. For the mathematical description of the model we use ordinary differential equations. To describe negative feedback oscillators one has to introduce a function describing the repressive action of the gene product on the promoter of its gene. We use Hill functions of the form to describe this negative feedback, wherein the Hill-coefficient h is a measure for the cooperativity of the repressor binding to the promoter and HR as well as HA are the thresholds determining half-inhibition or activation, respectively (see below). For transcription factors binding as homo-dimers we set the Hill coefficient to the value of 2. To describe activating gene action we use analogously Hill functions of the form
Oscillations start only when there is a delay between gene expression and negative feedback. This is often modeled with direct introduction of delayed arguments into the differential equations specifying the time used for transcribing a gene into mRNA and translating a mRNA into protein, resulting in a so-called delay differential equation system (for an example see [15], [47]). In the following, we specify the differential equations of our gene regulatory network. In all cases the gene indices on the variables written on the right side of the equations are not shown except when the variables refer to other genes. Decay rates are always given in min−1 and concentration values are given in arbitrary units.
The equations below describe the negative feedback oscillator at the core of our GRN:Here pC(t), pN(t), mC(t), and mN(t) designate concentrations of cytoplasmic protein, nuclear protein, cytoplasmic mRNA, and nuclear mRNA, respectively. The export rates of the protein from cytoplasm to nucleus, from nucleus to cytoplasm, and for the transport of mRNA from nucleus to cytoplasm are chosen as: epC = 0.007, epN = 0.001, and emN = 0.038. Furthermore, dmC = 0.067, dmN = 0.001, and dpC = 0.031 describe the degradation rates for cytoplasmic and nuclear mRNA, and cytoplasmic protein, respectively. Based on experimental evidence, we assume a rather low rate of mRNA degradation in the nucleus for all genes [48]. We suppose saturated protein decay in the nucleus characterized by threshold value F = 0.2 and maximum rate G = 0.96. The translation rate and the maximal transcription rate are given by K = 1.5 and k = 0.5, respectively. The Hill function with HR = 1.0 and HA = 4.5 describes the negative feedback of HES7 on its own promoter and the control of Hes7 transcription by the Notch intracellular domain (NICD). The bHLH-transcription factor HES7 binds as dimer to its own promoters thereby inhibiting transcription. The Hes7 gene contains only one N-box in its promoter [49]. If HES7 would bind also to the so-called E-boxes in the Hes7 promoter the Hill-coefficient could also be higher [50]. However, Chen et al. have shown that HES7 only binds to the N-box [33], so only one HES7 dimer binds. Therefore we chose a Hill-coefficient of 2. Furthermore, we subsume all interactions with co-factors of HES7 like Groucho/Tle1 in the basal transcription rate. HES7 is a target of D/N signaling. This entails NICD acting as transcriptional co-factor on the Hes7 promoter. As it was shown that two complexes comprising NICD, MAML1 and CSL bind as a dimer to the Hes1 promoter [51] and we assume a similar Hes7 promoter structure regarding activation by NICD, we also use a Hill-coefficient of 2 or higher for the Hill-function describing the effect of NICD in our simulations.
NICD is a fragment of the Notch receptor, which is generated after binding of the DLL1 ligand to the NOTCH1 receptor. Ligand binding enables access of proteases to cleavage sites in the intracellular part of NOTCH1 and subsequent transport of NICD from the cytoplasm to the nucleus [52].Here, rDN = 0.05 is the reaction rate between NOTCH1 receptors and the DLL1 ligands on the n neighboring cells, while raLfng describes the activation of D/N signaling by LFNG, and r0 is the reaction rate of DLL1 and NOTCH1 without LFNG action. For the simulations shown here the default value is 0.256. pMNotch1 designates NOTCH membrane protein concentration, pMDll1 DLL1 protein concentration in the membrane. epC = 0.12 and epN = 0.6 are the export rates for NICD from the cytoplasm to the nucleus and vice versa, and dpC = 0.2 is the NICD decay rate in the cytoplasm. As NICD acts as a co-transcription factor in the nucleus its import rate to the nucleus is chosen larger as the export rate. In the simulations without Lfng in the GRN raLfng is set to 1.
At least in the presomitic mesoderm it was demonstrated that Dll1 expression is dynamic [35]. So the mathematics of negative feedback systems necessitates the use of a transport equation system with at least three equations for oscillatory behavior to be possible [53]. We use two equations for Dll1 mRNA and protein, each in nucleus and cytoplasm, making four differential equations:In the PSM Dll1 is activated directly and indirectly via TBX6 by Wnt signaling [54]. Based on experimental evidence, we assume an additional control by HES7 (see [10] for an extensive discussion). In the spatially constant model system we disregard the control by TBX6 and WNT3A. Therefore, we chose a Hill function of the form , with HR = 1.0. We chose the rate constants as in [10]: K = 1.5, dpC = 0.09, epC = 0.1, epM = 0.1, dpM = 0, dmC = 0.12, emN = 0.09, dmN = 0.001 and k = 1.25. The rate constant rDNcis = 0.01 describing D/N cis-inhibition results in fast synchronization.
After binding of one DLL1 molecule in the membrane of one cell to a NOTCH1 receptor in the membrane of a neighboring cell, the intracellular part of NOTCH1 is cleaved off to release NICD. This results in the destruction of the NOTCH1 molecule in this reaction. Therefore, the reaction term is subtracted in the equation describing NOTCH1 in the membrane, while it is added to the NICD equation. Because the DLL1 ligand bound to the extracellular domain of NOTCH1 is endocytosed and probably degraded [55], the same reaction term is subtracted in the equation describing DLL1 in the membrane. We assume that the same applies to the intracellular complex formed by a DELTA and NOTCH molecule.
Since we assume Notch1 expression to be static it suffices to describe its mRNA concentration by one simple equation with a production and decay term i.e. without differentiating between nucleus and cytoplasm.We chose K = 1.5, dpC = 0.2, epC = 0.1, epM = 0.0, dpM = 0.1, dm = 0.02, and k = 0.5 for the rate constants.
The differential equation system for Lfng has essentially the same structure as the one for Hes7, except that HES7 exerts a repressive influence on the Lfng promoter while NICD activates it. This is described by the Hill function with HR = 1.0 and HA = 4.5.Here pC(t), pN(t), mC(t), and mN(t) designate concentrations of cytoplasmic protein, nuclear protein, cytoplasmic mRNA, and nuclear mRNA, respectively. The export rates of the protein from cytoplasm to nucleus, from nucleus to cytoplasm, and for the transport of mRNA from nucleus to cytoplasm are chosen as: epC = 0.007, epN = 0.001, and emN = 0.038. Furthermore, dmC = 0.067, dmN = 0.001, and dpC = 0.031 describe the degradation rates for cytoplasmic and nuclear mRNA, and cytoplasmic protein, respectively. Again we suppose saturated protein decay in the nucleus characterized by threshold value F = 0.2 and maximum rate G = 0.96. The translation rate and the maximal transcription rate are given by K = 1.5 and k = 0.5, respectively.
For the modeling of growth and geometry in the growing PSM we refer to [10]. We also use the same parameters and equations for the Wnt3a, Tbx6, and Fgf8 genes described therein. Dll1 and Notch1 induction by WNT3A and TBX6, i.e., their corresponding Hill functions, are also chosen as in [10].
Noise is introduced by shutting off transcription only for Hes7, Dll1, Notch1, and Lfng i.e., not for the gradient generating genes, because in our model there is no way this noise could be corrected by D/N signaling. To include this, one would have to simulate a full model of the Wnt3a and Fgf8 pathway with genes like Nkd1 or Dusp4 and others, which exert a negative feedback on their respective pathways and are known to be controlled by D/N signaling [56].
|
10.1371/journal.pgen.1003866 | Detection of Slipped-DNAs at the Trinucleotide Repeats of the Myotonic Dystrophy Type I Disease Locus in Patient Tissues | Slipped-strand DNAs, formed by out-of-register mispairing of repeat units on complementary strands, were proposed over 55 years ago as transient intermediates in repeat length mutations, hypothesized to cause at least 40 neurodegenerative diseases. While slipped-DNAs have been characterized in vitro, evidence of slipped-DNAs at an endogenous locus in biologically relevant tissues, where instability varies widely, is lacking. Here, using an anti-DNA junction antibody and immunoprecipitation, we identify slipped-DNAs at the unstable trinucleotide repeats (CTG)n•(CAG)n of the myotonic dystrophy disease locus in patient brain, heart, muscle and other tissues, where the largest expansions arise in non-mitotic tissues such as cortex and heart, and are smallest in the cerebellum. Slipped-DNAs are shown to be present on the expanded allele and in chromatinized DNA. Slipped-DNAs are present as clusters of slip-outs along a DNA, with each slip-out having 1–100 extrahelical repeats. The allelic levels of slipped-DNA containing molecules were significantly greater in the heart over the cerebellum (relative to genomic equivalents of pre-IP input DNA) of a DM1 individual; an enrichment consistent with increased allelic levels of slipped-DNA structures in tissues having greater levels of CTG instability. Surprisingly, this supports the formation of slipped-DNAs as persistent mutation products of repeat instability, and not merely as transient mutagenic intermediates. These findings further our understanding of the processes of mutation and genetic variation.
| Over 30 diseases are caused by the expansion of a trinucleotide repeat (TNR) in a specific gene, including the most common adult-onset form of muscular dystrophy, myotonic dystrophy (DM1). The mechanistic contributors to this unstable (TNR) expansion are not fully known, although since the discovery of these types of diseases over twenty years ago, the extrusion of the expanded repeats into mutagenic slipped-DNA conformations has been hypothesized. Here, we show the presence of slipped-DNA at the DM1 disease locus in various patient tissues. The allelic amounts of slipped-DNA in tissues correlate with overall levels of repeat instability. Slipped-DNA was also found to form in clusters along a tract of expanded repeats, which has been previously shown in vitro to impede DNA repair. This is the first evidence for slipped-DNA formation at an endogenous disease-causing gene in patient tissues.
| All models proposed to explain the instability of trinucleotide repeats involve DNA slippage at the repeats (Fig. 1) [1]–[12]. Slipped-DNAs were first hypothesized to exist in 1958 [13]. Slipped-DNAs are thought to contribute to more than 30 neuromuscular/neurodegenerative diseases caused by unstable microsatellite repeats, including myotonic dystrophy type 1 (DM1) and numerous cancers that show microsatellite instability [1]–[3], hence understanding slipped-DNAs in patient tissues is of great importance [14], [15]. Expansion mutations continue in DM1 patients as they age, coinciding with worsening symptoms. Patients exhibit inter-tissue repeat length differences as great as 5,770 repeats, with large expansions occurring in affected tissues such as brain, muscle and heart, indicating high levels of continuing expansions [4], [5]. The formation and aberrant repair of slipped-DNAs is a likely source of repeat instability and progressive disease severity in patients (Fig. 1) [6], [7]. An understanding of these DNA mutagenic intermediates in patients should provide insight as to how they may be processed and lead to mutations. The important questions demanding answers are 1) Do slipped-DNAs form at disease loci? 2) Do their levels vary in patient tissues that undergo variable levels of repeat expansion within a given individual? And, 3) What is the biophysical structure of these slipped-DNAs? These questions cannot be answered in a heterologous model system that shows repeat instability that does not reflect the instability ongoing in a patient, nor one lacking tissues. While slipped-DNAs have been characterized in vitro [8], [9], data supporting the presence of slipped-DNAs at an endogenous locus in biologically relevant tissues has been lacking. Previously it was reported that hairpin DNAs can form in a model cell line which contain CTG/CAG repeats integrated at an ectopic locus, where instability was contraction biased as opposed to the expansion bias present in affected patients (10). Although the authors report hairpin formation in living cells, that report could not comment upon either structure formation at an endogenous disease locus, their variation between tissues showing variable instability, upon their persistence in patients, or on the structural features of the DNAs. Furthermore, in that system, instability and hairpin detection depended upon DNA replication, contrasting with the high levels of instability arising in post-mitotic tissues of patients. It is imperative to study instability in patient tissues since repeat mutations are expansion-biased, arising by processes distinct from contractions, coupled with the known locus-specific effects, the tissue-specific variations of instability, and the fact that most expansions arise in non-replicating cells. To identify slipped-DNAs at a disease locus in patient tissues we devised a DNA-immunoprecipitation (IP) protocol that uses a highly specific monoclonal anti-DNA junction antibody (2D3) that recognizes 3-way DNA junctions [9], [11], [12], a structural feature of slipped-DNAs [8], [9]. The 2D3 antibody has been characterized extensively (summarized in Text S1 and citations therein). Briefly, 2D3 binds specifically to junction-containing DNAs with no sequence preference. Footprinting has mapped the 2D3 binding site at DNA junctions [9], [11], [12] where binding can occur from either of the angles sub-tended by the junction arms. For each junction 1–2 antibodies can bind per DNA junction. 2D3 does not bind single-stranded DNA, hairpins, Z-DNA, triplex or quadruplex, with no off-target binding reported in in vitro systems used (see Text S1 and citations therein). 2D3 binds best to slipped-DNAs [9], strengthening its use to isolate these structures.
To verify that the anti-junction DNA antibody recognizes slipped-DNA structures with varying slip-out sizes and slip-out numbers, we assessed 2D3 binding to various slipped-DNAs using an electrophoretic mobility shift assay. Various slipped-DNA structures can form at trinucleotide repeats, and our lab has structurally characterized them in detail by electrophoresis, chemical and enzymatic probing, and electron microscopy [8], [9], [16]–[18]. While a slip-out does not translationally move along the repeat tract (slip or slide), multiple junction conformations occur and these are in dynamic equilibrium [8], [9], [16]–[18]. Previously, we demonstrated that 2D3 binds homoduplex slipped-DNAs of 50 CTG/CAG repeats on complementary strands as well as slipped intermediate heteroduplex DNAs (with an excess of 20 CAGs or 20 CTGs) [9]. These putative mutagenic intermediates may involve isolated slip-outs of various sizes, determined by the length difference between the two repeat-containing strands. Slipped-DNAs can also arise in tracts where the number of repeats between complementary strands does not differ, where each molecule contains multiple clustered short slip-outs [8], [9], [16]–[18]. For example each slipped molecule formed by (CTG)50•(CAG)50 contained 2–62 short slip-outs (1–31 on each strand), each composed of one to three repeat units [8], [9], [16]–[18]. To determine if 2D3 will recognize and bind to smaller slip-outs or clustered slip-outs, DNAs were made with isolated slip-outs of 1- or 3- or 20-excess repeat units of either the CTG or CAG strand, as well as a substrate with multiple clustered slip-outs on both strands along a tract of 50 repeats (Fig. 2A, see schematics). The antibody was incubated with each of these radioactively labeled DNAs, and resolved on polyacrylamide gels to visualize antibody-DNA complexes, evident as slower migrating than the protein-free DNA. Each of the substrates gave rise to slow-migrating species, where only the highest concentration led to partial shifting of the fully-duplexed control (leftmost lanes, see hollow arrowheads). These non-specific complexes were completely lost upon the addition of increasing amounts of competitor DNA (25-fold of non-specific competitor; linearized plasmid). This low-level non-specific binding, typical of many DNA-binding proteins/antibodies, has previously been observed to be readily competed-out [9], [11], [12]. In contrast, the lowest concentrations of 2D3 bound and shifted all of the slip-outs with an excess of 20 repeats (50×30 and 30×50) evident as several species (hollow and gray arrowheads). With the addition of competitor, these substrates remained completely shifted, but migrated faster (black arrowheads), indicating slip-out specific interaction. The 3- and 1-excess slip-outs with variable 2D3 concentrations also yielded shifted species, which were partially resistant to non-specific competition. The multiple shifted species may reflect different antibody-junction conformations [8], [9], [16]–[18], since each antibody can bind at either of the three junction angles, each of these could migrate differently and yield a broad and smeared appearance. The multiple species may also reflect different numbers of antibodies, as it was previously demonstrated that one to two 2D3 antibodies could bind a DNA-junction [9], [11], [12]. The clustered slip-outs were readily shifted by 2D3 and progressively shifted with increasing amounts of 2D3; indicative of increasing numbers of bound antibodies to the many slip-outs/molecule. We have previously shown this interaction to be resistant to competition [9], [11], [12]. Thus, 2D3 bound all sizes of slip-outs tested, with only very minor off-target binding to linear DNA that is eliminated after competition (Fig. 2A). 2D3 binds most effectively to larger slip-outs and clustered slip-outs. Importantly, the anti-DNA junction antibody did not induce the formation of slipped-DNA in fully-duplexed (CTG)50•(CAG)50 DNA (Fig. 2A, leftmost lanes), consistent with its inability to induce cruciform or hairpin extrusion [11], [12], [19]. Further control experiments with longer repeats are described below. 2D3 binding to slipped-DNAs was sensitive to proteinase K. Non-specific antibody-DNA binding was not seen with an isotype-matched anti-actin antibody control, which should not bind DNA (detailed in Supplementary Fig. S1). Thus, the anti-junction DNA antibody specifically recognizes a variety of slip-out sizes as well as DNA containing isolated and multiple clustered slip-outs; supporting the antibody as a useful tool to detect slipped-DNA structures.
To determine whether slipped-DNAs are present in a disease locus of patient tissues we developed a DNA-IP strategy (Fig. 2C). To be sure that we were not inducing slipped-DNA formation through manipulations (genomic isolation, IP etc.) we used several precautionary measures and performed a series of controls. Precautionary measures included preparing genomic DNAs under conditions that avoided DNA denaturation or shearing. Genomic DNAs were prepared from patient tissues under non-denaturing conditions to avoid inducing the formation of unusual DNA structures (see Fig. S1 and Text S1). Slipped-DNAs in CAG/CTG repeats cannot be induced by DNA supercoiling, and their biophysical stability does not depend upon supercoiling [8], [20]. Slipped DNAs can stably exist in linear unrestrained DNA [8]. To reduce DNA shear, enhance immunoprecipitation and eliminate the binding of 2D3 to supercoil-dependent DNA structures (cruciforms), genomic DNAs were restriction digested with BbsI and BamHI (which cut 180 bp upstream and 244 bp downstream of the CTG repeat, respectively, Fig. 2C), and then IP'd using 2D3 antibody and protein G agarose beads. While the 2D3 antibody does not induce cruciform or hairpin extrusion (see above) or slipped-DNAs in short fragments (less than 50 repeats) (Fig. 2A) it was important to assess its effect upon disease-relevant CTG expansions. Control experiments using in vitro prepared DNAs with (CTG)500•(CAG)500 showed that the 2D3 antibody did not induce slipped-DNAs into the expanded tracts (Fig. S2A, detailed in Text S1). Furthermore, we tested the possibility that structures might arise during genomic isolation in two ways; first, by adding the synthetically formed, fully-paired (CTG)500•(CAG)500 or (CTG)50•(CAG)50 DNAs to tissues to follow them through the genomic isolation protocol (Fig. S1), which did not induce slipped-DNA formation, and second, by assembling and disassembling nucleosomes on a supercoiled (CTG)250•(CAG)250 repeat plasmid, which again did not induce slipped-DNAs (Fig. S1). Thus, removal of chromatin from supercoiled repeats and the IP conditions used does not structurally alter the repeats from a fully-paired Watson-Crick duplex, consistent with previous reports [20]. Using these conditions and controls, we proceeded to assess the presence of slipped-DNA in patient tissues.
Towards determining whether the anti-junction antibody could recognize slip-outs that may have arisen at the expanded locus in human patient DNAs we first characterized the CTG tract length in a set of tissues from DM1 patients and a control non-DM1 individual by Southern blot (summarized in Fig. 2B, this LNA-Southern blot analyses have been previously published in [4]). A tissue-specific length variation of the expanded allele was evident in both patients used throughout this study, with many tissues showing a heterogeneous range of repeat sizes (Fig. 2B). Increased length heterogeneity within a tissue and larger expansions between tissues are indicative of high levels of active CTG instability, while limited length heterogeneity and shorter expansions reflect less instability [4]. The largest and most heterogeneous expansions were present in the most affected tissues (cortex, muscle and heart; indicated by broad repeat size ranges), while the cerebellum showed the shortest expansion with little or no length heterogeneity (as indicated by a relatively distinct sized fragment). Control non-DM1 samples showed only non-expanded alleles. These tissues were used throughout the study.
To determine whether 2D3 would recognize and bind to slipped-DNAs formed at the endogenous DM1 locus, we immunoprecipitated (IP'd) DNAs from various tissues, using the IP protocol (Fig. 2C) which would not be expected to either introduce or remove any pre-existing slipped-DNA structures. IP'd DNAs were subsequently characterized by various means, as described below.
Immunoprecipitation of slipped-DNA structures would be expected to enrich for the expanded disease-DM1 allele, as the non-expanded allele would not be expected to contain slipped-DNAs. Slipped-DNAs were in fact present on the expanded but not the non-expanded allele, as outlined: Because the expanded CTG tract in DM1 patient tissues is beyond the PCR amplifiable size range (1310–6550 repeats), we devised a multiplex PCR to distinguish the presence of the disease-allele from the non-disease allele in the immunoprecipitated DNAs. As outlined in Figure 3A, two sets of PCR primers amplifying adjacent overlapping regions of the DM1 locus were used in the same PCR reaction. The regions included the CTG repeat, and the downstream CTCF binding site. In the presence of both the expanded disease allele and the non-expanded allele, we expect four PCR products of distinct sizes. The largest two products would be unique to the non-expanded template derived from amplification across both the short CTG tract and the downstream CTCF site (259 bp), and the short CTG tract alone (211 bp). Amplification of the CTCF site only (204 bp) or the product in the overlapping region (156 bp) could arise from either allele. We did not expect either a PCR product from across the very large CTG expansion or from across the expansion and the adjacent CTCF site. Thus, when only the expanded CTG allele is present, we would not expect PCR products unique to the non-expanded allele (259 bp and 211 bp), whereas when both expanded and non-expanded alleles are present we would expect all four PCR products. This multiplex PCR assay was applied to 2D3-IP'd samples derived from various tissues of a DM1 patient (ADM5) with an expanded CTG tract of 1310–6550 repeats, and a non-expanded allele of (CTG)4 (Fig. 3B). Each genomic DNA prior to IP revealed all four multiplex PCR products, expected from both DM1 alleles (Fig. 3B, all lanes indicated by “−”). The IP'd DNAs revealed only the CTCF and overlapping PCR products (Fig. 3B, all lanes indicated by “+”). The absence of the upper two products, unique to the non-expanded allele provides strong support for the IP'd DNA to be enriched for the expanded DM1 allele, and deficient in the non-expanded allele. Multiplex PCR analysis of 2D3-IP'd DNAs from a non-DM1 control individual with DM1 alleles of (CTG)5 and (CTG)25 failed to detect any PCR products (Fig. S3A) – indicating that the 2D3 antibody could not detect any slipped-DNAs at either of the non-expanded alleles. The specificity of the 2D3-IP protocol to enrich for slipped-DNAs was supported by the inability to detect DNAs derived from a control region (lamin B2) that is devoid of unusual structure forming DNA sequence motifs (Fig. S3B). Thus, the ability to detect only those multiplex PCR products common to both the expanded and non-expanded alleles, but not those unique to the non-expanded allele, supports a specific enrichment of the expanded DM1 allele through anti-DNA junction IP. This strongly supports the presence of slipped-DNA in only the mutant expanded DM1 allele.
We used an independent direct method to determine the enrichment of the expanded DM1 allele in the IP'd material. Triplet-primed PCR (TP-PCR), which has been used as a DM1 diagnostic tool, allows for amplification of short stretches of a large expanded CTG tract [21]. TP-PCR uses a primer that hybridizes to the repeat and a primer that hybridizes downstream of the repeat (Fig. 3C). TP-PCR of the expanded CTG allele typically yields a heterogeneous range of CTG sizes in the PCR products, electrophoretically visible as a smear. This range of PCR products arises by PCR amplification of the repeat-specific primer hybridized randomly along the CTG expansion but relatively close to the flanking primer. TP-PCR across the non-expanded allele yields a distinct but shorter size range of PCR products, due to the limited locations to which the repeat-primer can hybridize (Fig. 3C). TP-PCR analysis of the DM1 patient genomic DNAs prior to IP revealed both the expected smear of products for the expanded allele and the distinct band for the non-expanded allele (Fig. 3D, “Gen.” lanes). However, TP-PCR of the IP'd sample revealed predominantly the longer range smeared products (“IP” lanes) derived from the expanded allele, while the DNA in the supernatant following IP from the same experiment revealed predominantly the shorter products derived from the non-expanded allele (“SN” lanes). These results directly support the enrichment of the expanded DM1 disease allele by the anti-DNA junction antibody, consistent with the interpretation that slipped-DNAs are present along the expanded allele.
The levels of DM1 mutant alleles containing slipped-DNA might be expected to correlate with the levels of CTG instability, which varies between tissues of the same DM1 individual [4]. We quantified the amount of DM1 DNA being IP'd from two tissues from the same patient harboring high and low levels of CTG instability, indicated by the larger and heterogeneous CTG lengths in the heart (ranging from 4100–6000 repeats), and the more discrete and shorter length in the cerebellum (1100–1500 repeats, predominantly 1200 repeats), respectively (Fig. 2B). Quantification was accomplished using the highly accurate competitive quantitative PCR [22], [23] which involves the coamplification of a target DNA sample with known amounts of a cloned competitor DNA that shares most of the nucleotide sequence and primer sites with the target (detailed in Fig. S4A). Competitive PCR permits quantification of the absolute number of target molecules (a few representative examples are shown in Fig. S4B and S4C). Anti-DNA junction antibody IP'd DNA levels were significantly greater in the heart over the cerebellum (relative to input) of a DM1 individual (Fig. 4A), an enrichment that is consistent with a greater proportion of DM1 alleles with slipped-DNA present in tissues having greater levels of CTG instability. These findings indicate more DM1 alleles with at least one slipped-out, but cannot reveal the number of slip-outs per allele. DNA from a non-DM1 individual was not enriched for slipped-DNAs in either cerebellum or heart (Fig. 4A). These findings reveal a trend of increased levels of slipped-DNAs in tissues with higher levels of instability.
As a further test for the presence of slipped-DNAs in DM1 patient tissues, DNA was treated with enzymes that specifically digest features of slipped-DNA structures; a treatment that should eliminate slipped-DNAs. Mung bean nuclease (MBN) and T7 endonuclease I (T7endoI) have been shown to specifically cleave the single-stranded regions in slip-outs and across DNA-junctions of slipped-DNAs [8], [16] (Fig. S5D). Following enzyme digestions, DNAs were subjected to TP-PCR to assess the potential loss of the expanded alleles (Fig. 4B,C,D; additional examples in Fig. S5). If the expanded allele contains slipped-DNA, cleavage by these enzymes should decrease the amount of the smear representing the expanded allele. The amount of expanded repeat products in several DM1 patient tissues was significantly reduced following treatment with MBN or T7endoI (Fig. 4D, Fig. S5C). Non-DM1 control DNAs showed no difference after enzyme treatment (Fig. S5B,C). The significant reduction of the expanded allele by either structure-specific enzyme is consistent with a portion of the DM1 disease alleles being in the slipped-DNA conformation.
It was of interest to know if the slipped DNAs were present within tissues in the native chromatin context. Furthermore, the detection of slipped-DNAs directly on chromatinized DNA in tissues, where freeze and thawing of tissues does not affect chromatin packaging relative to fresh tissues [24], provides further evidence for the existence of slipped-DNAs prior to chromatin removal during DNA isolation in the above experiments. To assess the presence of slipped-DNAs in tissues, we used a slight modification of an established protocol of DNA nuclease accessibility assay on tissues, which assesses DNA in its native chromatin context [24] (see references in Text S1). Unusual DNA conformations have been detected in native chromatin through nuclease digestion [25]–[27]. For example, cruciform and stem-loops structures, similar to slipped-DNAs, were susceptible to nuclease digestion and found to stably reside in the inter-nucleosomal region [26], [27]. DM1 patient and control tissues were treated directly with either buffer only as a negative control, or both T7endonucleaseI and mung bean nuclease to test for digestion, or AluI restriction enzyme to test for off-target digestion (Fig. 5A, B; Fig. S6). Following treatment, DNA was isolated and subjected to TP-PCR to assess the potential loss by digestion of the expanded alleles. There was a significant reduction in the expanded allele from patient ADM9 muscle tissue when treated with T7 and MBN nucleases (p = 0.0038), but not when treated with AluI (p = 0.255). There was no significant reduction in either control cerebellum or heart, or patient cerebellum, after treatment with either T7 and MBN, or AluI (Fig. 5A,B; Fig. S6). These results are consistent with the above results in that the tissue with greater CTG instability contains more slipped-DNAs, and control tissues do not. Thus, slipped-DNAs were present at the mutant DM1 locus, in the native chromatin in DM1 tissues and at levels that correlated with levels of instability.
Towards biophysically characterizing the slipped-DNAs, we visualized these by electron microscopy (EM). EM analysis of the IP'd slipped-DNAs revealed multiple bends, kinks, bulges, branched arms, and regions with increased thickness relative to control DNA (Fig. 5C, D, Fig. S7). A significantly increased number of molecules with visibly detectable slip-outs were identified in tissues showing high levels of CTG instability compared to the cerebellum, which showed the lowest level of instability (skeletal muscle vs. cerebellum, two-sided t-test; p = 0.005; pancreas vs. cerebellum, two-sided t-test; p = 0.0202; all unstable tissues (heart, liver, pancreas, cortex, skeletal muscle) vs. cerebellum; p = 0.0386; Fig. S6C). The variation of visibly detectable slip-outs between IP'd material of different tissues might indicate a slip-out size variation between tissues. The size of the slip-outs presented a bimodal distribution ranging from 1–100 repeats with peaks at ∼30 and <10 repeats (Fig. 5D). Multiple slip-outs were clustered along a given DNA, with distances of <100 bp between slip-outs (Fig. 5D). These features were also present in in vitro induced slipped-DNAs in synthetic (CTG)800•(CAG)800 (Fig. S7A) and are consistent with previous electron micrographs showing multiple short slip-outs in (CTG)•(CAG) tracts with 50–250 repeats [16], [17]. The IP'd DNAs were shorter than expected, possibly due to IP-induced DNA shearing of the very long repeat expansions that we studied, a phenomenon previously observed for DNAs enriched in single-strand nicks, gaps or DNAse I hypersensitive sites. An enrichment of such strand breaks may be expected for arrested repair of clustered slip-out lesions [28], a sensitivity to site fragility [29], and susceptibility of slipped-DNAs to double-strand breaks [30].
We report the isolation of slipped-DNAs present at an endogenous disease locus from various DM1 patient tissues. Our observations are supported by multiple independent, complementary experimental approaches. To better understand processes leading to disease-causing mutations, it is crucial to study these structures at specific genomic loci and in relevant patient tissues. This is particularly important for trinucleotide repeat mutations as these events are expansion-biased, arising by processes distinct from contractions, with instability patterns varying between disease loci, and between mitotic and non-mitotic tissues. An understanding of these DNA mutagenic intermediates occurring in patients should provide insight as to how they may be processed and lead to mutations, as model systems do not accurately reflect what is ongoing in affected individuals, especially on a tissue-specific level. Here, we reveal slipped-DNAs at the expanded CTG/CAG DM1 disease locus in tissues including the brain and heart. The brain in humans is essentially post-mitotic after birth [31], with both the cerebellum and cortex developing from very early on in embryogenesis and continuing up until the first years after birth [32], [33]. Similarly, the majority of cells in the heart are post-mitotic in nature, with cell division ceasing in cardiac myocytes soon after birth and the dividing cell population in adults being less than 1% of the total number of cells [34]–[36]. Since CTG expansions continue to expand at a greater rate following birth compared to prenatal expansion [4], our detection of slipped-DNAs in non-mitotic tissues supports that they arose in the absence of DNA replication. However, it is also possible that slipped-DNAs could form during early development and persist throughout the lifetime of an individual. Despite the similarity in timing of the post-mitotic nature of the frontal cortex, cerebellum and the heart, the cortex and heart showed significantly more slipped-DNA in patients compared to cerebellum, further supporting the hypothesis of the presence of tissue-specific instability factors. In DM1 and other CAG/CTG repeat diseases the tissue selectivity of repeat expansions correlates with the tissue selectivity of pathogenesis in the brain regions [37]–[41], the heart [42] and muscle [4], [43]. Our evidence at a patient locus from various tissues serves as the only evidence supporting the existence of slipped-DNAs in patients. Our observations of slipped-DNAs at a disease locus in post-mortem patient tissues, coupled with the report of hairpin DNAs at exogenously integrated repeats in cultured cell models [10] supports the formation of slipped-DNAs in patients. Importantly, we reveal a strong difference in the allelic levels of slipped-DNA between tissues of the same individual, and these levels correlate with the levels of instability in the tissues assessed.
Surprisingly, slipped-DNAs are present as tandem clusters of multiple small slip-outs, with increased levels in DM1 patient tissues showing high repeat instability. This contrasts with the long-presumed concept that slipped-DNAs arise as isolated slip-outs on either strand of the duplex. That slipped-DNAs are clusters of short slip-outs gives new insight into the role they may play in the mutation process. The majority of analyses addressing the potential roles of slip-outs have centered upon the presumption that single isolated slip-outs may form transiently and be processed to fully-duplexed repeat-length mutations. Complementing our isolation of clustered slipped-DNA structures, it has been shown previously that clustered structures are poorly repaired compared to isolated single slip-outs [28]. The most striking implication of our findings is that slipped-DNAs do not appear to be merely transient in nature, as previously believed [44], but may be persistent in vivo products of instability, consistent with their poor ability to be repaired [28]. We propose that intrastrand slippage may occur after attempted repair events on slipped-DNAs become arrested due to adjacent slip-outs (Fig. 6). This shifts the repeats further out of register leaving a gap that, when filled, results in an increased number of repeats in that strand and producing slipped-DNAs with an excess of repeats on one or both strands. Reiterations of such events in the absence of proper repair would lead to instability, where products are slipped-DNAs.
Our IP protocol has identified slipped-DNAs at a CTG/CAG disease locus in patient tissues, structures that were not detected in appropriate controls (i.e., the antibody did not induce structure formation). While our findings are not definitive in vivo evidence, the fact that the allelic levels of slipped-DNAs varied between DM1 patient tissues, which showed varying levels of instability supports that these were formed in the patient. This concept is further supported by the detection of hairpins at integrated CTG/CAG repeats in cell models [10]. Different mutational paths involving slipped-DNAs may involve specific DNA repair pathways. For example, the mismatch repair complex MutSβ is required for CAG/CTG expansions [1]–[3]. Similarly, the levels and biophysical structure of slipped-DNAs forming at mono- and dinucleotide repeats may be distinct between tumors showing microsatellite instability compared to those that are microsatellite stable. Slipped-DNA identification and characterization could provide mechanistic and prognostic insights into medically important mutations in the many repeat-disease loci, hypermutable viruses, mitochondria, hypervariable genomic regions, and fragile sites [45]–[49] as it has done herein for DM1 repeat instability.
Human tissues analyzed in this study are listed in Fig. 2B. Autopsy tissues from a non-affected individual (ADN1) were obtained snap-frozen from the National Disease Research Interchange. DNA extractions from human tissues were carried out under conditions to minimize DNA denaturation, to avoid inducing unusual DNA structure formation. Procedures were performed in low binding tubes, which avoid DNA denaturation. Complete protocol can be found in Text S1. Human-CTG repeat lengths were assessed by Southern blot using an LNA probe [DIG-labeled (CAG)7-5′-gcAgCagcAgCagCagcAgca-3′], as described previously [4]. Non-expanded CTG alleles were sized by sequencing of the products obtained after PCR amplification (forward primer 409, reverse primer 407). Slipped-DNA structures for bandshift were made as previously described [9] with minor changes (see Text S1). Individual structures and varying concentrations of antibody were run on a 4% (w/v) polyacrylamide gel in 1× TBE buffer at a constant 150 V for 1.5 hours. Following electrophoresis gels were dried and autoradiographed. Fully-duplexed (CTG)50•(CAG)50 were used to test for specificity of binding, with linearized p-Bluescript plasmid additionally being used as a competitor during antibody binding. DNA-IP of slipped-structures was carried out using the anti-DNA junction monoclonal antibody 2D3. 1 ug of genomic DNAs were restriction digested with BamHI and BbsI overnight at 37°C. 50 ul of hybridoma 2D3 culture supernatant containing ∼5 ug/ml immunoglobulin was diluted 1∶1 with PBS and added to the restriction digested patient DNA the following day and incubated on ice for one hour. 80 ul Protein G beads were prepared and resuspended in 250 ul 1× TE, which were then added to the antibody-DNA mixture and incubated on ice for 1 hour, mixing occasionally. The antibody/DNA/bead complex was washed free of unbound DNA and the affinity purified DNA was eluted using TE+2% SDS. DNA was further purified by phenol-chloroform-isoamyl alcohol extraction, followed by 100% ethanol precipitation. IP'd DNA samples were analyzed by electron microscopy essentially as described [16]. Multiplex PCR and TP-PCR protocol conditions are described in Text S1. For structure specificity experiments, patient DNA was digested with MBN or T7EndoI overnight and used for TP-PCR with either a fluorescently labeled primer after which the products were analyzed by GeneScan, or with a non-fluorescently labeled primer after which the products were run on a 1% agarose gel for visual analysis. DNAs digested within their native chromatin context were subjected to a modified nuclease accessibility protocol before digestion and DNA isolation (Text S1). EM and nucleosome assembly and disassembly were carried out as previously reported (Text S1, references 9 and 42, respectively).
|
10.1371/journal.pcbi.1007275 | Characterizing and dissociating multiple time-varying modulatory computations influencing neuronal activity | In many brain areas, sensory responses are heavily modulated by factors including attentional state, context, reward history, motor preparation, learned associations, and other cognitive variables. Modelling the effect of these modulatory factors on sensory responses has proven challenging, mostly due to the time-varying and nonlinear nature of the underlying computations. Here we present a computational model capable of capturing and dissociating multiple time-varying modulatory effects on neuronal responses on the order of milliseconds. The model’s performance is tested on extrastriate perisaccadic visual responses in nonhuman primates. Visual neurons respond to stimuli presented around the time of saccades differently than during fixation. These perisaccadic changes include sensitivity to the stimuli presented at locations outside the neuron’s receptive field, which suggests a contribution of multiple sources to perisaccadic response generation. Current computational approaches cannot quantitatively characterize the contribution of each modulatory source in response generation, mainly due to the very short timescale on which the saccade takes place. In this study, we use a high spatiotemporal resolution experimental paradigm along with a novel extension of the generalized linear model framework (GLM), termed the sparse-variable GLM, to allow for time-varying model parameters representing the temporal evolution of the system with a resolution on the order of milliseconds. We used this model framework to precisely map the temporal evolution of the spatiotemporal receptive field of visual neurons in the middle temporal area during the execution of a saccade. Moreover, an extended model based on a factorization of the sparse-variable GLM allowed us to disassociate and quantify the contribution of individual sources to the perisaccadic response. Our results show that our novel framework can precisely capture the changes in sensitivity of neurons around the time of saccades, and provide a general framework to quantitatively track the role of multiple modulatory sources over time.
| The sensory responses of neurons in many brain areas, particularly those in higher prefrontal or parietal areas, are strongly influenced by factors including task rules, attentional state, context, reward history, motor preparation, learned associations, and other cognitive variables. These modulations often occur in combination, or on fast timescales which present a challenge for both experimental and modelling approaches aiming to describe the underlying mechanisms or computations. Here we present a computational model capable of capturing and dissociating multiple time-varying modulatory effects on spiking responses on the order of milliseconds. The model’s performance is evaluated by testing its ability to reproduce and dissociate multiple changes in visual sensitivity occurring in extrastriate visual cortex around the time of rapid eye movements. No previous model is capable of capturing these changes with as fine a resolution as that presented here. Our model both provides specific insight into the nature and time course of changes in visual sensitivity around the time of eye movements, and offers a general framework applicable to a wide variety of contexts in which sensory processing is modulated dynamically by multiple time-varying cognitive or behavioral factors, to understand the neuronal computations underpinning these modulations and make predictions about the underlying mechanisms.
| In many brain areas, particularly ‘associative’ regions including parietal and prefrontal cortex, sensory processing is affected by various intrinsic or extrinsic nonsensory covariates such as task or context variables, attention, learned associations, motor preparation, or cognition-related control signals. The fact that multiple such variables may simultaneously modulate sensory activity, and that their influence can change rapidly over the course of a task, poses challenges for precise experimental or computational quantification of their relative contributions to neuronal responses. In this paper, we develop a data-driven computational framework, which provides a rich statistical description of encoding time-varying sensory information by capturing and dissociating multiple time-varying modulations on the order of milliseconds.
We develop and test our model in the context of changes in visual sensitivity around the time of eye movements, which is an exemplar of such time-varying modulatory computations. There is a considerable literature demonstrating that visual neurons’ responses are modulated during rapid eye movements, known as saccades; these neurophysiological changes presumably underlie the biases in perception which also occur around the time of eye movements. Even during fixation, extrastriate cortical responses can show complex spatiotemporal dynamics which encode information about the stimulus [1]. Various types of perisaccadic response modulations have been reported. For example, many studies have shown that visual neurons lose their sensitivity to stimuli appearing in their receptive fields (RFs) shortly before a saccade (saccadic suppression) [2–7]. Other studies have demonstrated that visual neurons may preemptively shift their RF to the post-saccadic RF (future field remapping, or FF-remapping) [8–15], or to the saccade target (saccade target remapping, or ST-remapping) [16–18], even before a saccade is initiated. Moreover, there are several reports suggesting that the spatial distribution of the population of visual neurons’ RFs changes during or just prior to a saccade [19–21]. Taken together, these findings indicate that the perisaccadic responses evoked in visual neurons are modulated by several sources, i.e. stimuli perisaccadically presented at multiple locations in the visual field contribute to driving neurons’ responses.
Although existing experimental data identify those sources contributing to perisaccadic response modulation, they do not quantitatively characterize the contribution of each modulatory source to the response, alone or in combination with the other sources. A full understanding of visual perception during saccades may require the ability to reconstruct the visual scene across an eye movement based on neural activity, and this in turn necessitates a comprehensive understanding of how each modulatory source individually contributes to the perisaccadic representation, how the contributions of multiple sources are combined, and more importantly, how knocking out one of the sources may impact the reconstruction of the scene based on neural activity. Some of the limitations of the experimental data are practical: the short timescale on which the perisaccadic changes take place, combined with the limited number of trials that can be recorded in a single recording session, do not permit a full test of the contribution of different sources at each time point and in all combinations. This is where computational models come into play with two important roles, (1) making predictions of the neural responses to a wide variety of stimuli, and, (2) providing a quantitative description of how the modulatory sources contribute to response generation at different times relative to a saccade.
The fast changes in sensitivity around the time of saccades pose challenges for computational as well as experimental approaches. These changes make the stimulus-response relationship time-variant and create nonstationary responses and computations. This property makes many existing computational approaches, which are often based on time-invariant assumptions about the neural system, not applicable for modeling the nonstationary responses observed during a saccade. The approaches that have commonly been used to characterize nonstationary responses can be divided into three main categories. In the first approach, separate models are applied to several (overlapping or nonoverlapping) time intervals assuming that the stimulus-response relationship remains constant within each of those intervals [22–26]. This approach is more suitable for providing coarse snapshots of the neural states rather than analyzing how the states evolve over time. To address this limitation, the second approach provides methods for estimating the temporal variations of the model parameters to keep track of the temporal evolution of the underlying system. Among the methods using this approach, adaptive filtering solutions have widely been used in neural data analysis, especially in studying the temporal evolution of the spatiotemporal and spectrotemporal receptive field of neurons [27–32]; however, depending on the size of their parameter space, these models require a large amount of data for their parameter estimation, and as a result they fall short in the cases where the evolution of the underlying system happens on a very short timescale, which is the case in perisaccadic studies. The third approach uses state-space methods such as linear dynamical systems [33–35] or hidden Markov models [36–41], in which the next state of the system is determined based on its current state and its input. Although these methods have been successful in modeling dynamic neural data and especially for neural decoding applications, similar to adaptive filters, they require a large amount of data to work, which makes them insufficient for the resolution or precision required for modeling the perisaccadic responses. To address the need for a quantitative means to study nonstationary responses across a saccade, we recently developed an extension of the widely-used generalized linear models (GLMs) [42–47], termed the nonstationary generalized linear model (NSGLM, referred to as the N-model in this article) to describe response dynamics in visual neurons across a saccade [48]. The multiplicative spatiotemporal gain kernels introduced in the N-model recovered the rapid eye displacement signal and the resulting nonstationarity in responses solely based on the statistical relationship between the stimulus and response across a saccade. However, the N-model structure remained limited to the types of nonstationarity that could be described by gain factors modulating the filtered input stimuli, and could not be generalized to explain the range of response nonstationarities and their underlying modulatory computations observed during saccades.
None of the existing models are capable of tracking the dynamic changes in sensitivity which accompany saccades on millisecond timescales. In this paper, we present a new nonstationary approach to develop a sparse-variable generalized linear model (referred to as the S-model), which is capable of tracking the rapid changes occurring in the stimulus-response relationship across a saccade in the middle temporal (MT) cortex of nonhuman primates. The S-model is composed of a set of time-varying stimulus kernels which represent the time-varying spatiotemporal sensitivity of a neuron. Building on the success of the S-model in predicting perisaccadic responses, a circuit-inspired factorized version of the S-model (referred to as the F-model) was developed in order to dissociate and to emphasize the role of individual sources contributing to perisaccadic response modulation. In the F-model, each stimulus kernel is decomposed into a parsimonious set of multiple modulatory sources, combined to describe the spatiotemporal receptive field of the neuron. The F-model not only accurately captures the perisaccadic changes in neural sensitivity, but also provides a tractable computational model by which the contribution of various modulatory sources can be dissociated. This temporally precise, quantitative decomposition of a neuron’s perisaccadic responses offers unique opportunities for quantifying perisaccadic modulations and testing the perceptual effects of various perisaccadic modulatory sources. Furthermore, the computational framework can be applied to quantify and dissociate the effects of multiple modulatory factors on neuronal responses in a variety of brain areas and behavioral tasks.
The principal objective for the new computational framework developed in this study is to provide a statistical framework that will capture the encoding of time-varying information in higher brain areas. The desire for such a model was motivated by our findings about the perisaccadic response properties of neurons in area MT of macaque monkeys. The activity of 41 single neurons in MT cortex was recorded while animals performed a visually-guided saccade task with probe stimuli (Fig 1A). Each stimulus appeared on the screen for only 7 milliseconds (ms), allowing a high spatiotemporal resolution mapping of the neurons’ visual sensitivity. Neurons exhibited several types of changes in perisaccadic sensitivity, including saccadic suppression, FF-remapping, and ST-remapping. Fig 1B shows the saccadic suppression effect in an example neuron and the subpopulation of significantly modulated neurons (n = 8). Perisaccadic visual responses to a stimulus in the original RF are reduced compared to responses during fixation. For the example neuron, the average response to an RF stimulus dropped from 60.89 ± 2.82 (mean ± SE) spk/s during fixation to 40.49 ± 12.64 (mean ± SE) spk/s when the stimulus appeared just prior to saccade onset (example neuron, p < 0.001). For the subpopulation of neurons with significant saccadic suppression (see Methods), the average normalized response dropped from 2.53 ± 0.37 to 1.51 ± 0.29 (mean ± SE). Fig 1C shows the FF-remapping effect in an example neuron and the subpopulation of significantly modulated neurons (n = 23). During fixation there is no response to stimuli appearing in the FF; however, during the perisaccadic period neurons respond to stimuli in the FF. Note that the perisaccadic FF response occurs at longer latency than RF responses (neurons responding 50–75 ms after stimulus onset for the RF, vs. 80–150 ms after stimulus onset for the FF). For the example neuron, the average late response to an FF stimulus increased from 7.90 ± 1.02 (mean ± SE) spk/s during fixation to 18.42 ± 3.56 (mean ± SE) spk/s perisaccadically (example neuron, p < 0.001). For the subpopulation of neurons with significant FF-remapping, the average normalized response to FF stimuli increased from 0.90 ± 0.03 to 1.30 ± 0.08 (mean ± SE). Fig 1D shows the ST-remapping effect in an example neuron and the subpopulation of significantly modulated neurons (n = 37). During fixation there is no response to stimuli appearing around the ST; however, during the perisaccadic period neurons respond to stimuli around the ST. Like the FF-remapping effect, the perisaccadic ST response occurs at longer latency than RF responses. For the example neuron, the average late response to a stimulus near the ST increased from 23.80 ± 1.77 (mean ± SE) spk/s during fixation to 64.64 ± 7.91 (mean ± SE) spk/s perisaccadically (example neuron, p < 0.001). For the subpopulation of neurons with significant ST-remapping, the average normalized ST response increased from 0.83 ± 0.03 to 1.35 ± 0.06 (mean ± SE). The RF, FF, and ST effects could occur in different combinations or relative strengths across neurons. Fig 1E shows the FF- and ST-remapping effects for an example neuron and the subpopulation of neurons exhibiting both effects (n = 22). For the example neuron, the average late response to a stimulus in the FF increased from 15.83 ± 0.91 (mean ± SE) spk/s during fixation to 39.10 ± 5.14 (mean ± SE) spk/s perisaccadically (example neuron, p < 0.001), and the average response to a stimulus near the ST increased from 16.15 ± 0.94 (mean ± SE) spk/s during fixation to 36.17 ± 4.49 (mean ± SE) spk/s perisaccadically (example neuron, p < 0.001). For the subpopulation of neurons with significant FF- and ST-remapping, the average normalized FF response increased from 0.90 ± 0.03 to 1.32 ± 0.08 (mean ± SE), and the average normalized ST response increased from 0.80 ± 0.03 to 1.46 ± 0.08 (mean ± SE).
A set of novel variants on a classical GLM framework were developed with the goal of capturing perisaccadic modulations in visual sensitivity. The principal idea for using a GLM framework as a base structure in this study has several folds: (i) providing a rich, statistical description of stimulus-response relationship, (ii) computational tractability of the estimation procedure, (iii) providing biologically plausible response generation and modulation components, and finally, (iv) its flexibility to incorporate a variety of external and internal covariates depending on the experimental design and the estimation or prediction tasks at hand, which is crucial in this study. The structure of the models is largely similar (see Methods), with the exception of the stimulus kernels which are used to represent changes in spatiotemporal sensitivity. In the sparse-variable generalized linear model (S-model), the stimulus-response relationship in a neuron is characterized by time-varying stimulus kernels representing the time-dependent spatiotemporal receptive field profile of the neuron. The kernels of the S-model were estimated by maximizing the likelihood of the observed spikes under the instantaneous firing rate predicted by the model over the training set, and validating over the validation set. We next produced a parsimoniously factorized version of the S-model (F-model), in which the fitted kernels of the S-model were represented as a mixture of three time- and delay-dependent spatial skewed Gaussian kernels, where each captures the modulation arising from one of the RF, FF, or ST sources (corresponding to the modulations observed in perisaccadic responses as detailed in the last subsection). Finally, an aggregate model (A-model) was constructed by fitting the S-model components ten times (over subsets each including a randomly selected 65% of the data), creating an F-model from each of these S-models, and then taking the average of the F-model components, in order to obtain a model with kernels reflecting the entire dataset.
Fig 2 shows the full structure of the S- and F-models. Both models specify the probabilistic relationship between a sequence of input stimuli, defined in time and space, and the measured neural spike trains on the scale of single trials. Inputs to the models (Fig 2A) are convolved with a set of time-varying stimulus kernels (Fig 2B and 2C). For the S-model, each probe location has its own time-varying kernel (S-kernels, Fig 2B). In the F-model, each S-kernel is optimally approximated by a combination of the RF, FF, and ST modulatory sources added to a fixation kernel (F-kernels, Fig 2C). The output (Fig 2D) is then summed with an offset kernel (Fig 2E) and the feedback signal (Fig 2F) generated by the post-spike kernel (Fig 2G). The resulting generator signal (Fig 2H) is then passed through a nonlinearity (Fig 2I) to generate the instantaneous firing rate (Fig 2J). A Poisson spike generator (Fig 2K) is used to generate the spiking response (Fig 2L).
Fig 3 illustrates the ability of the S- and F-models to reproduce the time course of the neural activity across trials. Two sample trials are shown for three neurons indicating how well the models captured the instantaneous firing rate of the neurons on the unseen data. The trials on the left show examples of high prediction accuracy (from top to bottom, ΔLL/spk = 1.12, 0.80, and 0.63 bits/spk for the S-model; and ΔLL/spk = 0.81, 0.47, and 0.55 bits/spk for the F-model), and the trials on the right show examples of median prediction accuracy (from top to bottom, ΔLL/spk = 0.46, 0.41, and 0.19 bits/spk for the S-model; and ΔLL/spk = 0.41, 0.34, and 0.05 bits/spk for the F-model) for each neuron and model. (The A-model was omitted from this analysis as it uses all data when aggregating over the multiple F-model fits, and so, there is no unseen data for the A-model.)
A summary of the models’ structures and key properties is presented in Fig 4. The N-model was described in a previous publication [48] and is presented here for the purpose of comparison. The N-, S-, and F-models are all variations on the well-known GLM structure, using different approaches to capture the nonstationarities existing in the perisaccadic responses. In terms of model components (Fig 4, first row), the N-model has a gain kernel and time-invariant stimulus kernel at each probe location, as well as post-spike and offset kernels, while the S-model has time-variant stimulus kernels at each probe location, a post-spike kernel, and an offset kernel. The components of the F-model were three skewed Gaussian kernels, and post-spike and offset kernels taken from the S-model. In the N-model, the nonstationarity is modeled by multiplying a space- and time-varying gain by the result of the convolution of input stimuli and time-invariant stimulus kernels (Fig 4, second row). Although successful in capturing the shift of spatial sensitivity following an eye movement, the N-model fails to describe the perisaccadic responses arising from a range of modulatory computations beyond an instantaneous gain mechanism, e.g. changes in response latency. The S-model deals with this issue by employing time-varying stimulus kernels, providing more degrees of freedom for the model; in effect, the S-model can be considered as a set of GLMs, each one corresponding to a single time point and all fitted simultaneously. The S-model accurately characterizes the observed perisaccadic modulations, but does not lend itself easily to a mechanistic level interpretation. To find a way to interpret the S-kernels, and identify and dissociate possible sources that give rise to the spatiotemporal sensitivities revealed by the S-kernels, the fitted stimulus kernels of the S-model were factorized into sources of modulation in the F-model, such that stimulus kernels were reconstructed by a combination of three modulatory sources added to a fixation kernel representing the neuron’s behavior during fixation. This allowed the interpretation of changes in responses (especially perisaccadic responses) as resulting from changes in the characteristics of a few modulatory sources. Due to the structure of the N-model, the model components responsible for response generation (stimulus kernels) and response modulation (gain kernels) are fitted independently, while these components are interwoven in the S- (two-dimensional stimulus kernels) and F-models (skewed Gaussian kernels) (Fig 4, third row). While in the N- and S-models there are separate stimulus kernels corresponding to separate probe locations, the modulatory sources in the F-model have a spatial profile, and so, the spatiotemporal receptive field structure of the neuron in the F-model is described by just three sources at each time point relative to the saccade instead of a set of stimulus kernels (Fig 4, fourth row). Moreover, while the response latency cannot vary across time in the N-model, and can vary at each time and probe location in the S-model, the F-model determines the response latencies based on three modulatory sources (Fig 4, fifth row).
The ability of the N-, S-, and F-models to predict single spike trains (over test data) during the fixation vs. perisaccadic period is compared in Fig 5A (again the A-model was omitted from this analysis for the same reason described in Fig 3). The N-model is no better at predicting the spike times during the perisaccadic period compared to the fixation period (ΔLL/spk = 0.12 ± 0.02 (mean ± SE) bits/spk during fixation vs. 0.13 ± 0.02 (mean ± SE) bits/spk during saccade, p-value = 0.26). For the S-model, the prediction accuracy increases in the perisaccadic period (ΔLL/spk = 0.24 ± 0.02 (mean ± SE) bits/spk during fixation vs. 0.30 ± 0.02 (mean ± SE) bits/spk during saccade, p-value < 0.001). The F-model, despite being based on an approximation of the S-model, also shows greater prediction accuracy during the perisaccadic period compared to fixation (ΔLL/spk = 0.14 ± 0.01 (mean ± SE) bits/spk during fixation vs. 0.18 ± 0.02 (mean ± SE) bits/spk during saccade, p-value = 0.003). These improvements reflect the efficiency of the fitting strategy tailored to capture dynamic aspects of the response. The richer architectures of the S- and F- models allowed them to capture dynamic changes in the neuron’s spatiotemporal response as a function of time to the saccade, and as a result, both better predict the responses compared to the N-model. Both the S-, and F-models outperform the N-model during both the fixation (S-model vs. N-model: p-value < 0.001; F-model vs. N-model: p-value = 0.001) and perisaccadic (S-model vs. N-model: p-value < 0.001; F-model vs. N-model: p-value < 0.001) periods. In all comparisons, the statistical significance was determined by the Wilcoxon signed-rank test (n = 41 neurons).
As seen in Fig 5A, the F-model trails the S-model in terms of spiking time prediction, which is not surprising given the greater number of variables in the S-model; however, when it comes to the previously reported perisaccadic modulations (i.e., saccadic suppression, FF-remapping, and ST-remapping), the F-model performs as well as the S-model. In order to demonstrate this point, Fig 5B examines the ability of all four models to reproduce the saccadic suppression, FF-remapping, and ST-remapping effects (over all recorded trials, not only test trials, to avoid a bias toward a subset of data). For this analysis, neurons were categorized as displaying or not displaying each of three perisaccadic effects using their recorded responses. Then, the models’ predictions were used to re-classify neurons into the same categories, and the response-based and model-based classifications were compared using four different measures of classification performance (sensitivity, accuracy, the geometric mean of sensitivity and precision, and F-measure). The predictions of each model were compared using the one-tailed mid P-value McNemar test [49], with the null hypotheses that the S-/F-/A- and N-models have equal performance in terms of reproducing each of three perisaccadic effects (P-values less than 0.05 were considered a rejection of the null hypothesis). For the FF-remapping effect, the A-model significantly outperforms the N-model (p = 0.048); the S- and F- models did not significantly outperform the N-model (S-model vs. N-model: p = 0.17, and F-model vs. N-model: p = 0.08). The S-, F-, and A-models all significantly outperform the N-model in terms of reproducing the ST-remapping effect (S-model vs. N-model: p = 0.004, F-model vs. N-model: p = 0.006, and F-model vs. N-model: p = 0.006). Finally, we did not find any significant difference between the S-, F-, and A-models and N-model in terms of reproducing the saccadic suppression effect (S-model vs. N-model: p = 0.38, F-model vs. N-model: p = 0.17, and F-model vs. N-model: p = 0.11), indicating that a global change in gain is sufficient to reproduce the saccadic suppression effect.
Additionally, in order to compare the ability of the S- and F-models to accurately predict responses specifically during the perisaccadic period, we calculated the ratio of perisaccadic performance to fixation performance (in terms of ΔLL/spk over test data) for the S- and F-models. Since adding additional variables is generally expected to improve model performance, during both the fixation and perisaccadic periods, this ratio of perisaccadic to fixation performance was used to quantify the added perisaccadic predictive value of the model regardless of overall changes due to the degrees of freedom. The S-model has a lower perisaccadic to fixation ratio compared to the F-model (1.33 ± 0.99 (median ± SE) for the S-model vs. 1.51 ± 0.25 (median ± SE) for the F-model, Wilcoxon signed‐rank p‐value = 0.003), indicating that although overall performance is higher due to the S-model’s greater degrees of freedom, the F-model does better at specifically reproducing perisaccadic changes in the response. In order to assess the contribution of individual sources, represented by time- and delay-dependent spatial skewed Gaussian kernels, to the F-model’s performance, different variants of the F-model were created (as detailed in the Methods section). These variants included models in which only one of the three sources was included (+RF, +FF, and +ST), models in which one of the sources was eliminated (-RF, -FF, or -ST), and one in which all three modulations were eliminated (no-source). A linear regression was conducted in which the slope of a line fitted to the perisaccadic vs. fixation performance for the population of 41 neurons was considered as an estimate for the perisaccadic to fixation ratio of each model. In order to evaluate the role of each source in the model’s perisaccadic performance, the ratios of perisaccadic to fixation performance were compared across the model variants described above. As seen in Fig 5C, top panel, the no-source model has the lowest perisaccadic to fixation performance ratio (0.28 ± 0.08). The perisaccadic to fixation performance ratio increases when any individual source is added to the model (+RF: 0.53 ± 0.09, +FF: 0.95 ± 0.14, +ST: 0.48 ± 0.09). The F-model which incorporates all three sources has the highest perisaccadic to fixation performance ratio (1.35 ± 0.09), and eliminating any individual source decreases the perisaccadic to fixation performance ratio (-RF: 1.20 ± 0.11, -FF: 0.72 ± 0.08, -ST: 1.27 ± 0.11). Fig 5C, bottom panel, summarizes these results in terms of what percent of the total improvement (the difference between the F-model and the no-source model) is present after adding or eliminating each of the sources. As seen, adding the RF, FF, or ST source improves the perisaccadic to fixation ratio by 23.38%, 62.65%, and 18.89%, respectively, and eliminating the RF, FF, or ST source decreases the ratio by 14.79%, 59.38%, and 7.87%, respectively.
From now on, “model” refers to the A-model whenever not otherwise specified.
As seen in Fig 6, the model mimics the perisaccadic responses at both the level of individual neurons and the subpopulation of neurons which display each perisaccadic modulation. The model predictions are shown for the same sample neurons depicted in Fig 1B–1E.
Fig 6A, top panel, shows the saccadic suppression effect observed in Fig 1B (left) as predicted by the model. For this neuron, the average model-predicted firing rate over the early response window (50–75 ms after stimulus onset) in response to an RF stimulus dropped from 22.69 ± 0.78 (mean ± SE) spk/s during fixation to 15.72 ± 1.72 (mean ± SE) spk/s when the stimulus appeared just prior to saccade onset (example neuron, p < 0.001). The average of the perisaccadic normalized responses for the subpopulation of neurons with a significant saccadic suppression effect in both the experimental data and model prediction are shown in Fig 6A, middle and bottom panels respectively (n = 7 out of 8). For this subpopulation, the average normalized response over the early response window dropped from 2.48 ± 0.42 to 1.38 ± 0.30 (mean ± SE), and the average predicted firing rate over the same response window decreased from 1.59 ± 0.16 to 1.06 ± 0.09 (mean ± SE). Fig 6B, top panel, shows the FF-remapping effect observed in Fig 1C (left) as predicted by the model. For this neuron, the average model-predicted firing rate over the late response window (80–150 ms after stimulus onset) in response to an FF stimulus increased from 8.01 ± 0.06 (mean ± SE) spk/s during fixation to 11.64 ± 0.36 (mean ± SE) spk/s when the stimulus appeared just prior to saccade onset (example neuron, p < 0.001). The average of the perisaccadic normalized responses for the subpopulation of neurons with a significant FF-remapping effect in both the experimental data and model prediction are shown in Fig 6B, middle and bottom panels respectively (n = 20 out of 23). For this subpopulation, the average normalized response over the late response window increased from 0.89 ± 0.03 to 1.36 ± 0.08 (mean ± SE), and the average normalized predicted firing rate over the same response window increased from 0.93 ± 0.02 to 1.19 ± 0.04 (mean ± SE). Fig 6C, top panel, shows the ST-remapping effect observed in Fig 1D (left) as predicted by the model. For this neuron, the average model-predicted firing rate over the late response window in response to an ST stimulus increased from 24.53 ± 0.39 (mean ± SE) spk/s during fixation to 47.24 ± 2.25 (mean ± SE) spk/s when the stimulus appeared just prior to saccade onset (example neuron, p < 0.001). The average of the perisaccadic normalized responses for the subpopulation of neurons with a significant ST-remapping effect in both the experimental data and model prediction are shown in Fig 6C, middle and bottom panels respectively (n = 24 out of 37). For this subpopulation, the average normalized response over the late response window increased from 0.79 ± 0.03 to 1.48 ± 0.07 (mean ± SE), and the average normalized predicted firing rate over the same response window increased from 0.92 ± 0.02 to 1.19 ± 0.04 (mean ± SE). Finally, Fig 6D, top row, shows the FF- and ST-remapping effects observed in Fig 1E (left column) as predicted by the model. For this neuron, the average model-predicted firing rate over late response window to FF and ST stimuli increased from 17.39 ± 0.38 (mean ± SE) spk/s and 15.85 ± 0.36 (mean ± SE) spk/s during fixation to 32.12 ± 2.50 (mean ± SE) spk/s and 33.11 ± 1.98 (mean ± SE) spk/s when the stimuli appeared just prior to saccade onset respectively (example neuron, p < 0.001 for both). The average of the perisaccadic normalized responses for the subpopulation of neurons with significant FF- and ST-remapping effects in both the response and model prediction are shown in Fig 6D, middle and bottom rows respectively (n = 17 out of 22). For this subpopulation, the average normalized response to FF and ST stimuli over the late response window increased from 0.88 ± 0.03 to 1.43 ± 0.09 (mean ± SE) and from 0.78 ± 0.04 to 1.56 ± 0.09 (mean ± SE) respectively, and the average predicted firing rate in response to FF and ST stimuli over the same response window increased from 0.92 ± 0.02 to 1.22 ± 0.05 (mean ± SE) and from 0.91 ± 0.02 to 1.23 ± 0.06 (mean ± SE) respectively.
Next, we compared the actual and predicted prevalence maps of the saccadic suppression, FF-remapping, and ST-remapping effects as a function of time from saccade and from stimulus onset. Fig 7A shows the experimental prevalence maps of saccadic suppression, FF-remapping, and ST-remapping (left to right, respectively); each plot shows the percent of neurons displaying the corresponding effect at each time point relative to saccade and stimulus onset. Fig 7B displays the same prevalence maps based on the model predictions. The frequency of occurrence and the timing of effects across the population are similar for the experimental data and the model prediction. This fact confirms that not only can the model replicate the perisaccadic effects well, but also it follows the dynamics (timing) of those perisaccadic modulations very closely. Strong correlations between the experimental and model-predicted values confirmed the model ability to replicate the timing and frequency of occurrence of effects across the population (saccadic suppression, r = 0.65, p < 0.001; FF-remapping, r = 0.60, p < 0.001; and ST-remapping = 0.62, p < 0.001; Pearson product-moment correlation). As detailed in the Methods section, the experimental prevalence maps are drawn based on the empirical firing rate values obtained by smoothing the observed spike trains, and accordingly have higher variability than the model-predicted firing rate values. This high variability means that fewer effects reach statistical significance, as reflected in the low values of the experimental maps in comparison with the model-predicted ones.
The model allows us to disassociate the contributions of multiple modulatory sources in generating the instantaneous firing rate of a neuron at each point in time relative to a saccade, which was not possible with any previous computational approach.
The contributions of the ST, FF, and RF sources to the neurons’ response are illustrated schematically in Fig 8A and 8B. Fig 8A shows how the three sources contribute differently to the neuron’s firing activity at different times relative to saccade onset. As seen, the evoked response in the neuron at each time point is due to stimuli presented in the ST, FF, and/or RF probe locations at different latencies relative to the response onset. In addition, each source contributes with a specific gain. The latency and gain corresponding to each contributing source vary over time. Fig 8B simplifies this idea in a schematic: multiple modulation sources (here: the ST, FF, and RF sources) contribute to the response generation at each instant of time with time-varying gains and latencies.
Fig 8C illustrates one’s neurons changing ability to detect stimuli presented at different locations over time relative to saccade. To quantify a neuron’s ability to detect stimuli, the receiver operating characteristic (ROC) curve analysis was used: for each location and latency, the detectability of the neuron was defined as the ROC between the model-predicted responses and the trial-shuffled responses (see Methods). The neuron’s maximum detectability at a specific location and the latency at which that maximum occurs correspond, respectively, to the gain and latency values shown in Fig 8A and 8B.
Fig 8C shows the detectability values of an example neuron for visual stimuli appearing in the RF, FF, and ST locations as a function of time relative to saccade and stimulus onset. The neuron can detect stimuli presented in the RF location with a latency of ~60 ms; it loses its ability to detect stimuli in the original RF location after the saccade (~60 ms after saccade onset). At almost the same time that the neuron loses the ability to detect stimuli at the RF location, the neuron becomes able to detect stimuli presented at the FF location with a latency of ~100 ms, and stimuli presented at the ST location with a latency of ~100 ms. Finally, at ~100 ms after a saccade, the neuron can detect stimuli presented in its new RF location (former FF location) with a latency of ~60 ms (the normal latency of MT neurons). In brief, Fig 8C shows that a visual neuron can detect stimuli presented at different locations at different latencies during the perisaccadic period, while it is only sensitive to the stimuli presented at the current RF location at normal latency during the fixation period. In order to make this point clear, Fig 8D examines the stimulus detectability and latency for different locations at a single time point relative to the saccade (responses measured at 55 ms after saccade) for the example neuron. As seen, peak detectability at the FF and ST locations occurs at longer latencies than peak detectability for the RF location; however, peak detectability values for the FF and ST locations are nearly equal to the original RF. So, the stimuli which appeared in the RF ~60 ms earlier and stimuli which appeared in the FF or ST ~100 earlier are equally detectable based on the neuron’s activity at that time point, demonstrating how multiple sources can simultaneously drive the neuron’s response with different latencies and gains.
The temporal evolution of the example neuron’s ability to detect stimuli at different locations, and the latency at which the neural response reflects the presence of a stimulus, are shown in Fig 8E, left panel. In this figure, the maximum detectability and the corresponding latency value for a single neuron for a stimulus presented at the RF, FF, or ST probe location are displayed. The stimulus detectability at the RF, FF, and ST probe locations is represented by the red, blue, and green circles, respectively. The x-axis represents time from saccade onset, and the y-axis represents the latency at which peak detectability occurs. The value of peak detectability at each time point is indicated by the size of circles. As seen, at first, the example neuron can only detect stimuli presented at the RF location ~60 ms earlier (red circles). Next, ~65 ms before the saccade, the neuron can also detect stimuli presented at the FF location ~90 ms earlier (blue circles). This ability to detect FF stimuli disappears 25 ms later, then reemerges ~13 ms after the saccade (with a latency of ~80–100 ms) and persists for ~60 ms. At that point the former FF becomes the RF, and the neuron responds to stimuli in this location with the normal latency of MT neurons, i.e. ~60 ms. Beginning ~80 ms after the saccade, the neuron can no longer detect stimuli presented at the RF location (no more red circles). Meanwhile, from 66 to 130 ms after the saccade, the example neuron detects ST stimuli which were presented ~130 ms earlier (green circles). Taken together, Fig 8E (left panel) confirms that while the neuron can only see the stimuli presented at the (former or new) RF location at normal latency during the fixation period, it can detect stimuli presented at different locations at different latencies during the saccade. To provide a sense of how Fig 8E, left panel, was generated, an animated 3D movie depicting the same information was created (S1 Movie: see Supporting Information). This movie shows the temporal evolution of the maximum detectability and the corresponding latency value for the same neuron shown in Fig 8E, left panel, in response to stimuli presented at the RF, FF, and ST probe locations as a saccade is prepared and executed. The detectability to the RF, FF, and ST locations is represented by the red, blue, and green discs, respectively. In this 3D space, the x and y values represent horizontal and vertical coordinates within the visual field, and the z-axis represents the latency at which peak detectability occurs. Time relative to the saccade changes over time (at a much slower speed), shown by the time label values on top of the plot. The color intensity of the discs represents the peak detectability value.
To understand how the neural detectability of FF and ST stimuli evolves over time in a population of neurons, synthetic neurons were constructed corresponding to the subpopulation average FF- and ST-remapping effects. These synthetic neurons were constructed by averaging the Gaussians of the neurons displaying each effect in both the response and model prediction (as detailed in the Methods section). These synthetic neurons illustrate the population time course of the changes in detectability across the saccade. Fig 8E, middle panel, illustrates how the stimulus detectability at the FF location changes over time relative to a saccade. From ~50 ms before to ~60 ms after a saccade, the population can detect FF stimuli with latencies of ~60–90 ms, indicating that the neural population is detecting stimuli appearing in the FF ~140 to ~0 ms before the saccade. This figure also shows that at around 60 ms after a saccade, the latency of the stimulus detectability at the FF location returns to ~60 ms (i.e., the RF latency value), indicating that the FF has become the new RF. Fig 8E, right panel, illustrates how the population’s ability to detect stimuli presented at the ST location changes over time. Similar to the FF-remapping effect, the population can detect ST stimuli at longer latencies (~100–120 ms) from 45 before to 15 ms after a saccade and from 74 to 114 ms after a saccade. 3D depictions of Fig 8E, middle and right panels, are presented as animated S2 & S3 Movies (see Supporting Information) respectively. These figures (and movies) reveal that FF and ST stimuli contribute to response generation well before saccade initiation, influencing neurons’ responses both during the saccade and after the eyes have landed.
We developed a time-varying model framework capable of precisely capturing fast changes in neuronal responses–induced by dynamic task and behavioral variables–at the resolution of individual neurons, spikes, and trials, as well as dissociating and quantifying the modulatory sources contributing independently to these changes. This new framework yielded several new insights into the coding of time-varying sensory information when applied to the spike trains in the visual cortex measured across rapid eye movements. We tested our model’s ability to capture and dissociate multiple time-varying sources of modulation in the context of perisaccadic changes in visual sensitivity. Neurons’ visual sensitivity changes dramatically around the time of saccades [3–6, 11–16]. The fast timescale on which these changes occur creates challenges both for experimental approaches and for computational models aiming to quantitatively characterize the relationship between various extrinsic or intrinsic system covariates and the modulations in sensitivity to visual input. Our approach combines high spatiotemporal resolution visual stimulation across many locations within the visual field of a neuron with a novel GLM-based model structure in order to accurately capture these fine timescale modulations. The model successfully predicted the responses of visual neurons in area MT, including during the perisaccadic period, at the level of single trials and with a temporal resolution beyond that of existing approaches. Moreover, the model could accurately reproduce the perisaccadic modulations observed in perisaccadic responses, including saccadic suppression, FF-remapping, and ST-remapping, at the level of both single neurons and the population; a goal which, to our knowledge, was unattainable using any previously reported computational framework. These modulations are consistent with previous neurophysiological studies; for example, the FF-remapping effect seen here is consistent with previous reports of memory remapping, but no predictive/anticipatory remapping, in MT [50–52].
The combination of the computational framework designed to capture dynamic changes in sensitivity and a high spatiotemporal resolution stimulus presentation paradigm allowed us to investigate how the spatiotemporal receptive field of a visual neuron evolves over time with a resolution on the order of <10 ms. Importantly, the stimulus presentation paradigm makes no assumptions about the time windows or spatial locations to which neurons will respond, and therefore in comparison with conventional experimental techniques, which mostly rely on a-priori selection of relevant stimulus locations and comparatively large temporal windows, offers both a more complete and higher spatiotemporal resolution picture of dynamic changes in neural sensitivity across saccades. The fitted kernels in the models represent the time-varying spatiotemporal receptive field structure of the neuron as captured by the models. Beyond merely tracking the changes in neurons’ RFs, this research provides a means to quantitatively study how multiple modulatory sources interact in generating the response of a neuron at each moment in time, and accordingly how the visual scene is encoded by neurons. This in turn can provide insight into the problem of how the pre- and post-saccadic scenes are integrated across saccades.
The computational framework developed in this paper opens up a plethora of opportunities for future research and applications; here we discuss some of these possible future uses of the model. First of all, the model can be used to examine the neural basis of psychophysical phenomena observed during eye movements. There are several reports indicating that saccades alter the perception of space and time [53–64]. Several research groups have tried to find explanations for these perceptual changes using abstract models designed to simulate the spiking behavior of visual neurons [21, 65–68]. The present computational framework provides a data-driven, more complete, and quantified description of perisaccadic response modulations, solely based on the statistical features of real spiking data and their relationship with a variety of covariates, allowing examination of the contribution of independent components to various perisaccadic perceptual phenomena. Secondly, the model can provide a read-out of visual information based on neural responses over the time course of a saccade. The choice of pseudorandom visual probe patterns across space and time generated an unbiased estimation of the encoding model, which is crucial for decoding arbitrary temporal and spatial information, which in turn has not been provided to the model during training. This model-based decoding of the spiking activity allows us to directly and quantitatively link neural activity to perception. Thirdly, the decomposition of modulatory effects provides tools for identifying the sources and consequences of their associated modulations. Sources can be selectively eliminated from the model to test their effects on the perceptual read-out, and the effect of source elimination on neural responses can be compared to the results of inactivation experiments. An important question in visual neuroscience is which area or areas in the brain are responsible for the perisaccadic changes in visual neurons. Our model provides an opportunity to explore this question by dissociating sources of modulation, quantifying their strength over time, and correlating them with the activity of various areas of the brain with the aim of finding which area might control each of the modulatory sources, and how inactivation of that area may alter perception across saccades. The current version of the F-model makes some assumptions about the different types of perisaccadic modulatory signals, based on previously reported neurophysiological effects; future work, however, will screen the S-models’ response functions to identify perisaccadic modulations without the assumptions built into the F-model. Finally, the proposed computational framework can serve to generate artificial populations of neurons and test their responses to a larger number of perisaccadic stimuli than would be experimentally feasible, allowing investigation of population-level visual representations. In total, the model presented here provides a blueprint for how sensory stimuli and internal covariates evoke different neural responses due to changes in the neural state. This framework can be applied not only to perisaccadic visual responses, but to a wide variety of brain areas and behavioral contexts in which sensory responses are modulated by combinations of factors such as attentional state, context, reward history, motor preparation, learned associations, and other cognitive variables.
Two adult male rhesus monkeys (Macaca mulatta) were used in this study. All experimental procedures were in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and the Society for Neuroscience Guidelines and Policies. The protocols for all experimental, surgical, and behavioral procedures were approved by the Montana State University Institutional Animal Care and Use Committee. Animals were pair-housed when possible and had daily access to enrichment activities. During the recording days, they had controlled access to fluids, but food was available ad libitum. All surgical procedures were carried out under Isoflurane anesthesia and strict aseptic conditions.
Prior to undergoing behavioral training, each animal was implanted with a stainless steel headpost, attached to the skull using orthopedic titanium screws and dental acrylic. All surgical procedures were carried out under Isoflourane anesthesia and strict aseptic conditions. Following behavioral training, custom-made PEEK recording chambers (interior 22 × 22 mm) were mounted on the skull and affixed with dental acrylic. Within the chamber a 22 × 22 mm craniotomy was performed above the extrastriate visual areas including areas V4 and MT (extrastriate craniotomies were centered at −6 mm A/P, 23 mm M/L and −13 mm A/P, 23 mm M/L).
Monkeys performed a visually guided saccade task during which task-irrelevant random dot stimuli flashed on screen. Two adult male rhesus monkeys were trained to fixate on a fixation point (FP; a red dot) located in the center of the screen. After they fixated, a saccade target (ST; a red dot) appeared 10 degrees away horizontally. Then, after a randomized time interval between 600 and 750 ms (drawn from a uniform distribution), the fixation point disappeared, cuing the monkeys to make a saccade to the ST. After remaining fixated on the ST for 600 ms monkeys received a reward. During this procedure, a series of randomly located probe stimuli were presented on the screen in a 9 by 9 grid of possible locations. Each stimulus was a white square (full contrast), 0.5 by 0.5 degree of visual angle (dva), against a black background. Each stimulus lasted for 7 ms and stimuli were presented consecutively without any overlap, such that at each time point there was exactly one stimulus on the screen. The locations of consecutive probe stimuli followed a pseudorandom order, called a condition. In each condition, a complete sequence of 81 probe stimuli was presented throughout the length of a trial. Conditions were designed to ensure that each probe location occurred at each time in the sequence with equal frequency across trials. The pseudorandom presentation of the probe stimuli made it possible to track the temporal evolution of the neurons’ spatiotemporal receptive field (RF) using an unbiased set of stimuli [69] and independent of their relative timing to the saccade events. For each recording session, the grid of the possible locations of the probes was positioned such that it covered the estimated pre- and post-saccadic receptive fields of the neurons under study, as well as the fixation point and saccade target. The spatial extent of the probe grids varied from 24 to 44.8 (mean ± SD = 30.25 ± 6.13) dva horizontally, and from 16 to 28 (mean ± SD = 18.11 ± 3.97) dva vertically. The (center-to-center) distance between two adjacent probe locations varied from 3 to 5.6 (mean ± SD = 3.78 ± 0.77) dva horizontally, and from 2 to 3.5 (mean ± SD = 2.26 ± 0.50) dva vertically. Each trial lasted between 2100 to 2300 ms. Throughout the entire course of the experiment, the spiking activity of the neurons in area MT was recorded using a 16-channel linear array electrode (V-probe, Plexon Inc., Dallas, TX) at a sampling rate of 32 kHz, and sorted offline using the Plexon offline spike sorter. The spike sorter program was employed to perform a principal component analysis, clusters of spikes with similar waveform properties were manually classified as belonging to a single neuron (single unit). The sorted spikes were then read into Matlab to verify the presence of a visually-sensitive receptive field. From a population of 49 well-isolated neurons, 8 neurons were discarded because they did not respond to any probe stimuli before and/or after the saccade, and the rest were used for analyses. The eye position of the monkeys was monitored with an infrared optical eye tracking system (EyeLink 1000 Plus Eye Tracker, SR Research Ltd., Ottawa, CA) with a resolution of < 0.01 dva, and a sampling frequency of 2 kHz. Stimuli presentation in the experiment was controlled using the MonkeyLogic toolbox [70]. Visual stimuli were presented on a 24-inch ASUS VG248QE LED monitor with a resolution of 1920X1080 pixels with a refresh rate of 144 Hz, positioned 28.5 cm in front of the animal’s eyes. A photodiode (OSRAM Opto Semiconductors, Sunnyvale CA), mounted on the lower left corner of the monitor, was used to record the actual onset and offset times of stimuli appearing on the screen with a continuous signal sampled and stored at 32 kHz. In total, data were recorded from 41 MT neurons during 11 recording sessions. In 9 recording sessions, the saccades were made to the left (27 of 41 neurons), and in 2 recording sessions, the saccades were made to the right (14 of 41 neurons). The positioning of the probe grids, the spatial distribution of the receptive field of the neurons, and the average of the photodiode signal for an example session is provided in the Supporting Information (S1 Fig). The experimental data are available at http://dx.doi.org/10.6080/K0FB514J and https://github.com/nnategh/SFA-Models.
The RF location refers to the probe location which generated the maximum firing rate during the fixation period (-500 to -100 ms from saccade onset) in the early response window, i.e. 50 to 75 ms from probe onset. The future field (FF) location was then set to the probe location shifted away from the RF probe in the direction and (rounded) size of the saccade vector. Finally, the saccade target (ST) location was defined as the probe location, out of the 4 × 4 probe locations centered around the ST, which generated the maximum firing rate during the perisaccadic period (-50 to 0 ms from saccade) in the late response window, i.e. 80 to 150 ms after probe onset, compared to the fixation period. To avoid overlap between the ST and FF locations, the FF probe location and all probe locations surrounding it were excluded from the potential ST probe locations (if they fell within the 4 × 4 probe locations centered around the ST).
In this paper, three perisaccadic response modulations were studied: saccadic suppression, FF-remapping, and ST-remapping. Saccadic suppression was defined as a significant decrease in the spike count of the neuron over the early response window (50–75 ms after stimulus onset) in response to a stimulus presented in the RF probe location shortly before a saccade (-30 to 0 ms from saccade onset), compared to the same stimulus presented during fixation (-500 to -100 ms from saccade onset). In the same way, the FF- and ST-remapping in a neuron were defined as increases in the spike count of the neuron in the late response window (80–150 ms after stimulus onset) to a stimulus presented in the FF or ST probe location shortly before a saccade (-50 to 0 ms from saccade onset) compared to the same stimulus presented during fixation (-500 to -100 ms from saccade onset). In all three cases, the statistical significance was tested by comparing the perisaccadic and fixation spike counts (during the same stimulus-aligned window) using the Wilcoxon one-sided signed-rank test, and a p-value of less than 0.05 (with no multiple comparison adjustment) was considered statistically significant.
Note that all statistical analyses were performed on spike counts between windows of the same duration (not estimates of firing rate). For graphical purposes, however, the spike trains were smoothed by convolving with a Gaussian with full width at half-max (FWHM) 13 ms. The response of a neuron to a stimulus presented in a given time interval was estimated by averaging the stimulus-aligned spike trains from 0 to 150 ms after stimulus onset. Due to differences in the mean firing activity of MT neurons, their perisaccadic and fixation responses were normalized by dividing by the grand mean of their firing rates (from 0–150 ms after onset of a stimulus, across both conditions) before averaging over a subpopulation of neurons. The grand mean was defined as the mean of the means of the perisaccadic and fixation responses.
We developed a sparse-variable generalized linear model framework, termed the S-model, which is able to track the saccade-induced rapid changes in the spatiotemporal sensitivity of the neurons on the order of 7 milliseconds (which is the resolution of stimuli presentation). The principal idea of the S-model is that the stimulus-response relationship in a neuron is characterized by a set of two-dimensional stimulus kernels (k(t,τ)), which represent the spatiotemporal receptive field of the neuron as varying along the time dimension (t). Fixing the stimulus kernels along the time dimension results in the conventional one-dimensional stimulus kernels (k(τ)) used in ordinary generalized linear models (GLMs) [43, 71]. More specifically, the conditional intensity function (CIF) of the S-model, representing the instantaneous firing rate of an MT neuron, i.e., λ(l)(t), under our experimental paradigm is described by,
λ(l)(t)=f(∑x,y,τkx,y(t,τ).sx,y(l)(t−τ)+∑τh(τ).r(l)(t−τ)+b(t)+b0),
(1)
where sx,y(l)(t)∈{0,1} denotes a temporal sequence of probe stimuli presented at probe location (x,y) on trial l where 0 and 1 represent, respectively, an off and on probe condition, r(l)(t)∈{0,1} indicates the spiking response of the neuron on that trial, kx,y(t,τ) represents the two-dimensional stimulus kernel corresponding to the stimulus sequence being presented at probe location (x,y), h(τ) is the post-spike kernel applied to the spike history, b(t) is the offset kernel which captures the saccade-induced changes in the baseline activity (activity in the absence of a visual stimulus and feedback from spiking responses) of the neuron over the time course of the experiment, b0 = f−1(r0) where r0 is defined as the measured mean firing rate (spikes per second) across all trials in the experimental session, and finally,
f(u)=rmax1+e−u,
(2)
is a static sigmoidal function representing the response nonlinear properties where rmax indicates the maximum firing rate of the neuron obtained empirically from the experimental data. Compared with the empirical nonlinearity estimated nonparametrically from data, this choice of model nonlinearity adequately captured the nonlinear properties of the neurons’ response. All trials were saccade aligned, i.e., t = 0 refers to the time when a saccade was initiated. In order to reduce the high dimensionality of the problem, all the kernels were parameterized as a linear combination of a set of basis functions defined across time, delay, or both variables as follows,
kx,y(t,τ)=∑i,jκx,y,i,j.Bi,j(t,τ),
(3)
h(τ)=−∑iηi2.Hi(τ),
(4)
b(t)=∑jβj.Oj(t),
(5)
where {κx,y,i,j}, {ηi}, and {βj} are the basis parameters of the stimulus kernels, post-spike kernel, and offset kernel respectively. Since the basis functions {Hi(τ)} were set to be positive and the post-spike kernel was set to reflect only the response refractory effects [72], the negative square of the basis parameters {ηi} were used such that the resulting h(τ) produced a non-positive kernel. The basis functions representing the two-dimensional stimulus kernels were set as follows,
Bi,j(t,τ)=Ui(τ)Vj(t),
(6)
where Ui(τ) and Vj(t) were chosen to be B-spline functions of order two. {Ui(τ)} span over the delay variable τ, representing a 150 ms-long kernel using a set of 26 knots uniformly spaced at {-13, -6, …, 155, 162} ms (in total, 23 basis functions), and {Vj(t)} span over the time variable t, representing a 1081 ms-long kernel centered at the saccade onset using a set of 159 knots uniformly spaced at {-554, -547, …, 545, 552} ms (in total, 156 basis functions). The basis functions {Hi(τ)} representing the post-spike kernel were chosen to be B-spline functions of order two with non-uniformly distributed 23 knots over the delay variable τ; the spacing of the knots around zero, which indicates the spike time, was smaller and increased further away from the spike time and its associated refractory period (the knots were spaced at {1, 2, 3, 4, 6, 8, 15, 22, 29, 36, 43, 50, 57, 64, 71, 78, 92, 106, 120, 134, 148, 162, 176} ms, in total 20 basis functions). Finally, {Oj(t)} span over the time variable t, representing a 1081 ms-long kernel centered at the saccade onset using a set of 77 knots uniformly spaced at {-570, -555, …, 555, 570} ms (in total, 74 basis functions). A visualization of basis functions is presented in the Supporting Information (S2 Fig). From (Eqs 1 and 3–5) together,
λ(l)(t)=f(∑x,y,τ,i,jκx,y,i,j.Bi,j(t,τ).sx,y(l)(t−τ)−∑τ,iηi2.Hi(τ).r(l)(t−τ)+∑jβj.Oj(t)+b0).
(7)
Eq (7) denotes the CIF of the spiking process described by the S-model. The probability of a spike train associated with this Poisson process is thus given by,
p(r(l)|s)=∏tp(r(l)(t)|s)∝∏t(λ(l)(t).Δ)r(l)(t)e−λ(l)(t).Δ,
(8)
where s is the sequence of input stimuli, and r(l) = {r(l)(t)} represents the sequence of binned spike counts with bins of size Δ ms on trial l. Here, the bin size was chosen equal to 1 ms which ensures that at most one spike can fall in each time bin. The point process log-likelihood (LL) [44] of the observed spike trains given the model is,
LL({κx,y,i,j},{ηi},{βj})=∑l,t(r(l)(t).log(λ(l)(t).Δ)−λ(l)(t).Δ).
(9)
{κx,y,i,j}, {ηi}, and {βj} are estimated by maximizing the log-likelihood function given in Eq (9). The choice of nonlinear function in the S-model’s CIF made the problem of log-likelihood maximization non-convex; the block coordinate ascent method was used to solve the optimization problem (detailed later in this subsection).
To avoid overfitting despite the high dimensionality of the S-model, multiple computational approaches were adopted. First, representing each model kernel using a linear combination of smooth basis functions resulted in an optimization process in a lower dimensional space and with well-behaved search paths. Second, a parameter selection strategy was used to identify the subset of parameters most important for mediating the stimulus-response relationship; only those parameters were included in the model fitting procedure, in order to reduce the high dimensionality of the problem. In this strategy, the basis parameters {κx,y,i,j}, which comprise the majority of the model parameters, were ranked according to their significance in response prediction, and the less significant ones were eliminated by the following procedure: first, the model’s CIF was assumed to be obtained by a single κx,y,i,j and with no dependency on the spike train history or the fluctuations in the baseline activity (i.e., without post-spike and offset kernels). To fit this simplified model, an MLE procedure was performed over 100 subsets of the data, each one obtained by randomly selecting 65% of the data (35% as training and 30% as validation) to generate a distribution of the estimated values for κx,y,i,j. To evaluate the significance of κx,y,i,j, a control distribution of this parameter was constructed using the same strategy but with a set of shuffled responses. A κx,y,i,j parameter was assessed as a significant parameter for response prediction if the mean of its distribution (μx,y,i,j) satisfied the following condition:
|μx,y,i,j−μ¯x,y,i,j|≥1.5σ¯x,y,i,j,
(10)
where μ¯x,y,i,j, and σ¯x,y,i,j are the mean and standard deviation of the control distribution. Those κx,y,i,j parameters that were detected as significant were included in the model fitting and all others were set to zero.
Lastly, to help prevent overfitting of the S-model to the training dataset, a cross-validation approach was used to regularize the model parameters. In this approach, the data were randomly split into a training set (35%), a validation set (30%), and a test set (35%). Then, in order to estimate the model parameters, the likelihood function in Eq (9) was maximized using the block coordinate ascent method as follows: (initialization) The model parameters were set to a very small non-zero value (here: 10−6, in order to have a non-zero gradient); (step 1) The likelihood function was maximized with respect to the selected parameters (chosen based on the parameter selection strategy detailed above) describing the stimulus kernel corresponding to the probe location (x,y) = (1,1) while the rest of model parameters were held fixed at their current estimates. The selected parameters were iteratively updated (using only the training data) to maximize the log-likelihood function over both the training and validation data, until the relative change in the root mean of squares of the selected parameters was less than 1%. Then the same procedure was repeated for the parameters describing the stimulus kernel corresponding to the next probe location, until all probe locations were completed. (step 2) the same as step 1, but for the parameters describing the post-spike kernel. (step 3) the same as step 1, but for the parameters describing the offset kernel. (step 4) steps 1–3 were repeated until no update in the model parameters was observed during these steps. Then, the estimation process was terminated, and fitting was considered complete. The test data were withheld from the model fitting procedure and were used to measure the model goodness-of-fit, ensuring the generalizability of the fitted model to unseen data. The block coordinate ascent method gave a stable solution for the data and with regard to different initializations of the parameters, and benefited the convergence time of the estimation procedure.
Employing basis functions as well as the parameter selection strategy reduced the dimensionality of the parameter space; this reduced dimensionality, along with the cross-validation approach, provided a detailed map of the neuron’s time-variant spatiotemporal sensitivity with less concern of overfitting. A visualization of a few sample stimulus kernels, post-spike kernels, and offset kernels obtained from the S-model fitting are provided in the Supporting Information (S3 Fig).
Although the S-model can capture the dynamics of the spatiotemporal receptive field, and thus can characterize the response modulation on the timescale of a saccade, it does not identify explicitly what sources contribute to the response modulation. The idea of identifying modulatory sources was inspired by the perisaccadic modulations observed in our experimental data and in many previous studies (see Introduction), i.e. saccadic suppression, FF-remapping, and ST-remapping. In fact, the stimulus kernels fitted using the S-model, S-kernels, can be approximated by time- and delay-dependent mixtures of spatial skewed Gaussians where each Gaussian captures the response modulation arising from one of the RF, FF, or ST sources at a given time and delay (for more details, see S4 Fig in the Supporting Information). To quantitatively dissociate the effects of the RF, FF, and ST sources at different times and delays, a factorized sparse-variable generalized linear model, called the F-model, was developed which approximates the fitted S-kernels kx,y(t,τ) by k^x,y(t,τ) as:
k^x,y(t,τ)=k˜x,y(τ)+∑srG(x,y;φsr(t,τ))+c(t,τ),
(11)
where k˜x,y(τ), termed the fixation kernel, represents the average spatiotemporal receptive field of the neuron over the fixation period, defined as
k˜x,y(τ)=1t2−t1∑t=t1t2kx,y(t,τ);
(12)
where t1 = −400 ms, and t2 = −300 ms from saccade; c(t,τ) represents the time- and delay-dependent baseline profile which is uniform across spatial dimensions; and finally, each G(x,y;φsr(t,τ)) indicates a time- and delay-dependent spatial skewed Gaussian representing the modulation source sr∈{RF,FF,ST}, which is parameterized by φsr(t,τ)={asr(t,τ),μxsr(t,τ),μysr(t,τ),σxsr(t,τ),σysr(t,τ),ρsr(t,τ),γxsr(t,τ),γysr(t,τ)} as follows,
G(x,y;φsr(t,τ))=asr(t,τ).e−12(1−ρsr(t,τ)2){(x−μxsr(t,τ))2σxsr(t,τ)2+(y−μysr(t,τ))2σysr(t,τ)2−2ρsr(t,τ).(x−μxsr(t,τ))(y−μysr(t,τ))σxsr(t,τ).σysr(t,τ)}.Φ(γxsr(t,τ)(x−μxsr(t,τ))).Φ(γysr(t,τ)(y−μysr(t,τ))),
(13)
where asr(t,τ), (μxsr(t,τ),μysr(t,τ)), (σxsr(t,τ),σysr(t,τ)), ρsr(t,τ), and (γxsr(t,τ),γysr(t,τ)) represent the amplitude, the x – and y-coordinate of the center, the horizontal and vertical spread, the orientation, and the horizontal and vertical skewness of the Gaussian kernel G(x,y;φsr(t,τ)) corresponding to the modulation source sr at time t and delay τ, and Φ(∙)indicates the standard normal cumulative distribution function.
The time- and delay-dependent parameters φsr(t,τ) and c(t,τ) were estimated by minimizing the sum square difference between the F-kernel k^x,y(t,τ), as specified in Eq (11), and the S-kernel kx,y(t,τ) obtained from the S-model for each time t and delay τ. In order to eliminate noise included in the S-kernels as well as to alleviate overfitting, the Gaussian parameters were estimated over non-overlapping bins across the delay dimension (instead of single values of delay), i.e. φsr and c corresponding to time t from saccade and delays τb≤τ<τb+1 were estimated by minimizing
∑x,y∑τb≤τ<τb+1(k˜x,y(τ)+∑srG(x,y;φsr)+c−kx,y(t,τ))2
(14)
where {τb} = {1, 20, 40, 50, 53, 56, 59, 62, 65, 68, 71, 74, 77, 80, 85, 90, 95, 100, 105, 110, 115, 120, 125, 130, 135, 140, 145, 151} ms. In order to constrain each Gaussian kernel within an area surrounding the modulation source which the Gaussian kernel represented (i.e. surrounding the RF, FF, or ST probe location as defined before), and to avoid overlap between Gaussian kernels, the Gaussian parameters were subject to a set of bounded limits during the estimation process, such that (1) the center of Gaussian kernels could only move one probe location away from the probe location corresponding to the effect they were intended to represent (i.e., one probe location away from the RF probe location for the center of the Gaussian kernel for the RF, etc.); (2) the horizontal and vertical spread of the Gaussian kernels could not exceed 2 probe locations; (3) their orientations were limited to between -1 and +1; and finally, (4) the absolute values of the horizontal and vertical skewness of the Gaussian kernels were bounded to 5. In order to remove the discontinuities created by estimating the F-kernels over delay bins, the reconstructed F-kernels were smoothed in the delay dimension using a moving average filter with a span of 10 ms. As a result, the F-kernels had lower values in comparison with the corresponding S-kernels, and so, the firing rates predicted by the F-model were lower in magnitude compared to the corresponding S-model.
The models’ performance was evaluated over test data, which was used neither for training the model parameters nor for validating the fitted ones, in terms of the log-likelihood of the observed spike trains given the estimated instantaneous firing rate. In order to estimate the instantaneous firing rates, the sequences of stimuli presented to the neuron were given to the model according to Eq (1). To simulate the effect of spiking history, the recorded spike trains were used. Using the true spike history always raises the concern of getting an unfairly good estimate for the model performance due to the presence of a strong refractory period or self-excitatory component in the post-spike kernel. However, neither of these was an issue here: the post-spike kernels were defined to reflect only the refractory effects (as detailed in the “S-model framework” subsection), and in this dataset only a small number of MT neurons fired at rates in which the absolute refractory period came into play. As a result, including the post-spike kernels had no significant impact on the model performance, as demonstrated by comparing the model performance using the true history, the simulated history (as detailed in [73]: algorithm 2), and no history. The results indicated that there is no significant difference in the model performance measures between these three scenarios (S5 Fig in the Supporting Information). Throughout this paper, the true history was used whenever the model was employed to simulate the recorded spiking responses (Figs 3 and 5–7, except Fig 5B).
The log-likelihood evaluates how well the spike times are predicted by the model, and can do so at the level of individual trials. In the log-likelihood formula (see Eq (9)), the first term is larger when the spikes are observed at high values of the estimated firing rate, and the second term is larger when there are fewer spikes at low values of the predicted firing rate. The log-likelihood was normalized by spike counts, as reported in Williamson et al. [74] and Cui et al. [72], to indicate the amount of information being conveyed by individual spikes. One problem with the log-likelihood is understanding the meaning of a given log-likelihood value in isolation. To address this issue, the log-likelihood of the models were compared with the log-likelihood of a NULL model in which the instantaneous firing rate of the neuron was set to its average firing rate; so, the amount of improvement compared to the NULL model normalized by spike counts, i.e. the log-likelihood per spike (ΔLL/spk), was reported instead of the raw likelihood value.
To prove that the S- and F-models better predict the perisaccadic responses compared to alternative models, the quality of perisaccadic response prediction by the S- and F-models was compared with the N-model -which is the state-of-the-art model in the perisaccadic response modeling as detailed in [48]- in terms of the log-likelihood per spike. For each model, the log-likelihood of the perisaccadic spikes was scatter plotted vs. the log-likelihood of the fixation spikes for the population of 41 MT neurons (Fig 5A). Because the stimuli presented during the perisaccadic period (-50 to 0 ms from saccade onset) evoke responses after 50–150 ms from stimulus onset, the perisaccadic spikes were defined as those spikes observed from 0 to 150 ms after saccade onset. By the same reasoning, the fixation spikes were defined as those spikes observed from 450 to 0 ms before saccade onset.
In order to robustly capture the fast perisaccadic modulations (saccadic suppression, FF-remapping, and ST-remapping), it is critical to reduce the variability of the model prediction over the limited perisaccadic data; for that purpose, an aggregate version of the F-model, termed the A-model, was developed. The A-model was constructed by (1) fitting the S-model ten times as explained earlier, on different randomly selected subsets of the data, (2) fitting ten F-models corresponding to each fitted S-model, (3) averaging the model variables (φsrs, and c) obtained from each F-model in order to gain a set of aggregate variables, and finally, (4) constructing the A-model kernels using the aggregate variables through Eq (11). Our code for estimating the S-, F-, and A-model parameters and performing model evaluation analysis is provided at https://github.com/nnategh/SFA-Models.
After confirming the goodness-of-fit of the S- and F-models over test data, all trials (and not only those withheld for model testing) for each cell were employed for the rest of the analyses in this study to provide more accurate empirical measures from the experimental data. In order to evaluate the models’ ability to capture the perisaccadic effects observed in the experimental data, i.e. saccadic suppression, FF-remapping, and ST-remapping, sequences of stimuli (n = 1000) were presented to the fitted (N-, S-, F-, and A-) models and the corresponding sequences of instantaneous firing rates were generated (based on Eq (1)). The spatial frequency and timing of the stimulus sequences were the same as those for the experimental paradigm. To model the effect of spiking history, the algorithm developed by Chen et al. (algorithm 2, [73]) was employed to simulate the spike history (self-history). Then, the saccadic suppression effect was defined as a significant decrease in the mean firing rate predicted by the model over the early response window (50–75 ms after stimulus onset) in response to a stimulus being presented in the RF probe location shortly before a saccade (-30 to 0 ms from saccade onset) compared to the same stimulus presented during fixation (-500 to -100 ms from saccade onset). The FF- and ST-remapping were defined as significant increases in the mean firing rate predicted by the model in the late response window (80–150 ms after stimulus onset) in response to a stimulus being presented in the FF or ST probe location shortly before a saccade (-50 to 0 ms from saccade onset) compared to the same stimulus presented during fixation (-500 to -100 ms from saccade onset). In all cases, the statistical significance was tested by comparing the perisaccadic and fixation predicted firing rates using the one-sided Wilcoxon rank-sum test. A p-value less than 0.05 was considered statistically significant.
The model-predicted firing rate in response to a stimulus presented in a given time interval was estimated by averaging the stimulus-aligned sequences of firing rates predicted by the model from 0 to 150 ms after stimulus onset. The perisaccadic and fixation firing rates predicted by the model were normalized by dividing by the grand mean of the model-predicted firing rates (from 0–150 ms after onset of a stimulus, across both conditions) before averaging over a population of neurons. The grand mean was defined as the mean of the means of the perisaccadic and fixation firing rates predicted by the model.
A dichotomous analysis was performed for each model and for each perisaccadic effect to assess in how many neurons the effect is significantly observed/not observed in both the experimental data and the model prediction (as detailed in the “Statistical analysis of model predictions” subsection). To this purpose, for each model (N-, F-, S-, and A-) and each perisaccadic effect (saccadic suppression, FF-remapping, ST-remapping), a confusion matrix was constructed which reported the presence or absence of the statistical significance of an effect measured from the experimental data versus from the model prediction. The confusion matrix contained the following values: the number of neurons in which the corresponding effect was statistically significant in both the response and the model (true positive, TP), statistically significant in the response but not the model (false negative, FN), statistically significant in the model but not the response (false positive, FP), and statistically significant in neither the response nor in the model (true negative, TN). There are multiple measures to evaluate the overall classification accuracy of a model using a confusion matrix, including sensitivity, accuracy, and precision. The agreement between the experimental data and the model prediction in terms of displaying/not displaying each of the perisaccadic effects was assessed using sensitivity, accuracy, the geometric mean of the sensitivity and precision (GSP), and F-measure (with α = 1/2), defined as follows:
sensitivity=TPTP+FN
(15)
accuracy=TP+TNTP+FN+FP+TN
(16)
precision=TPTP+FP
(17)
GSP=sensitivity.precision
(18)
F‐measure=(1+α2).sensitivity.precisionα2.precision+sensitivity
(19)
Note that the dichotomous analysis was performed over the entire data set (including the training and validation trials, as mentioned before) in order to use sufficiently large number of trials recorded for each cell for measuring firing rates. The McNemar test [49] was employed to evaluate the statistical significance of different models’ abilities to classify neurons as displaying each of the perisaccadic effects.
In order to evaluate a model’s ability to capture perisaccadic modulations specifically, the perisaccadic performance to fixation performance ratios of the models were calculated and compared. The performance was quantified in terms of ΔLL/spk (as detailed in the “Model evaluation” subsection), resulting in a unitless performance ratio value. Since the perisaccadic modulations are only measurable if a stimulus appears in one of the RF, FF, or ST probe locations during the perisaccadic period, the perisaccadic and fixation performance values were calculated over those trials from the test data which met this criterion.
The perisaccadic to fixation ratio was employed to compare the S- and F-models. In addition to the S- and F-models, different variants of the F-model were created to assess the contribution of individual sources, represented by time- and delay-dependent spatial skewed Gaussian kernels, to the F-model’s performance. For this purpose, one, two, or all of the sources were eliminated by nulling the parameters (φsr) of the relevant Gaussian kernel. To null the parameters of a Gaussian kernel, its parameters were replaced by parameters randomly selected from the fixation period (400 to 300 ms before saccade). By eliminating the RF, FF, or ST Gaussian kernel, the F-model without an RF, FF, or ST source was constructed (-RF, -FF, -ST respectively). By preserving only the RF, FF, or ST Gaussian kernel, the F-model with only the RF, FF, or ST source was constructed (+RF, +FF, +ST respectively). By eliminating all of the RF, FF, and ST Gaussian kernels, the F-model with no source was constructed (no-source).
In order to handle outliers created when taking the ratio, the perisaccadic performance values of the population of neurons were regressed as a linear function (with a fixed intercept of zero) of the fixation performance values for a model. The slope of the fitted line (using a robust-fit algorithm to remove outliers) was considered as an estimate for the perisaccadic to fixation ratio for the corresponding model. S6 Fig shows the regression fits and values. For each of the partial models as well as the F-model, the perisaccadic to fixation performance ratio obtained by this way are plotted in Fig 5C, top panel. The error bars indicate the standard errors of the estimated slopes.
Finally, in order to quantify how much adding one source contributes to the performance of the F-model, the increase percentage was defined for the +RF, +FF, and +ST models as the difference between the performance of the corresponding model and the no-source model divided by the difference between the performance of the F-model and the no-source model, stated as a percentage. Also, to quantify how much eliminating one source reduces the performance of the F-model, the decrease percentage was defined for the -RF, -FF, and -ST models as the difference between the performance of the corresponding model and the F-model divided by the difference between the performance of the no-source model and the F-model, stated as a percentage. All instances of the ‘model performance’ here refer to the perisaccadic to fixation performance ratio.
The prevalence of the perisaccadic effects at different times relative to a saccade as well as stimulus onset was assessed for both the experimental data and the model prediction (Fig 7).
For the saccadic suppression effect, the prevalence map using the experimental data was constructed as follows: for each neuron, the set of measured firing rate values generated by the neuron in response to a stimulus presented at the RF probe location at time t relative to the saccade and latency τ from stimulus onset was collected, r^t,τ. In the same manner, the set of measured firing rate values generated by the neuron in response to the same stimulus presented during fixation (-500 to -100 from saccade) was collected for the latency of τ, called r¯τ. Then, mt,τ, a map of 0’s and 1’s, was built for the corresponding neuron:
mt,τ={1,ifr^t,τissignificantlylessthanr¯τ0,otherwise
(20)
By averaging all mt,τs obtained from those neurons displaying saccadic suppression in both the response and the model prediction, the prevalence map of the saccadic suppression effect was generated. This map illustrates the percentage of the neurons displaying suppression at each time point relative to the saccade as well as to the stimulus onset. The same approach was employed to draw the prevalence map for the FF- and ST-remapping effects using the experimental data, except that mt,τ was built as:
mt,τ={1,ifr^t,τissignificantlygreaterthanr¯τ0,otherwise
(21)
Similarly, the prevalence maps were drawn for each of the perisaccadic effects using the model predictions. In the model-based prevalence maps, the instantaneous firing rate predicted by the model was used instead of the measured firing rate employed in the data-based maps. The instantaneous firing rate of a neuron was measured by smoothing spike trains through convolving with a Gaussian with FWHM 33 ms. The statistical significance was tested by the one-sided Wilcoxon rank-sum test. A p-value less than 0.05 was considered statistically significant. To quantify the similarity between the data-based and model-based prevalence maps for each perisaccadic effect, a Pearson product-moment correlation was used.
The receiver operating characteristic (ROC) curve analysis [75] was used in this paper to investigate how the spatiotemporal detectability of a neuron changes over time relative to a saccade. For this purpose, sequences of stimuli (n = 1000) were presented to a fitted A-model, and the corresponding sequences of instantaneous firing rates were generated (based on Eq (1)). The spatial frequency and timing of the stimulus sequences were as the same as those for the experimental paradigm. To model the effect of spiking history, the algorithm developed by Chen et al. (algorithm 2, [73]) was employed to simulate the spike history. Then, for each probe location (x,y), each time point t relative to saccade, and each latency τ, the set of predicted firing rate values at time point t was collected if a stimulus was presented at probe location (x,y) at time t−τ; this distribution of firing rates was called rx,y,t,τ. This procedure was repeated after pairing the same sequences of firing rate with shuffled sequences of stimuli, in order to have an estimate of the firing rate distribution when there is no causal relationship between the stimuli and responses (NULL hypothesis); this shuffled distribution of firing rates was called r˜x,y,t,τ. Afterwards, an ROC curve was constructed based on the rx,y,t,τ and r˜x,y,t,τ values, and the area under the curve was considered as the detectability of the neuron, roc(x,y,t,τ), to probe location (x,y) at time t when a stimulus was presented there τ ms earlier (or equivalently, with a latency of τ ms). The detectability of a neuron indicates how reliably a stimulus presented at a given probe location at a given time and latency can be detected by the neuron. To assess the significance of the detectability value, the ROC between two distinct shuffled firing rate distributions was calculated 100 times to obtain a null detectability distribution. The model-predicted detectability value and null detectability distribution were compared (two-sample t-test), and considered statistically significant if p was less than 10−9.
To better visualize the changes in neurons’ detectability over time relative to the saccade, a plot showing the maximum detectability and the corresponding latency over time is shown for an example neuron (see Fig 8E, left panel). For this purpose, at each time point t relative to saccade, the maximum detectability of the neuron to each probe location (x,y), Ix,y,t, and the corresponding latency at which the maximum detectability occurred, Tx,y,t, were determined as follows:
Ix,y,t=maxτ{roc(x,y,t,τ)},
(22)
Tx,y,t=argmaxτ{roc(x,y,t,τ)},
(23)
if Ix,y,t was a statistically significant detectability value. Then, Tx,y,t for the RF probe location (red), FF probe location (blue), and one of the closest probe locations to the ST (green) were plotted vs. time to saccade, and the maximum detectability reached at each time point, Ix,y,t, was indicated by the size of the markers. Tx,y,t and Ix,y,t were smoothed over time with a moving average filter spanning 7 ms. Note that values of Ix,y,t were included only for corresponding Tx,y,t values greater than 50 ms, i.e. the maximum detectability had to occur at or after the normal latency of MT neurons. In addition to this plot, an animated movie was created (S1 Movie: see Supporting Information) for the same neuron. In this movie, for each time point t, the corresponding Tx,y,t was visualized by the height of a disc centered at probe location (x,y) in a 3D space. In this 3D space, the x- and y-axis represent the horizontal and vertical position in the visual field, and the z-axis represents the latency at which the maximum detectability to a probe location occurs. The amount of the maximum detectability, Ix,y,t, is represented by the intensity of the discs’ color. Only the detectability to the three main probe locations displayed in the plot (indicated by color) is shown here.
Finally, in order to have a sense of how the perisaccadic effects were displayed by the population of MT neurons, similar plots and movies were generated for artificially generated neurons representing the population statistics (see Fig 8E, middle and right panels; and S2 and S3 Movies in the Supporting Information). A population analysis was done as follows: the variables (φsrs, and c) obtained from the A-models of the neurons in which a perisaccadic effect was significantly observed in both the response and model prediction were normalized by the distance between the adjacent probes, and averaged. Then, a synthetic neuron was built for each perisaccadic effect using the variables corresponding to the relevant Gaussian kernel (for example, the average of the normalized φFFs was used to construct the kernels corresponding to a synthetic neuron representing the FF-remapping effect, while φRFs and φSTs were set to zeros for that neuron). Then, the same procedure used for the sample neuron was employed to produce a 2D plot as well as an animated movie for that effect (p-value less than 10−7 was considered significant). Due to the low number of neurons displaying the saccadic suppression effect in both the response and model prediction (n = 7), this effect was excluded from the analysis.
|
10.1371/journal.pcbi.1002976 | Automated Analysis of a Diverse Synapse Population | Synapses of the mammalian central nervous system are highly diverse in function and molecular composition. Synapse diversity per se may be critical to brain function, since memory and homeostatic mechanisms are thought to be rooted primarily in activity-dependent plastic changes in specific subsets of individual synapses. Unfortunately, the measurement of synapse diversity has been restricted by the limitations of methods capable of measuring synapse properties at the level of individual synapses. Array tomography is a new high-resolution, high-throughput proteomic imaging method that has the potential to advance the measurement of unit-level synapse diversity across large and diverse synapse populations. Here we present an automated feature extraction and classification algorithm designed to quantify synapses from high-dimensional array tomographic data too voluminous for manual analysis. We demonstrate the use of this method to quantify laminar distributions of synapses in mouse somatosensory cortex and validate the classification process by detecting the presence of known but uncommon proteomic profiles. Such classification and quantification will be highly useful in identifying specific subpopulations of synapses exhibiting plasticity in response to perturbations from the environment or the sensory periphery.
| Synaptic connections are fundamental to every aspect of brain function. There is growing recognition that individual synapses are the key sites of the functional plasticity that allows brain circuits to store and retrieve memories and to adapt to changing demands and environments. There is also a growing consensus that many neurological, psychiatric, neurodevelopmental and neurodegenerative disorders may be best understood at the level of specific, proteomically-defined synapse subsets. Here, we introduce and validate computational analysis tools designed to complement array tomography, a new high-resolution proteomic imaging method, to enable the analysis of diverse synapse populations of unprecedentedly large size at the single-synapse level. We expect these new single-synapse classification and analysis tools to substantially advance the search for the specific physical traces, or engrams, of specific memories in the brains synaptic circuits. We also expect these same tools to be useful for identifying the specific subsets of synapses that are impacted by the various synaptically-rooted afflictions of the brain.
| Synapses are fundamental to every aspect of brain function. They are recognized today as being highly complex structures and highly diverse in both function and molecular composition. At the structural level, individual synapses of the mammalian central nervous system are thought to comprise hundreds of distinct protein species [1]–[3], and genomic and gene expression data available implies very strongly that there are multiple isoforms of many of these proteins and that their expression is differentially patterned across the brains diverse cell types [4]. It thus seems inescapable that synapses of the brain, even within traditional transmitter-defined synapse categories (e.g., glutamatergic, GABAergic, cholinergic, etc.), must be highly diverse in protein composition [5]. This conclusion is consistent with the available functional data, where physiological studies report wide differences in synaptic transmission as different brain regions and pathways are examined (again, even when results are compared only within traditional neurotransmitter categories). Moreover, the well-known functional plasticity of both synapse structure and synapse function in response to electrical activity implies directly that even an otherwise homogeneous synapse population must become heterogeneous or diverse after individual synapses experience differential activity. In this light, it seems likely that synapse diversity per se may be critical to the proper function of neural circuitry. For instance, there is now widely believed that the plasticity (and therefore resulting diversity) of individual synapses is fundamental to memory storage and retrieval and to many other aspects of neural circuit adaptation to environmental change [6], [7].
Unfortunately, the measurement of synapse diversity has been restricted by the limitations of available methods capable of resolving individual synapses. Array tomography (AT) is a new high-resolution, high-throughput proteomic imaging method that has the potential to very substantially advance the measurement of unit-level synapse diversity across large and diverse synapse populations. AT uses multiple cycles of immunohistochemical labeling on thin sections of resin-embedded tissue to image the proteomic composition of synapse-sized structures in a depth-invariant manner. We have applied AT to freshly-fixed mouse cerebral cortex, where our volumes have typical sizes of thousands to millions of of tissue, contain millions of individually-resolved synapses, and label over a dozen multiplexed proteomic markers.
With proper analysis, the informational density of array tomographic volumes has numerous potential applications. Synapse-level resolution of large volumes of tissue is an ideal tool for addressing interesting hypotheses concerning principles like synaptic scaling [6], structural arrangement [8], and novel synapse types [9], [10]. Combined with connectomic data [11], [12], genetic models [13], [14] or dye filling techniques [15], [16], array tomography can also address questions regarding proteomic distributions in specific subsets of cells. We are interested in investigations of this nature and others in the mouse cerebral cortex, where the anatomical distribution of synapses, aside from cortical layer cytoarchitectonics, is currently largely unexplored.
Utilizing array tomography to its fullest extent requires the development of new synapse detection and classification capabilities. Simple analysis, using repeated human observation of a fraction of the channels available in the full volume, may be acceptable for analyzing fragmentary subsets of a few hundred synapses but cannot scale beyond that. We have developed tools and methods to assist in handling the high proteomic dimensionality of array tomographic volumes (Figure 1), principally the synaptogram [17]; a means of visualizing small pieces of highly multiplexed data by splaying out the 3-D volume surrounding a region of interest (ostensibly a single synapse) into a larger 2-D image.
An example of a synaptogram in action can be seen in Figure 1-C,D, both of which visualize the same synaptic volume. 1-C attempts to render the volume in three dimensions, assigning a different color to each channel, and running out of easily separable colors in the process, even for this one example. It also falls prone to the usual pitfalls of obscuration and optical confusion common to snapshots of rendered scenes, such that splitting the image into multiple ones displaying subsets of colors helps visualization considerably. Contrast this with a synaptogram of the same synapse in Figure 1-D. Each row of thumbnails displays a different channel (plus synapsin, included to serve as a reference channel), each column shows a different z-section; left is below, right is above. Unlike the render, position and depth relationships are presented clearly, and the synaptogram can be extended to include an arbitrarily large number of simultaneous imaging channels by appending new rows vertically.
With only a bit of exposure to synaptograms, human experts can use them to tell at a glance exactly what they're seeing. This eases the difficulty of per-synapse manual classification such that the effort of classifying a set of few hundred synapses is no longer excessive, but no matter how convenient they are to analyze individually, the sheer number of synapses makes manual analysis of the entire data set effectively impractical.
Given that just a few hundred analyzed examples can be obtained with a reasonable expenditure of effort, there are two approaches to consider. The first is to use those examples as a representative sample, in a manner similar to stereology. That may work well for some questions, but not others. Rare or novel synapse types and cortical laminar distributions would be especially difficult to study. An alternative, which this paper will present, is to take that sample of accurately classified synapses and extrapolate its decision-making information to the much larger population of unclassified individuals.
The first necessary step in our classification process is to locate the sites which may contain synapses. Despite their appreciable proteomic diversity [18], cortical synapses are small: from the ostensible midpoint of the synapse, all relevant synaptic protein labeling can fit within a 500 nanometer radius for mouse cortex [19]. Given a reliable method of locating synapses, all information needed to verify and type those synapses can be measured from the local volume surrounding them, greatly reducing the spatial analysis needed per synapse. To avoid confusion with actual synapses, we refer to these sorts of putative synapse locations as “synaptic loci.” They are specific places which might be synaptic.
In order to find putative synapses to help limit the necessary search space, we are using an antibody targeting Synapsin I. Synapsin is a scaffolding protein reportedly found in all cortical synapses [20], and labeled antibodies targeting synapsin have previously been used on their own to estimate synapse counts [21]. A Millipore Rabbit anti-Synapsin I antibody (Millipore AB1543P) demonstrates robust and reliable labeling, and is likely to be colocalized with all relevant synaptic markers [17]. For these reasons the core of our analysis uses Synapsin I labeling to derive a list of locations likely to contain synapses from which to begin small volumetric searches for confirmation. Our approach is to use the brightest point of each Synapsin I punctum as the site of a possible synapse to designate a local volume for further analysis, without attempting to explicitly determine the punctum boundaries.
We prefer our local maxima-based approach over thresholding-based segmentation because the latter has a number of issues arising from AT's largely anisotropic resolution (∼200nm × ∼200nm × 70nm). This anisotropy, combined with (often unknown) epitope density and labeling variance means that any segmented punctum boundary is at best an estimate. An approach using local maxima, paired with a voxel-based rotation-invariant feature set, is not affected by the exact boundaries of the puncta of interest, but by the puncta themselves.
While our approach to synapse discovery sidesteps segmentation, it does so at the cost of introducing potential false positives: background local maxima which segmentation would have discarded, but whose peak brightness rises over our low threshold for consideration. However, it is possible to filter those out in later classification. Conversely, this method is ideal for teasing apart “clumps” of synaptic labeling, where multiple synapses exist in close proximity but can be resolved by the Rayleigh criterion and thus having separate local maxima.
Machine learning methods come in two broad categories. Supervised learning algorithms, trained using a sufficient number of human rated synapses, are capable of producing numerical descriptions of human judgment as it is applied to synapse classification, as well as extrapolating that judgment to the hundreds of thousands of synapses which comprise an average data set. Unsupervised clustering, on the other hand, when applied to raw synaptic loci or already classified synapses is a great approach to the discovery of marginal classes or subtle subtypes.
Although visual analysis is the traditional and preferred method of examining biological data, long strings of numbers such as our feature vectors are difficult for humans to visualize. In response, high-dimensional numerical measurements have often been approached using some form of dimensionality reduction as a first step in numerical analysis. Simply put, reducing a long string of numbers to a short string of numbers makes them easier graphically display and understand. Principal Component Analysis (PCA) is a venerable method of dimensionality reduction which has seen use in similar applications [22], [23], and has proven useful in ours as well.
Our PCA result, illustrated in Figure 2, identifies some synaptic populations but does not separate them sufficiently for classification. The loci tend to aggregate in clusters which correspond to a few of the broader synaptic categorizations, namely GABAergic and two common subtypes of glutamatergic synapses. We identified the clusters using multivariate regression, that is, taking a few of the more distant examples and inferring the contribution of channels which brought them from the mean. The PCA demonstrates the existence of separate populations corresponding to each class, but in the reduced dimensionality of PCA, simple thresholds are insufficient for proper class discrimination. The two components plotted explain 50.4% of the variance between them. An additional 20.9% of the variance is present in the second principal component (not shown), which appeared to represent synapse size.
Ideally, the dimensionality reduction accomplished by the above methods would have proven amenable to simple thresholding. If that where the case, multivariate regression might have led to identification and, combined with a measure of the statistical significance of cluster separation, classification of unknown synapses based solely on where they fell in the unsupervised plot. Since our clusters were not so cleanly separable, we resorted to a more subtle stratagem involving supervised learning.
The “supervision” of supervised learning refers to the supervised training set, a random or semi-random collection of human-rated examples from which the machine learning algorithm (MLA) infers the rules for classification to extrapolate onto novel synapses. To generate each item of the set, we presented a synaptogram to a human trainer, who rated the synaptogram in one or more binary categories representing the presence or absence of channels relevant to synapse classes of interest. We could then associate those categorizations with the already-derived feature vectors of those examples, compiling them into a library of “correct” classifications for training.
After classification, the predicted presence of each channel for a given locus can be derived from the percentage of decision trees in the random forest ensemble which attest to its presence. This effectively serves as a confidence metric for the entire ensemble, and is generally referred to as the “posterior probability.” An instance with a posterior probability of 1.0 is unequivocally positive for the class in question, one of 0.0 is undeniably negative. In this manner, we reduce the 4-long numeric feature vector to a -long numeric posterior vector, representing the presence or absence of all relevant channels. We can then use these vectors in a combinatorial fashion to recreate synaptic classes. Glutamatergic VGluT1-expressing synapses, for example, should at a minimum be positive (posterior probability 0.5) for VGluT1 and PSD95.
When we began the class discovery process as shown in Figure 6, we expected relationships based on our preconceived notions of a few synapse classes: that GAD, VGAT and parvalbumin should all be copresent to some extent [28], and that VGluT1 and VGluT2 should each colocalize with PSD95, but not with each other [31]. Since the algorithm was partially developed with these relationships in mind, it is unsurprising that we found them.
The other channels we examined, VGluT3, VAChT and TH, had fewer performance expectations, largely due to their paucity in the data set. Of the three, VGluT3 displayed the most interestingly unexpected behavior, avoiding the common glutamatergic markers and colocalizing instead with all GABAergic synaptic markers. This is most likely a rare but suspected role of VGluT3 in the cortex [30], co-expressed with GAD in a minor population of excitatory interneurons. If so, a followup experiment with larger volumes may be able to more effectively study this sparse population, but its detection given even a small number of examples lends us a degree of confidence that our analysis returns usable results sufficient to detect novel synaptic phenomena.
Another interesting result with even fewer candidate synapses is a cortex-only localization of VGluT2 and TH. Dopaminergic (TH) neurons have been reported to express VGluT2 in rat cultures [32], midbrain and hypothalamus [33]. It is possible that these are afferent projections from subthalamic nuclei, in which case their localization within the cortex and further proteomic differentiation would be interesting to examine in more detail. However, with current volume sizes we only find a dozen of these appositions, so at this time it would be problematic to assert certain confirmation of their existence, much less their distribution.
The variance of human raters raises a few interesting questions to look into in the future. Two of the six raters (#2 and #6) self-reported using a stricter standard of classification than the rest: when an example was at all doubtful, they classified it as being negative. Effectively, these raters elected to position themselves on the left side of the ROC curve, trading an increase in false negatives for reduced false positives. Depending on the application, stricter classification may be preferable. We tested an example of this sort of premeditated error ourselves by training a number of MLAs with various classification criteria, and comparing their output. These results are presented in Table 3. As one might expect, we found that a bit of bias in the training process could go a long way to reducing either Type I or II errors, at the cost of increasing the other, and that this effect is exaggerated when processing examples human raters find difficult.
Based on our experiences, we would recommend taking time to discuss questionable examples and reasons for rating them one way or another. Such conversations are rather illuminating and very effective at getting everyone to agree on a common standard of classification.
There are two significant limitations to the questions which can be asked using this method. The first and strictest: an array tomography volume is a decidedly terminal snapshot of a piece of tissue. This precludes experiments which would watch a particular cell or dendrite change over time, or in response to learning [34], except in animal models which are stereotyped enough for different animals to have equivalent nervous systems, namely C. Elegans [35] and Drosophila [36]. Synapse populations are assumed to be fairly invariant between individual mice (and presumably humans), however, which allows us to study changes to synaptic classes as a whole in response to plasticity or disease.
The second limitation is more easily rectified. Our analysis partially depends on limiting the scope of the problem to that required to identify synapses at locations already suspected to contain a synapse. For common synapse classes this is easy. They all express Synapsin I, so wherever we find our Synapsin I marker, there may be a synapse. As mentioned, we have already begun to abut the usefulness of Synapsin I, which may not be expressed in dopaminergic synapses [25]. Using a pan-Synapsin antibody would be a straightforward solution to catching all dopaminergic synapses, but it is fully possible that other, more exotic synapse types may not express Synapsin at all, instead relying on some currently unknown mechanism to perform the same function.
Establishing a robust system for synapse classification in array tomographic volumes opens up a number of avenues for addressing biological questions. It allows us to conduct single-synapse analyses in large regions of tissue, which lets us study rare or spatially-segregated populations. It helps us discover new synaptic populations and novel variations on known synapse types, and gives us an unprecedented level of control over the proteomic complexity we can bring to bear.
All procedures related to the care and treatment of animals were approved by the Administrative Panel on Laboratory Animal Care at Stanford University. All volumes were acquired from mouse cortex, line C57BL/6J, using the methodology given in [17].
One adult mouse was used for this study. The animal was anesthetized by halothane inhalation and its brain quickly removed and placed in 4% formaldehyde and 2.5% sucrose in phosphate-buffered saline (PBS) at room temperature. Its cerebral hemisphere was sliced coronally into three pieces and fixed and embedded using rapid microwave irradiation (PELCO 3451 laboratory microwave system with ColdSpot; Ted Pella, Redding CA) as described in [37]. The tissue was dehydrated up to 70% ethanol.
Ribbons of serial ultrathin (70 nm) sections were cut with an ultramicrotome (EM UC6, Leica Microsystems, Wetzlar, Germany) as described in [37]. The ribbons were mounted on subbed coverslips (coated with 0.5% gelatin and 0.05% chromium potassium sulfate) and placed on a hot plate (60 C) for 30 min. For SEM imaging, the subbed coverslips were also carbon coated using a Denton Bench Top Turbo Carbon Evaporator (Denton Vacuum, Moorestown, NJ). Subbed and carbon coated coverslips were also prepared for mounting ribbons of sections to be used for multiple immunostaining rounds (6).
Staining was performed as described in [37]. The coverslips with sections were mounted using SlowFade Gold antifade with DAPI (Invitrogen, Carlsbad CA). To elute the applied antibodies, the mounting medium was washed away with dH2O and a solution of 0.2 M NaOH and 0.02% SDS in distilled water was applied for 20 min. After an extensive wash with Tris buffer and distilled water, the coverslips were dried and placed on a hot plate (60C) for 30 min.
The primary antibodies and their dilutions are listed in [17], Table 1. Only well characterized commercial antibodies were used and they were evaluated specifically for AT as described in the Supplemental Experimental Procedures of [17]. For immunofluorescence, Alexa Fluor 488, 594, and 647 secondary antibodies of the appropriate species, highly preadsorbed (Invitrogen, Carlsbad CA) were used at a dilution 1∶150. The sequence of antibody application in the multiround staining is presented in [17], Table S1.
Sections were imaged on a Zeiss Axio Imager.Z1 Upright Fluorescence Microscope with motorized stage and Axiocam HR Digital Camera as described in [37]. Briefly, a tiled image of the entire ribbon of sections on a coverslip was obtained using a 10 objective and the MosaiX feature of the software. The region of interest was then identified on each section with custom-made software and imaged at a higher magnification with a Zeiss 63/1.4 NA Plan Apochromat objective, using the image-based automatic focus capability of the software. The resulting stack of images was exported to ImageJ, aligned using the MultiStackReg plugin and imported back into the Axiovision software to generate a volume rendering. When a ribbon was stained and imaged multiple times, the MultiStackReg plugin was used to register the stacks generated from each successive imaging session with the first session stacks based on the DAPI channel, then a second within-stack alignment was applied to all the stacks. Since DAPI was stained in all imaging sessions it made an ideal candidate for alignment, and the alignment transformation of each imaging session's DAPI channel was propagated to the other members of that session to bring the entire channel set into the same coordinate space.
To reconstruct the larger volume of tissue used in this study, we first used Zeiss Axiovision software to stitch together individual high-magnification image tiles and produce a single mosaic image of each antibody stain for each serial section in the ribbon, creating a z stack of mosaic images for each fluorescence channel instead of a single field of view stack. To coarsely align the image stacks, we used the MultiStackReg plugin with the DAPI channel, as described above and in [37].
To analyze synapse-level structures an additional alignment step was needed to remove a minor non-linear physical warping introduced into the ribbons by the sectioning process. We used a second ImageJ plugin, autobUnwarpJ (available at http://www.stanford.edu/~nweiler), which adapts an algorithm for elastic image registration using vector-spline regularization [38]. As before, we aligned only a single channel, Synapsin, and propagated the generated transformation to the other channels. Synapsin proved ideal for this purpose because it is a dense, high-frequency channel whose labeled objects are still considerably thicker than a single section, creating good fiducial markers for the alignment process.
Finally, data used for Table 3 and Figure S2 were processed after imaging using a method of deconvolution recently published by our lab [39]. This does not seem to affect MLA performance, but the smaller, more discrete puncta do cause an increase in the number of synapsin local maxima, and therefore generates more extracted synapsin loci. Future work using deconvolved volumes may benefit from incorporating an additional filtering step in the extraction process to either smooth the data before finding local maxima or segment puncta more directly.
Before analyzing imaged volumes, we subtracted the background from each fluorescent channel using a 10×10 pixel (1 ) rolling ball filter to remove systematic non-punctate background fluorescence, then normalized each slice of the stack without saturating any pixels, such that the brightness histogram of each section was stretched as much as possible without loss of information. No other image processing, including removal of fluorescence due to foreign material, nonspecific staining, etc, was performed before analysis.
To extract a list of putative synapse locations from raw volume data, we first identified individual synapsin puncta by convolving the synapsin channel with a 3×3×3 local maxima filter; retaining all voxels with a brightness those of its 26-voxel neighborhood. Then, we passed the synapsin maxima through a connected component filter to reduce peak voxel clumps (caused by discretization of the fluorescence data) to centroids, and discarded those below a deliberately low threshold (10% of the total brightness range) as being too dim to represent a real synapse. What remained was a list of putative synapse locations, or “synaptic loci,” so named for their central role in later classification steps.
In order to gauge the reliability of any single human expert's rating, we performed a qualitative test of the consistency of human classification. We presented a set of one hundred randomly-selected synaptograms to a group of six human raters who were familiar with the task of interpreting synaptograms, and instructed them to classify the set based on whether or not the synaptogram was centered on a glutamatergic synapse. Once collated, we considered the true classification of a given synaptogram to be that of a simple majority vote of the first five raters (to prevent ties). When we compared each rater's performance relative to the average, we found an average accuracy rate of 77.7%, with a standard deviation of 10.1% (Figure 4). The largest source of variance arose from the self-reported stringency of the raters, in how much ambiguity they found acceptable when classifying a locus as positive.
To test the influence that training stringency and classification difficulty have on MLA performance, we repeated the above test with three classifiers trained by rater 1. In addition to the classifier trained with the default strategy, classifiers that would attempt to guess “yes” or “no” in ambiguous cases were trained and their mutual performance compared, using their average agreement to establish a gold standard as was the case for the humans. We also subdivided the data set further into “easy” versus “hard” cases through the use of a fourth MLA, and compared those conditions as well.
The color of each point in the PCA figure was determined by taking the extreme outliers of the three clusters, determining their feature composition via multivariate regression, taking the dot product of the feature weight vectors with the feature vector of each locus, and assigning that to red, green or blue for the VGluT1, VGluT2, and GABA clusters respectively. Colors were manually normalized to be of approximately equal intensity, and synaptic loci not strongly represented in any of the three colors were removed to better visualize cluster relationship.
The training of our machine learning algorithm differs from standard supervised learning, in which training examples are chosen at random, by instead selecting examples which together compose a varied training set. We also added a preprocessing step to facilitate the training of very rare classes, on our case VAchT and VGluT3. Thus the training set generation occurs in two phases. The first phase is to “prime” the training set data for rare classes by choosing one of each class's requisite presynaptic channels and randomly sampling a subset from the loci for which the channel's local brightness is more than two standard deviations above the mean. A number of class subsets generated in this manner are collated, each class contributing to the negative examples of the rest. The second phase is an “active” training process in which a human rater and the MLA being trained work in tandem to speed training, a technique known as active learning [40], [41]. At each step, the half-trained classifier selects a few examples, half of which it thinks are positive and half negative, to present to the rater for verification and feedback.
In pseudocode, the training proceeds according to the following algorithm:
while Human wishes to train do
Load training synaptogram population,
Human selects a synaptic category
Train RFE using partially classified training set , display predicted error rate
while Human wishes to add training examples do
Randomly choose , where
Randomly choose a synaptogram from subpopulation , the elements of classified as
Display and to human for verification
Add/Update in to reflect human input
end while
end while
To produce the pairwise channel copresence map, for each marker pair we calculated the probability of co-occurrence , where is the number of loci found to be positive for , and is the number of total loci in the population. Multiplying by gives us the expected population, . We compared this number with the observed population using difference over sum normalization to find the normalized pairwise relationship . These relationships made pairwise comparisons easy to interpret, with one minor counter-intuitive exception: markers which comprised a substantial proportion of the synaptic loci population (VGluT1 and PSD95) had reduced values, even with themselves, owing to their high . To bring those into the same reference frame as the rest, we normalized again using the reciprocal of the sum of the relationship identity reciprocals, that is, . Finally, since the previous step disrupted negative relationship scaling such that the most negative pairs (VGluT1 vs GABAergic markers) reached nearly −3.0, we multiplied the positive ratings by 3 to match once more.
To simplify the calculation of the cortical depth-dependent metrics used in Figure 5, such that any given Y-value represented tissue at the same cortical depth, we needed to correct a minor slant in the raw volume. We measured the degree of tissue slope using the pial surface and the white matter/striatum boundary, and imposed an affine transformation on the loci, linearly interpolating them to be level. The underlying data and the features used to classify the loci were not changed as a result of this process.
Image normalization, locus discovery and feature extraction were implemented and performed using Fiji (http://fiji.sc/). Training set generation was implemented as a browser-based application, coded in Python, to permit our experts to work at their leisure. We used R for interactive classification for its ease of Python integration, but the final random forest classifiers, trained on the complete training set alone, used MATLAB (the TreeBagger class). Imaris was used to render the data for visualization of Figure 1.
All implemented code used in this analysis is available at http://code.google.com/p/smithlabsoftware/ under a GPL v3 license.
|
10.1371/journal.pntd.0005615 | Lp25 membrane protein from pathogenic Leptospira spp. is associated with rhabdomyolysis and oliguric acute kidney injury in a guinea pig model of leptospirosis | Acute kidney injury (AKI) from leptospirosis is frequently nonoliguric with hypo- or normokalemia. Higher serum potassium levels are observed in non-survivor patients and may have been caused by more severe AKI, metabolic disarrangement, or rhabdomyolysis. An association between the creatine phosphokinase (CPK) level and maximum serum creatinine level has been observed in these patients, which suggests that rhabdomyolysis contributes to severe AKI and hyperkalemia. LipL32 and Lp25 are conserved proteins in pathogenic strains of Leptospira spp., but these proteins have no known function. This study evaluated the effect of these proteins on renal function in guinea pigs. Lp25 is an outer membrane protein that appears responsible for the development of oliguric AKI associated with hyperkalemia induced by rhabdomyolysis (e.g., elevated CPK, uric acid and serum phosphate). This study is the first characterization of a leptospiral outer membrane protein that is associated with severe manifestations of leptospirosis. Therapeutic methods to attenuate this protein and inhibit rhabdomyolysis-induced AKI could protect animals and patients from severe forms of this disease and decrease mortality.
| Rhabdomyolysis is a syndrome that results from the disruption of skeletal muscle integrity, leading to a massive release of the intracellular contents into the blood stream, including myoglobin, creatine phosphokinase, aspartate transaminase, lactate dehydrogenase, aldolase, and electrolytes. Complications of rhabdomyolysis include acute kidney injury (AKI), hyperkalemia, hyperphosphatemia, and hypovolemia, which may result in death without early treatment. The most frequent causes of this syndrome are trauma, excessive muscle activity, drugs, toxins, electrolyte imbalance, muscle ischemia, metabolic disorders, and infectious diseases. Among leptospirosis cases, the AKI induced by rhabdomyolysis has been described almost exclusively in patients with severe form of leptospirosis. However, the role of rhabdomyolysis in the pathogenesis of AKI due to leptospiral infection is not understood.
| Leptospirosis is an emerging zoonosis that is caused by pathogenic spirochetes of the genus Leptospira. Approximately 1.03 million cases of the disease occur in humans worldwide, with approximately 60,000 deaths annually [1]. Many species of wild and domestic animals are leptospirosis reservoir hosts and eliminate leptospires to the environment via urinary shedding. Infection may result from direct contact with carrier animals or indirect contact with contaminated soil and water [2,3].
Human leptospirosis ranges from an asymptomatic or self-limited febrile illness (80 to 90% of cases) to a life-threatening illness (5 to 10% of cases). The life-threatening manifestation is characterized by Weil´s syndrome (renal failure, hemorrhage and jaundice) or severe pulmonary hemorrhagic syndrome [2–5].
Leptospirosis-induced acute kidney injury (AKI) is typically nonoliguric at the beginning of renal failure evolution or during mild forms with high frequency of hypokalemia [4]. In a prior study, higher serum potassium levels were observed in patients with more severe renal dysfunction concomitant with rhabdomyolysis. In addition, an association between creatine phosphokinase levels (CPK) (a marker of muscle injury) and maximum serum creatinine levels has been reported. This suggests that rhabdomyolysis is associated with severe AKI in leptospirosis [6].
Various bacterial, viral, fungal and protozoal infections lead to rhabdomyolysis [7–10], but the mechanism of muscle damage has not been established for many infections, including leptospirosis. Muscle injury may result from a direct pathogen invasion of skeletal muscle, tissue hypoxia, high lysosomal enzymatic activity or the release of toxins [11,12]. The identification of proteins that act as toxins in the host during leptospiral infection is essential to understanding the pathophysiological mechanisms of rhabdomyolysis and the mechanisms that contribute to severity of AKI.
The subsurface lipoprotein LipL32 is present in pathogenic leptospires, and it is the most abundantly expressed protein (40,000 copies per cell) [13,14]. However, the role of this protein in the pathogenesis of leptospirosis remains unknown [15]. Lp25 is a putative outer membrane lipoprotein of pathogenic Leptospira spp., but its function is not known. No sequences similar to this protein were identified in saprophytic Leptospira spp. [16–18].
The present study investigated whether the LipL32 and Lp25 proteins expressed by pathogenic Leptospira were associated with rhabdomyolysis and oliguric AKI in guinea pigs. To our knowledge, this study is the first characterization of a leptospiral protein associated with renal and muscular manifestations of leptospirosis.
Leptospira biflexa serovar Patoc strain Patoc I, Leptospira noguchii serovar Panama strain CZ214K, Leptospira borgpetersenii serovar Javanica strain Veldrat Batavia 46, Leptospira borgpetersenii serovar Tarassovi strain 17, Leptospira kirschneri serovar Cynopteri strain 3522C, Leptospira interrogans serovar Hardjo strain Hardjoprajitno, Leptospira interrogans serovar Pomona strain 13A, and Leptospira interrogans serovar Copenhageni strain Fiocruz L1-130 were obtained from the Laboratory of Bacterial Zoonosis, School of Veterinary Medicine and Animal Science, University of São Paulo, Brazil. Leptospira strains were cultured at 29°C under aerobic conditions in liquid EMJH medium (Difco, Thermo Fisher Scientific, Boston, MA, USA.) with 10% rabbit serum, enriched with L-asparagine (0.015%), sodium pyruvate (0.001% [wt/vol]), calcium chloride (0.001% [wt/vol]), magnesium chloride (0.001% [wt/vol]), peptone (0.03% [wt/vol]), and meat extract (0.02% [wt/vol]) [17].
LipL32 and Lp25 proteins were chosen for this study because no research has been performed to investigate their effects on renal function experimentally in animals. Lp25 was identified by bioinformatics analyses using the L. interrogans serovar Copenhageni strain Fiocruz L1-130 genome sequence previously described in studies published by our group [17]. The selection was based on the prediction of protein localization in the outer membrane. We gave priority to Lp25 because its function is not known. Leptospiral immunoglobulin-like protein A (LigA) [19] and LpL31 [20] were used as controls in the immunoblot analysis. LigA is a known outer membrane protein, and LipL31 is an inner membrane-associated protein [19, 20].
Open reading frames LIC10009 (encoding a protein designated Lp25, for leptospiral protein 25, based on its molecular mass) [21] and LIC11352 (LipL32) were cloned into pAE [17,22] and pDEST-17 (Invitrogen, Carlsbad, CA, USA -or- Paisley, Scotland, UK.) vectors, respectively, as previously described [23]. The coding sequence of the carboxy-terminal portion of LigA (LigAC), corresponding to nucleotides 1891–3675 (LIC10465), was cloned into a pAE vector as previously described [21]. The coding sequence of the LipL31 (LIC11456) was amplified using PCR from genomic DNA of L. interrogans serovar Copenhageni strain Fiocruz L1-130 using the following primers: F: CTCGAGGGAGATAATTCCG and R: CTGCAGTTACTGCCCAGTAG. Sequences were digested using XhoI and HindIII restriction enzymes, and fragments were cloned into the pAE vector [22]. Competent cells of the E.coli BL21(DE3) strain were transformed with pAE-Lp25, pAE-LigAC, pDEST-LipL32, and pAE-LipL31 constructs and cultivated until the optical density at 600 nm reached 0.6. The expression of recombinant proteins was induced with 1 mM isopropyl-1-thio-β-D-galactopyranoside (IPTG) at 37°C for 3 h. The His-tagged Lp25, LipL32, LigAC, and LipL31 proteins were purified using metal affinity chromatography, as previously described [21].
A New Zealand White rabbit was immunized for each protein via subcutaneous injection of 2 mg of purified recombinant protein absorbed in aluminum hydroxide as an adjuvant. The rabbits were immunized more two times with the same antigen preparation with fifteen day of interval. The rabbits were bled two weeks after last immunization. The IgG fraction from sera was precipitated using caprylic acid, as previously described [24].
OMPs from L. interrogans serovar Copenhageni strain Fiocruz L1-130 were extracted with Triton X-114 according to a previously described method [25]. Three distinct fractions were recovered: (P) the detergent-insoluble pellet that corresponds to inner membranes, cytoplasmic components, and non-lysed cells; (A) the aqueous phase that contains periplasmic content, and, (D) the Triton X-114 detergent phase with outer membranes. The whole cell lysate (W) and the obtained fractions were analyzed using immunoblots with LipL32, Lp25, LigA, and LipL31 antisera.
L. interrogans serovar Copenhageni strain Fiocruz L1-130 (5×108 cells/ml) were harvested via centrifugation at 2000 × g for 7 min, gently washed with PBS containing 5 mM MgCl2, and collected via centrifugation at 2000 × g for 10 min. Washed spirochetes were gently resuspended in PBS-5mM MgCl2, and the evaluation of surface protein localization on intact leptospires was performed by treatment of proteinase K (PK—Sigma-Aldrich, St. Louis, MO, USA) as previously described [13]. Immunoblot analyses were performed using antibodies against LipL32 and Lp25.
Proteins were transferred to a nitrocellulose membrane and probed with LipL32, Lp25, LigA, or LipL31 rabbit polyclonal antisera (diluted 1:500). The membrane was incubated with a secondary peroxidase-conjugated anti-rabbit antibody at a 1:5,000 dilution. The positive signals were detected using enhanced chemiluminescence (Thermo Fisher Scientific, Boston, MA, USA).
The guinea pigs were housed one per metabolic cages with food and drinking water freely available. The animals were acclimated to the housing conditions for 1 day prior to experimental procedures. Three groups of animals were studied: (1) control, n = 10 (1 ml of PBS); (2) LipL32, n = 13 (1 ml of PBS plus 400 μg of LipL32) and (3) Lp25, n = 14 (1 ml of PBS plus 400 μg of Lp25). The amount of protein used per dose (400 μg) was the greatest amount maintained in solution without precipitation. The solutions were injected intraperitoneally for 4 days. The animals were placed in metabolic cages for 12 h for urine collection on day 5. The volume of each 12-h urine sample was measured gravimetrically (UV ml/12 h). Urine samples were centrifuged in aliquots to remove suspended material, and urinary creatinine and sodium in the supernatants were measured. After urine collection all animals were anesthetized with a dose of sodium pentobarbital, and whole blood was collected by cardiac puncture. The animals were then killed with an overdose of anesthesia. Serum potassium and sodium were measured using flame photometry. The enzymatic colorimetric method (Labtest, Lagoa Santa, Brazil) was used to measure urinary and serum levels of creatinine, CPK and uric acid. The molybdate method was used to measure serum phosphate. The creatinine clearance (CrCl) was used to estimate the glomerular filtration rate (GFR) by formula: CrCl = Ucr (mg/dL) x UV (ml/min) / Pcr (mg/dL), corrected by 100 g of body weight (ml/min/100 g body weight) [26]. Fractional excretion of sodium was calculated using the formula: FENa = UNa x PCr / PNa x UCr x 100%. AKI was defined as a decrease in the GFR of more than 50% from the mean value obtained in the control group (PBS), and oliguria was defined as a urinary output of less than 50% of mean value of control group (PBS). Rhabdomyolysis was defined as an elevation in serum CPK of at least 3 times the mean value obtained in the control group. The occurrence of hyperkalemia, hyperphosphatemia, and hyperuricemia was defined as an increase of phosphate, potassium, and acid uric levels of more than the mean value of control group (PBS).
Muscle fragments from legs and paravertebral regions were collected at the time of euthanasia, routinely fixed in 10% buffered formalin (pH 7.2), embedded in paraffin and sectioned at 3 μm. Fragments from kidney were also collected and submitted to routine histological procedures. Sections were analyzed using EnVision (Dako, Glostrup, Denmark)-based immunohistochemistry methods, as previously described (5). The antigen retrieval step was performed by pressure cooking in 10 mM sodium citrate pH 6. Following the overnight incubation with primary rabbit polyclonal antisera (diluted 1:3,000–1:4,000) at 4°C and with secondary antibody (Envision peroxidase Dako K4002) for 30 min at room temperature. The presence of nonspecific staining was assessed using preimmune sera. Tissue sections for morphological analyses were stained with hematoxylin and eosin (H&E) and Gomori trichrome stain in selected sections. Muscular lesions were graded on a scale from 0 to 2: 0 as normal (without lesions), 1 as mild (chiefly the presence of individual hyaline contraction change and focal inflammatory interstitial reactivity), and 2 as severe (presence of necrosis, multiple lesions of individual myocytes and interstitial inflammatory infiltrated). Kidney sections were also fixed in 10% buffered formalin and stained with H&E for morphological analyses. Images were captured on an Axiophot Zeiss Axio microscope and analyzed using AxionVision 4.6 software.
The Ethic Committee on Animal Use of the Butantan Institute (CEUAIB), São Paulo, Brazil, previously approved the experimental protocols under the license numbers 55708 for the rabbit procedure and 99112 for guinea pig procedure. All animal procedures were conducted following the rules issued by the Brazilian National Council for Control of Animal Experimentation (CONCEA).
All quantitative data are expressed as the means ± SEM. Differences between the means of multiple parameters were analyzed using ANOVA followed by Student-Newman-Keuls test. Histological scores were compared using Student’s t-test. Values of p < 0.05 were considered statistically significant. All analyses were performed using GraphPad Prism 5 (Graphpad, La Jolla, CA).
A total of 37 guinea pigs were assigned into one of three treatment groups: control (n = 10), LipL32 (n = 13) or Lp25 (n = 14). Initial body weights were similar between the 3 groups: 194±4 g (control), 195±4 g (LipL32) and 191±3 g (Lp25). Weight gain was lower in the LipL32 group (18±3 g, p<0,05) and Lp25 group (13±3 g, p<0,01) than the control group (28±2 g) on day 5. Fig 1 shows that the GFR, evaluated as creatinine clearance (CrCl), was significantly lower in the Lp25 group (0.47±0.03 mL/min/100 gBW) than the control group (1.05±0.13 mL/min/100 gBW) and LipL32 (0.87±0.10 mL/min/100 gBW). The urinary volume was lower in the Lp25 group (12.0±1.3 UV mL/12h) than the control group (23.0±3.8 UV mL/12h) and the LipL32 group (17.3±3.6 UV mL/12h). Notably, the serum potassium level in the Lp25 group (6.7±0.5 mEq/L) was elevated, compared to the control group (4.7±0.2 mEq/L) and the LipL32 group (5.7±0.3 mEq/L). The fractional excretion of sodium was similar in the three groups (control 0.82±0.18%; LipL32 0.60±0.09%; Lp25 0.83± 0.15%).
The Lp25 group had significantly higher levels of serum CPK, phosphate and uric acid (2060±338 CPK U/L; 8.36±0.32 P mg/dL and 2.75±0.56 acid uric mg/dL) (Fig 2). These parameters in the LipL32 group (726±216 CPK U/L; 6.70±0.41 P mg/dL and 1.06±0.28 acid uric mg/dL) were similar to the control group (763±197 CPK U/L; 7.06±0.30 P mg/dL and 1.07±0.20 acid uric mg/dL).
AKI was observed in all animals of Lp25 group (100%) and 7/14 (50%) of these also had oliguria. In the LipL32 group, 2/13 animals (15.38%) had AKI oliguric and 1/13 (7.69%) presented AKI nonoliguric. Hyperkalemia was seen in 13/14 (92.85%) guinea pigs in the Lp25 group vs 8/13 (61.53%) in the LpL32 group.
In the Lp25 group, rhabdomyolysis, hyperphosphatemia, and hyperuricemia were encountered in 12/14 (85.71%), 10/14 (71.42%) and 9/14 (64.28%) animals, respectively. Seven animals of this group (50%) showed the three outcomes concomitantly. In the LipL32 group, 2/13 (15.38%) had rhabdomyolysis, 3/13 (23.07%) had hyperphosphatemia and 2/13 (15.38%) had hyperuricemia.
We performed immunohistological and morphological analyses to assess the effect of the LipL32 and Lp25 proteins on muscle tissues. The anti-rabbit Lp25 antibodies labeled isolated or small groups of muscular fibers with granular, faintly brownish antigen deposits that may partially delineate the sarcolemma and spread to the cytoplasm below (Fig 3A). The inflammatory infiltrate was generally discrete and primarily composed of small groups of monocytes that were present as focal isolated interstitial groups or small groups of isolated muscular fibers, frequently with antigenic linear deposits that partially circumscribed muscular fibers or inside their cytoplasm. Cytoplasmic antigenic granules were also detected occasionally in mononuclear phagocytic cells when the inflammatory infiltrate is more conspicuous. Immunohistochemistry with the anti-rabbit LipL32 antibodies was also positive on the sarcolemma and in the cytoplasm of isolated muscular fibers. Scarce mononuclear inflammatory interstitial reactions near the damaged muscle were also present (Fig 3B). Histological analyses of the muscle fragments revealed essentially similar muscular lesions were present in animals inoculated with both proteins, but with different degrees of severity (Fig 3C–3H). No wide range of muscular fiber sizes (small or large groups of atrophic or hypertrophic fibers) were observed in either protein group. Internal sarcolemmal nuclei were not detected in isolated muscle fibers. However, isolated nonspecific muscular lesions were present and ranged from hyaline contraction cytoplasmatic changes (Fig 3E and 3F) to small vacuoles, which may progress to hypercontracted isolated fibers to necrosis prior to phagocytosis (Fig 3G) or to intermediate damage, which was characterized by staining changes in myofibrils (Fig 3F), including the appearance of pale necrotic cells in H&E and Gomori trichrome stains (“ghost cells “) (Fig 3F and 3H). Irregular areas of muscle necrosis were the end result of these muscular disturbances (Fig 3C and 3D). Mild inflammatory infiltrate composed of macrophages were also observed around necrotic areas. Fragments from the LipL32 group revealed essentially similar findings to the Lp25 group, but with less frequent focal muscular damage and areas of necrosis (Fig 3D).
Muscular lesions severity scores were significantly lower (p<0.05) in the LipL32 group (0.66 ± 0.21) than the Lp25 group (1.43 ± 0.20), and no muscular lesions were observed in the control group. The difference between LipL32 and Lp25-inoculated animals was statistically significant (p<0.05, LipL32 vs. Lp25) (Fig 4). Specifically, 53.85% of animals treated with LipL32 shown no muscular lesions, whereas the lesions of severity stage 1 and 2 were observed in 30.77% and 15.38%, respectively, of guinea pigs inoculated with this protein. Otherwise, all Lp25-inoculated animals presented muscular lesions, of these 42.86% had mild manifestations (stage 1) and 57.14% had severe signs (stage 2). Histological examination of the kidneys revealed no lesions (S1 Fig).
Immunoblots of whole cell lysates revealed that the Lp25 protein was expressed by all strains of pathogenic leptospires tested, and it was not detected in the non-pathogenic strain (Fig 5A). Equal results were obtained in immunoblot assays using LipL32 antiserum as a positive control and demonstrated that the Lp25 protein, like the LipL32 [25], was conserved and only found in pathogenic species of Leptospira.
All control proteins were detected in whole cell extracts. LigA was completely solubilized by the detergent and fractionated into the aqueous and detergent phases. LipL31 was detected in the insoluble pellet fraction and aqueous phase, and it was completely absent from the detergent phase (Fig 5B). We performed an additional experiment using the LipL32 antiserum because a previous work reported that LipL32 was solubilized by Triton X-114 and mostly detected in the detergent fraction [25]. Fig 5B shows that the presence of LipL32 in the Triton X-114 fraction was confirmed [25]. These results demonstrated the correct functioning of the Triton X-114 fractionation method.
We also investigated the surface localization of Lp25 using proteolysis of intact cells of the L. interrogans serovar Copenhageni strain L1-130 using proteinase K. Fig 5C shows that Lp25 was susceptible to protease treatment in a dose-dependent manner, and the subsurface LipL32 was not susceptible, which suggests that Lp25 is exposed on the surface. These results are consistent with a previously published study that demonstrated that LipL32 was not exposed on the leptospiral surface, despite its localization in the outer membrane [13].
The renal manifestations of leptospirosis are variable and range from mild symptoms, such as low urinary protein excretion and sediment changes, to fatal AKI [4,27,28]. Severe cases of AKI are generally oliguric and hyperkalemic with a prolonged course and high mortality rate. Nonoliguric and normo- or hypokalemic AKI-forms are associated with a better prognosis [4, 27–29]. The renal pathophysiology that is consequent to leptospirosis infection is not clearly known despite advances in the knowledge of AKI epidemiology [4,5, 28–32]. Different factors may be involved, such as inflammatory processes, rhabdomyolysis, hemodynamic alterations, immune responses, and direct effects of leptospires and their products [30, 33, 31].
This study evaluated the effect of the LipL32 and Lp25 proteins on renal function in normal guinea pigs. We found that only Lp25 was associated with the development of oliguric AKI and rhabdomyolysis-induced hyperkalemia (elevated CPK, uric acid and serum phosphate).
Lp25 decreased the GFR compared LipL32 and control experiments (Fig 1). These results demonstrate, for the first time, that a specific protein from pathogenic Leptospira spp. plays an important role in the establishment of the AKI that is observed in Weil’s syndrome. In contrast, LipL32 protein did not produce a decrease in the GFR, despite a report that LipL32 induced interstitial nephritis-mediated gene expression in cultured mouse proximal tubule cells [34] and acute tubular injury in proximal pronephric ducts from zebrafish larvae kidneys [35]. These experiments were performed using in vitro preparations and another animal species. The decrease in GFR may not be directly dependent on the tubular damage because the FENa was not different between the three groups. Previous clinical studies also demonstrated that rhabdomyolysis-induced acute renal injuries did not modify sodium excretion [36–38].
Our results from cellular localization assays agree with previous studies that demonstrated that LipL32 was a subsurface protein that was not accessible on the leptospiral surface [13]. Notably, the role of LipL32 in Leptospira biology is not defined despite its abundant expression [14] in all pathogenic serovars. Murray and colleagues (2009) demonstrated that LipL32 was not required in acute (hamster) or chronic (rat) infection models for leptospirosis [39].
The results in Figs 1 and 2, such as the increase in potassium, creatine phosphokinase, uric acid and phosphate serum levels, are characteristics of the presence of rhabdomyolysis. The classification of the muscular lesions observed in histological preparations using a score (Fig 4) revealed that Lp25 induced the same type of muscular lesions as LipL32, but the Lp25 lesions were much more severe. Fig 3 shows areas of necrosis and moderate inflammatory infiltrate in the Lp25 group, and the LipL32 group exhibited small areas of necrosis and few inflammatory areas. Immunohistochemistry with anti-rabbit Lp25 and LipL32 antisera were positive for the two proteins and both proteins exhibited the same histological patterns. The deposition of the antigens was more intense in the Lp25 group. These results suggested that the LipL32 and Lp25 proteins reached the muscle tissue and induced lesions. However, lesions in the Lp25 group were more diffuse and apparently more accentuated than the LipL32 group.
We also demonstrated that Lp25 was a surface-exposed and conserved protein in pathogenic species of Leptospira. Previously published immunoblot studies using sera from leptospirosis patients and infected hamsters showed that Lp25 protein was expressed during the course of leptospiral infection [40,21]. This protein was recently included in the Leptospira endostatin-like (Len) family by the automatic NCBI prokaryotic genome re-annotation pipeline [41]. The members of the Len family bind plasminogen, laminin and human complement regulator factors [42–44, 18]. However, we previously demonstrated that Lp25 did not exhibit extracellular matrix-binding properties or play a role in immune evasion via interacting with the human complement regulator C4BP [17,45]. Comparative proteomic analyses of leptospira outer membrane proteins also demonstrated that the Lp25 protein, encoded by the LA0009 gene in L. interrogans serovar Lai, was up-regulated (1.3-fold) after an overnight upshift to 37°C [46]. These features are also compatible with one of the potential roles of the Lp25 protein, which is causing muscular damage that consequently is associated with oliguric AKI and hyperkalemia. These data demonstrated, for the first time, that Lp25 is associated with rhabdomyolysis, which is an important sign in leptospirosis and may underlie the muscular pain, which is a pathognomonic symptom of this disease.
|
10.1371/journal.pgen.0030187 | Adaptive Gene Expression Divergence Inferred from Population Genomics | Detailed studies of individual genes have shown that gene expression divergence often results from adaptive evolution of regulatory sequence. Genome-wide analyses, however, have yet to unite patterns of gene expression with polymorphism and divergence to infer population genetic mechanisms underlying expression evolution. Here, we combined genomic expression data—analyzed in a phylogenetic context—with whole genome light-shotgun sequence data from six Drosophila simulans lines and reference sequences from D. melanogaster and D. yakuba. These data allowed us to use molecular population genetics to test for neutral versus adaptive gene expression divergence on a genomic scale. We identified recent and recurrent adaptive evolution along the D. simulans lineage by contrasting sequence polymorphism within D. simulans to divergence from D. melanogaster and D. yakuba. Genes that evolved higher levels of expression in D. simulans have experienced adaptive evolution of the associated 3′ flanking and amino acid sequence. Concomitantly, these genes are also decelerating in their rates of protein evolution, which is in agreement with the finding that highly expressed genes evolve slowly. Interestingly, adaptive evolution in 5′ cis-regulatory regions did not correspond strongly with expression evolution. Our results provide a genomic view of the intimate link between selection acting on a phenotype and associated genic evolution.
| Changes in patterns of gene expression likely contribute greatly to phenotypic differences among closely related organisms. However, the evolutionary mechanisms, such as Darwinian selection and random genetic drift, which are underlying differences in patterns of expression, are only now being understood on a genomic level. We combine measurements of gene expression and whole-genome sequence data to investigate the relationship between the forces driving sequence evolution and expression divergence among closely related fruit flies. We find that Darwinian selection acting on regions that may control gene expression is associated with increases in gene expression levels. Investigation of the functional consequences of adaptive evolution on regulating gene expression is clearly warranted. The genetic tools available in Drosophila make functional experiments possible and will shed light on how closely related species have responded to reproductive, pathogenic, and environmental pressures.
| Changes in gene expression are governed primarily by the evolution of cis-acting elements and trans-acting factors. Several single-gene studies have combined data on expression, protein abundance, function, and sequence evolution to make powerful statements about the role of adaptive evolution in effecting phenotypic change [1,2]. These case studies of single genes focused on well-described pathways that were known, a priori, to have remarkable expression differences. As such, they may provide a biased view of the population genetic mechanisms controlling gene expression evolution. Thus, the question remains as to which forces, neutral or adaptive, predominate on a genomic level to bring about changes in gene expression.
Recent studies have tried to discern the causes of genome-wide expression evolution solely from patterns of gene expression variation within and among species [3–5]. Patterns of constant expression levels across several species combined with significantly elevated or reduced expression in a single species have been taken as evidence of lineage-specific adaptive evolution [3,4]. Alternatively, low levels of within-population variation in expression compared to divergence in expression among species has also been taken as evidence of adaptive evolution [5–7]. As these studies are based strictly on phenotypic data—expression variation—they are indirect indicators of the underlying genetic and population genetic phenomena. For example, elevated lineage-specific expression divergence can be explained equally well by directional selection or by reduced functional constraint. These studies highlight the importance of direct tests of the mechanisms of evolution. For example, Good et al. [8] used statistical inferences of adaptive protein evolution along with expression evolution to investigate the connection between the two. Their highly conservative test suggested that no significant connection existed. In an attempt to unite population genetic inference with expression data, Khaitovich et al. [9] found a positive correlation between linkage disequilibrium and expression divergence in genes expressed in the human brain. This result is consistent with recent adaptive evolution of cis-acting regulatory elements associated with brain-expressed genes, but could also be due to selection on protein function.
A global understanding of the population genetic processes acting on expression phenotypes requires both genomic expression data and genomic sequence variation and divergence data. Combining these data allows for the use of molecular population genetic tests to identify the underlying evolutionary mechanism. To this end, we combined expression data from three closely related species, D. simulans, D. melanogaster, and D. yakuba [6,10], with population genomic sequence data from D. simulans [11], and genome sequence data from D. melanogaster [12] and D. yakuba [11]. These data allow us to polarize both expression and sequence evolution to particular lineages. Additionally, we used the sequence data to mask expression probes (which were developed using the D. melanogaster reference) with sequence mismatches in D. simulans and D. yakuba. This approach has the critical advantage that it does not confound expression divergence with sequence evolution across lineages.
DNA polymorphism and divergence data allow one to directly test for both recent and recurrent directional selection on genes and noncoding regions associated with rapid changes in expression. If expression evolution were due to recent directional selection on cis-acting elements, we predict a reduction in the DNA heterozygosity to divergence ratio in flanking regions of genes showing expression evolution relative to genomic averages [13]. Alternatively, if recurrent directional selection has acted on cis-regulatory sequences controlling expression levels, one might observe excess fixations at regulatory sites relative to nearby “neutrally” evolving sites [14]. Finally, if gene expression diverges primarily due to trans-acting factors or neutral processes at cis-acting sites, one would expect no evidence of directional selection on noncoding sequences near genes showing expression divergence.
Here, we use population genomic and gene expression data from Drosophila to address the following questions: Is expression evolution associated with adaptive evolution of cis regions? Are genes with modified expression patterns also evolving modified protein function under directional selection? Are genes that change expression over short time scales clustered into distinct functional groups?
We reanalyzed previously collected expression data from adult male D. melanogaster, D. simulans, and D. yakuba from the Drosophila v1 Affymetrix GeneChip Array [6,10]. Sequence divergence of probe targets in D. simulans and D. yakuba could confound expression analysis [15], so mismatched probes were masked before analysis. After masking procedures, 4,427 probe sets remained, with an average of 3.81 (SE ± 1.01) probes per set. We defined genes that are increasing and decreasing in expression in D. simulans as those in the 5% tails of expression divergence from the D. melanogaster–D. simulans ancestor (see Materials and Methods).
Cis-regulatory element evolution directly affects transcription and mRNA half-life (see [16,17]). Cis-acting elements, such as core promoters, that regulate transcription are predominantly located in 5′ regions and those that control mRNA stability and degradation are primarily located in 3′ regions [16,17], although there is considerable variation among genes. We tested for evidence of an association between recent and recurrent directional selection in 5′ and 3′ flanking regions (which include UTRs and putative regulatory regions) and significant changes in expression levels.
Reductions in polymorphism relative to divergence indicate the action of recent directional selection [13]. Flanking regions with polymorphism to divergence ratios in the lowest 5% tail of the distribution were taken as having evidence of recent selective sweeps. Figure 1 depicts mean levels of polymorphism and divergence in 5′ and 3′ noncoding sequence. Flanking regions and UTRs have lower levels of polymorphism and divergence than silent sites, which is in agreement with previous findings that noncoding regions are under greater constraint than silent sites [13]. Genes with increased expression levels show more variability in levels of polymorphism and divergence over different features, but no strong pattern emerges. There is no evidence of hitchhiking effects in either 5′ or 3′ UTR or flanking regions in association with changes in expression (Figure 2; Table S1).
Using an extension of the McDonald-Kreitman test [14] for noncoding sites, we compared flanking polymorphic and fixed sites to synonymous sites of the corresponding gene to infer the action of recurrent directional selection. Genes with significant expression evolution show more evidence of recurrent directional selection in 3′ UTRs and 3′ flanking regions than expected by chance (Figure 2; Table S1). Genes with increases in expression drive this relationship. Although genes with reduced expression have more 3′ UTR and flanking region divergence than genes with no change in expression, the tests provide no strong evidence of recurrent adaptation associated with reduced gene expression (Figure 2; Table S1). The 5′ regulatory regions of genes with increased expression show the same trend, but again the result is not statistically significant (Figure 2; Table S1). Thus, recurrent adaptive evolution of 3′ cis-regulatory regions likely plays a critical role in adaptive expression increases.
The 3′ regulatory regions are bound by elements, such as microRNAs, that can stabilize or destabilize mRNA (see [18]). Given the linkage between adaptive evolution of 3′ regulatory regions and expression evolution, we hypothesized that microRNAs may be coevolving with their target genes. We retrieved information on known microRNAs and their targets in D. melanogaster from miRBase [19,20]. We found that those microRNAs that regulate a greater number of genes with changes in expression have faster, but not significantly faster, rates of evolution (Spearman's ρ = 0.2065, p = 0.1073). Rapid evolution of microRNAs and adaptive expression divergence associated with 3′ regions strongly motivate in-depth investigation of the 3′ flanking regions to uncover the functional mechanisms for transcriptional regulation of genes with significant expression evolution.
Increases in gene expression were more often associated with adaptive evolution than decreases in expression (Figure 2). This observation does not appear to be due to a bias in analysis of the data because expression changes are normally distributed and there is no correlation between estimated ancestral divergence and change in expression (see Materials and Methods). However, continually increasing expression levels cannot persist over long evolutionary time scales. In fact, expression levels are typically under strong stabilizing selection ([5], and see Materials and Methods). A speculative hypothesis for this observation relies on relaxation of codon bias. Begun et al. [11] documented an accumulation of fixations for unpreferred codons in D. simulans. If these unpreferred codons are slightly deleterious and reduce translational efficiency, regulatory regions may be under directional selection to compensate for this phenomenon by making more transcript available for translation.
As seen in previous research [6,8], genes with greater absolute levels of expression divergence evolve faster at the protein level (mean dN ± SE 0.0046 ± 0.0003 and 0.0034 ± 0.0001, for genes changing in expression and not changing, respectively; Wilcoxon: p < 0.0001; Table S2). Genes with rapid expression evolution are also represented by fewer expression probes per set (mean number of probes ± SE 2.98 ± 0.076 versus 3.90 ± 0.033; Wilcoxon: p < 0.0001). A rapid rate of sequence evolution would lead to more probe mismatch, which explains the observed pattern. This also renders our expression divergence analysis conservative, as our power to detect a significant expression difference is reduced for the most rapidly evolving genes. Interestingly, even though genes with significant increases in expression in D. simulans have higher average dN, they show decelerating dN in D. simulans relative to D. melanogaster and D. yakuba (resampling test: p = 0.023; method for relative rates described in Begun et al. [11]). The same is not true of genes with decreasing expression (p = 0.861). While higher average rates of amino acid evolution in genes with expression divergence could have been indicative of relaxed purifying selection, the deceleration in dN certainly speaks against that hypothesis. Previous work showed that high levels of expression correlate with lower rates of protein evolution [21–23], which may reflect selection for translational robustness [23] or translational accuracy [22]. The deceleration in protein evolution of genes with increases in expression is consistent with the idea of stronger translational selection on highly expressed genes, but overall, we see only a weak relationship between expression level and protein divergence (Spearman's ρ = −0.1821, p < 0.0001).
Genes adaptively evolving modified expression patterns may also be adaptively evolving modified protein function. We estimated the proportion of genes in each expression class—increasing, decreasing, and no change—with evidence for recurrent directional selection using the McDonald-Kreitman test [14]. For all genes in this analysis, the proportion undergoing recurrent adaptive evolution was similar to the genome-wide estimate [11]. The prevalence of recurrent adaptive evolution was not significantly different for genes showing expression evolution versus those showing no expression evolution (p = 0.4438; Figure 2 and Table S1).
We also tested for evidence of recent directional selection as measured by a reduction in the ratio of silent polymorphism to silent divergence [13]. Coding regions with ratios in the lowest 5% tail of the distribution were taken to have evidence for recent selective sweeps. A higher proportion of genes showing expression evolution have significantly reduced ratios of silent site polymorphism to divergence, which is consistent with recent selective sweeps (p = 0.0445; Figure 2 and Table S1). Genes with increased expression levels explain more of this relationship than genes with decreased expression (increase p = 0.0328, decrease p = 0.2530), although both sets have greater reductions of silent polymorphism to divergence ratios than genes that are not changing in expression.
The targets of these putative hitchhiking events may have been nearby regulatory regions in an intron or upstream or downstream of the protein coding region. Alternatively, one possible explanation for the association between upregulation and recent selection on coding regions is codon bias. Gene expression is positively correlated with codon bias [22]. Given this association, hitchhiking effects of preferred codons might increase with increasing levels of expression due to stronger selection for translational accuracy [22]. While there is a higher ratio of preferred to unpreferred polymorphisms and fixations in genes evolving increases in expression versus those that show no expression evolution, the difference is not statistically significant (Fisher's Exact Test: p ≫ 0.05 for both tests; Table 1). There may be a time lag between expression evolution and the fine-tuning of translation via codon bias. Thus, our data might mean that genes with the most extreme expression differences have recently increased expression. Alternatively, the hitchhiking events may result from adaptive evolution acting on one or a few amino acids or on nearby regulatory regions.
We used gene ontology information from Flybase and from the generic Gene Ontology Slim set of terms to determine whether certain functional classes of genes were more likely to evolve expression differences. Six ontology terms are significantly enriched for genes both with significant increases and decreases in expression (Table S3). Two of those terms, chymotrypsin and trypsin activity, have completely overlapping genes and are part of a larger category, serine-type endopeptidase activity. These genes have many functions, including reproduction, digestion, and immunity [24]. Three other categories, courtship behavior, negative regulation of transcription, and sex determination appear to be unrelated on the surface, but closer inspection of the genes in these categories reveals that all are involved in regulation of transcription or chromatin remodeling. These functions frequently evinced adaptive protein evolution in the genome-wide analysis of adaptive evolution in D. simulans [11]. This suggests that there may be a connection between adaptive protein evolution and expression divergence for some biological functions.
Because adaptive evolution of 3′ cis-regulatory regions may be driving expression divergence, at least for genes with increased expression, we examined the classes of genes associated with genes that have both evidence for adaptive 3′ evolution and significant expression divergence (Tables S4 and S5). We also investigated ontology terms associated with genes showing evidence of hitchhiking events and significant expression divergence (Table S6). Generally, genes with adaptive 3′ or protein evolution are found in the cytoplasm or are integral to the membrane. Their molecular functions are predominantly protein binding, nucleic acid binding, and translation related. The most common biological processes are related to response to stimuli, RNA regulation (binding, splicing, degradation), and metabolism.
In this study, we link adaptive sequence evolution to phenotypic change on a genome-wide scale. Several recent studies have illustrated the importance of adaptive evolution acting on noncoding DNA [11,25,26], and our data reinforce this point. More critically, we show that adaptive evolution of cis-acting elements in 3′ regions is clearly associated with and may be driving lineage-specific increases in expression that lead to phenotypic differences among species. Recent work suggests that genes with certain 5′ promoter elements show an increased interspecies variability in expression in yeast as well as Drosophila [27]. In contrast, our data implies that 3′ regulatory regions are playing a more critical role in adaptive expression divergence. Functional genomic investigation of these 3′ cis-regulatory regions is clearly warranted. The question now becomes, how and why do genes involved in important processes such as chromatin remodeling change their expression patterns through 3′ cis-acting regulatory adaptive evolution?
We reanalyzed expression data from 3-d-old virgin adult males of one isogenic line of D. melanogaster, ten isogenic lines of D. simulans, and one isogenic line of D. yakuba [6,10]. Three replicate chips for each line were used. All data were collected at the same location under standard conditions using the Affymetrix GeneChip Arrays (Drosophila 1.0), which contain 13,966 features representing the genome of D. melanogaster. Because the D. melanogaster gene annotation has been updated since the array was developed, we compared probe sequences to the D. melanogaster genome to determine which genes were targeted with each probe set.
The probes representing features on the Affymetrix GeneChip Arrays are constructed for D. melanogaster and are not expected to perfectly match other species. Prior research suggests that such imperfect matches cause incorrect measures of expression due to poor hybridization [10,15,28]. To account for the confounding effect of probe sequence divergence among species on gene expression measures, only probes that were identical matches to the genome sequences of D. melanogaster, D. simulans, and D. yakuba were included in analyses. Probes showing any divergence among the probe sequence on the array and the genome sequences of the three species were masked. Probe sets with fewer than two probes remaining after masking (out of the original 14) were removed before downstream analyses. Finally, probe sets that bound to overlapping genes or homologous sequence of multiple genes were also removed, as the signal could not be attributed to a single gene.
After probe-masking procedures, all chips were normalized and expression intensities were calculated using gcrma from the affy package available in Bioconductor [29,30]. The mean of the log2 expression intensity for each probe set was then calculated for each species. Probe sets for which the log2 mean intensity of at least one species was not greater than three were considered absent. Of the original 195,944 probes from 13,996 probe sets, 16,850 probes representing 4,427 probe sets remained after masking and removing probe sets with no detectable expression in either D. melanogaster or D. simulans (all expression data are in Table S7). The distribution of expression intensities was highly similar between species (Figure S1) and probe set intensities were highly correlated between species (Spearman's ρ = 0.92 between D. simulans and D. melanogaster and ρ = 0.89 between D. simulans and D. yakuba). However, probe sets with fewer probes have higher coefficients of variation in D. simulans and in D. melanogaster (Kruskal-Wallis tests: p < 0.0001 for all four tests). We tested whether probe sets with fewer probes gave reliable estimates of mean expression intensity. We randomly sampled four probes from probe sets that had all 14 probes remaining after masking. The mean expression intensity of the sample was highly correlated with the mean intensity estimated from all 14 probes (Spearman's ρ = 0.869). The mean expression level varied by +/− 7%, and the variance in expression among replicates increased by 22%.
Ancestral expression states were reconstructed using AncML v 1.0 [31] using the average of normalized log2 expression values for each species. Expression divergence was calculated as follows:
where Esim is the expression level of D. simulans and EAncmel-sim is the estimated expression level of the D. simulans/melanogaster ancestor. Figure S2 depicts the distribution of expression change along the D. simulans lineage. The distribution is not significantly different from normally distributed. Additionally, there is no correlation between change in expression along the D. simulans branch and the expression level of the inferred ancestor (Figure S3). The conical nature of Figure S3 reflects the negative correlation between expression level and expression divergence over short evolutionary time scales. We defined genes that are increasing and decreasing in expression in D. simulans as those in the 5% tails of expression divergence from the D. melanogaster–D. simulans ancestor. We calculated confidence intervals (CI) around the expression values for D. simulans and determined whether the D. melanogaster expression estimate fell within the D. simulans CI. Intraspecific expression divergence values in the tails are not normally distributed, so we calculated CIs in R using bias correction and acceleration [32]. One probe set (of 221) with increasing expression and four probe sets (of 221) with decreasing expression along the D. simulans lineage had mean intensities in D. melanogaster within the 95% CIs of D. simulans.
Drosophila simulans and D. yakuba syntenic assemblies are described in Begun et al. [11] and information on the D. yakuba genome project can be found at http://genome.wustl.edu. From light-shotgun sequencing of six lines of D. simulans, a total of 109 Mbp of euchromatic sequence were covered by at least one of the six lines. Each line had 43%–90% coverage of that 109 Mbp with an average of 3.6 alleles per site. However, coverage of genic regions was somewhat higher at 3.9 alleles per site.
Genes and Affymetrix probes were localized using the Flybase v.4.2 annotation (http://flybase.org/annot). Genes included were from two categories. The first set maintained the gene model of D. melanogaster meaning that, in D. simulans, they have canonical translation initiation codons (or that matched the D. melanogaster noncanonical codon), canonical splice junctions at the same position as D. melanogaster (or noncanonical splice junctions that were identical to the D. melanogaster nucleotides at splice sites), no premature termination, and a canonical termination codon. The second set was less conservative in that the gene could have a different gene model with respect to only one of the aforementioned criteria (i.e., either a noncanonical translation initiation codon at the D. melanogaster initiation site, or noncanonical splice junctions, or lack a termination codon at the D. melanogaster termination). Additionally, genes with premature terminations in the last exon were included. There were very few genes with imperfect models in any of the expression groups (10/212 with increased expression, 14/210 with decreased expression, and 173/3,814 with no change in expression). Only gold collection UTRs (i.e., those with completely sequenced cDNAs) were used in analyses (http://www.fruitfly.org/EST/gold_collection.shtml). Flanking regions consisted of sequence 1,000 bases upstream and downstream of any annotated UTR sequence for each gene (or initiation/termination codons for genes without annotated UTRs). Flanking sequence was truncated if the coding sequence of a neighboring gene was within the 1,000 bases. We also investigated 300 bases upstream of the 5′ UTR (see Table S1), which would target core promoter regions, and recovered the same results as with 1,000 bases upstream.
Some statistical tests were performed using JMP IN v5.1 (SAS Institute). PERL scripts for calculations of estimated nucleotide diversity (π), McDonald-Kreitman tests, and resampling tests were written by and can be obtained from AKH. Nucleotide diversity was estimated as in Begun et al. [11] for each genomic feature (exon, intron, UTRs, flanking) that had a minimum number of nucleotides represented [i.e., n (n − 1) × s ≥ 100, where n = average number of alleles sampled and s = number of sites]. The measure of nucleotide diversity, π, is the coverage-weighted average expected heterozygosity of nucleotide variants and is therefore an unbiased estimate of polymorphism. For coding regions, the numbers of silent and replacement sites were counted using the method of Nei and Gojobori [33]. The pathway between two codons was calculated as the average number of silent and replacement changes from all possible paths between the pair. Estimates of π on the X chromosome were corrected for sample size [π w = π × (4/3)] under the assumption that males and females have equal population sizes. Lineage-specific divergence was estimated by maximum likelihood using PAML v3.14 [34] and was reported as a weighted average over each D. simulans line with greater than 50 aligned sites in the segment being analyzed. PAML was run in batch mode using a BioPerl wrapper [35]. For noncoding regions, we used baseml with HKY as the model of evolution to account for transition/transversion bias and unequal base frequencies [36], and for coding regions we used codeml with codon frequencies estimated from the data. For all genes, 0.001 was added to heterozygosity and divergence values so that we could calculate ratios for genes with entries of zero. We did not analyze genes with zero values for both heterozygosity and divergence. Even after correction for smaller effective population sizes, heterozygosity at silent sites is significantly lower on the X chromosome than on autosomes (Kruskal-Wallis test: p < 0.0001, Tukey's HSD shows X is different from all autosomes), so we defined significantly low heterozygosity/divergence ratios separately for the X and autosomes. For each feature, genes in the lowest 5% tail of silent site heterozygosity/divergence ratios were defined as being significantly low and therefore showing evidence of a recent selective sweep. Those ratios defined as having evidence of recent selective sweeps were at least 10-fold lower than the mean ratio for all features. D. simulans–specific accelerations/decelerations in protein evolution were calculated as described in Begun et al. [11].
Polarized MK tests minimized the numbers of nonsynonymous substitutions and required that D. melanogaster and D. yakuba share the same codon to ensure that fixations and polymorphisms were attributable to evolution along the D. simulans lineage. We used a derivative of the McDonald-Kreitman test [14] to evaluate evidence for recurrent directional selection in noncoding regions. Polymorphic and fixed sites of noncoding DNA were compared to polymorphic and fixed silent sites of the gene. Again, we only analyzed sites where D. melanogaster and D. yakuba shared the same nucleotide.
With very few polymorphisms and fixations there is little power to detect the action of directional selection. Therefore, we imposed a minimum row and column count for tests to be included in downstream analyses. We required that each row and column in the 2 × 2 table have a sum of at least five observations. We also removed any tests that had a significant test result but that had a neutrality index value greater than one, (which indicates excess amino acid/noncoding polymorphism not directional selection [37]) in order to calculate the proportion of genes that are experiencing recurrent directional selection. All data for D. simulans heterozygosity, lineage-specific divergence and MK tests are listed in Table S8. Substitutions to preferred and unpreferred codons were estimated by a parsimony method developed by Y.-P. Poh [11].
For each category of interest (e.g., increasing or decreasing expression levels), we calculated the proportion of genes with a significant test result (for MK tests, p ≤ 0.05, for heterozygosity/divergence ratios were considered significant if they fell in the 5% tail). We then tested whether this proportion was significantly greater than the random expectation using resampling tests. We randomly drew n p-values from the set of all genes where n is the number of genes in the category. We repeated this procedure 10,000 times to get the empirical distribution of proportion genes with significant tests.
We obtained cellular component, molecular function, and biological process ontology terms from the Flybase gene ontology terms (http://flybase.org/genes/lk/function) in combination with the generic Gene Ontology Slim set of ontology terms (http://geneontology.org/GO.slims.shtml#avail). The proportion of genes with significant expression evolution was calculated for each ontology term. We determined whether each ontology term had a higher proportion of genes with significant D. simulans expression divergence than would be expected from the empirical distribution. We derived the empirical distribution for each ontology term by drawing the same number of genes as was in the term from all genes with expression data. We then calculated the proportion in the resampled dataset with significant expression evolution. We used 10,000 resampled data sets to derive the empirical distribution for each term. |
10.1371/journal.pntd.0003099 | Pre-miR-146a (rs2910164 G>C) Single Nucleotide Polymorphism Is Genetically and Functionally Associated with Leprosy | Mycobacterium leprae infects macrophages and Schwann cells inducing a gene expression program to facilitate its replication and progression to disease. MicroRNAs (miRNAs) are key regulators of gene expression and could be involved during the infection. To address the genetic influence of miRNAs in leprosy, we enrolled 1,098 individuals and conducted a case-control analysis in order to study four miRNAs genes containing single nucleotide polymorphism (miRSNP). We tested miRSNP-125a (rs12975333 G>T), miRSNP-223 (rs34952329 *>T), miRSNP-196a-2 (rs11614913 C>T) and miRSNP-146a (rs2910164 G>C). Amongst them, miRSNP-146a was the unique gene associated with risk to leprosy per se (GC OR = 1.44, p = 0.04; CC OR = 2.18, p = 0.0091). We replicated this finding showing that the C-allele was over-transmitted (p = 0.003) using a transmission-disequilibrium test. A functional analysis revealed that live M. leprae (MOI 100∶1) was able to induce miR-146a expression in THP-1 (p<0.05). Furthermore, pure neural leprosy biopsies expressed augmented levels of that miRNA as compared to biopsy samples from neuropathies not related with leprosy (p = 0.001). Interestingly, carriers of the risk variant (C-allele) produce higher levels of mature miR-146a in nerves (p = 0.04). From skin biopsies, although we observed augmented levels of miR-146a, we were not able to correlate it with a particular clinical form or neither host genotype. MiR-146a is known to modulate TNF levels, thus we assessed TNF expression (nerve biopsies) and released by peripheral blood mononuclear cells infected with BCG Moreau. In both cases lower TNF levels correlates with subjects carrying the risk C-allele, (p = 0.0453 and p = 0.0352; respectively), which is consistent with an immunomodulatory role of this miRNA in leprosy.
| In spite of the successful drug therapy, leprosy is still affecting people worldwide. It is well known that host genetic background influences leprosy development and that genetic variants have been associated with the disease. Therefore we conducted a study to evaluate the role of microRNAs (miRNAs) polymorphisms in leprosy. We observed that a polymorphism in miR-146a is associated with the risk to develop leprosy in Brazilians. Based on the analysis of clinical specimens, we found that the genetic variant was correlated with elevated levels of miR-146a and it is also a negative regulator of tumor necrosis factor (TNF), an important inflammatory mediator in the leprosy context. These findings provide tenable evidences that miR-146a is important in the control of gene expression during M. leprae infection and also may contribute with leprosy development by controlling TNF levels.
| Leprosy is an ancient disease caused by Mycobacterium leprae. Once infected, the majority of individuals may clear the bacilli through a natural resistant response. Nevertheless, some patients could develop a latent infection that eventually evolves to one of the clinical symptomatic forms of leprosy. Peripheral nerves Schwann cells and skin macrophages are preferentially invaded, evoking a chronic infection that may take years to become active. Once the disease is established, a range of immune responses occur in spite of M. leprae been genetically conserved [1]. The spectral clinical manifestations are classified in a five-group system proposed in the 1960s by Ridley and Jopling [2]. A classic view of predominant Th1 for tuberculoid (TT) pole where a localized form of the disease is observed, in contrast to a major Th2 profile, where a disseminated form, called lepromatous (LL) pole is verified [3]. This classification system also comprises intermediate phenotypes, known as borderline, that interpose those two well characterized poles. Also, a variable percentage of the patients can experience an abrupt inflammatory episodes during the natural course of the disease, which are called type I (reversal) or type II (erythema nodosum leprosum) reactions [4], [5]. Patients at the onset of the episodes exhibit high cytokine levels that are decreased once anti-inflammatory drugs are effective [6]–[8], while genetic association might also be important [9].
Host susceptibility or protection is associated with the complex interaction between environment and genetic background, leading to different outcomes. Several publications aimed to understand the genetic contribution to leprosy risk or protection using different approaches including: twin studies, family-based linkage analysis, candidate gene association and genome wide association studies [10]–[13]. In fact, studies are linking or associating genes that have been generating a compelling amount of evidence to confirm the genetic influence in leprosy outcome. For instance, genes associated with innate immune response, like TLR1, NOD2, and PARK2 [11], [14]–[16] or adaptive immune responses, such as IL10, IFNG and LTA/TNF/HLA have been consistently associated with leprosy [9], [13], [17]–[19].
Recently, microRNAs (miRNAs) have been described as novel regulators of innate and adaptive immune responses, although a few data reported its involvement in leprosy. MiRNA genes are transcribed by RNA polymerase II [20], resulting in a hairpin primary-miRNA (pri-miRNA) that is processed, in a cascade, by different RNAses [21] generating pre-miRNA, and finally the mature miRNA strand facilitating the miR-RISC (RNA-induced silencing complex) assembly [20], [22]. The miRNAs control gene expression at post-translational level by pairing with 3′-untranslated regions [23] leading to mRNA cleavage or translational repression [24]. Given that, it is possible to assume that the presence of polymorphisms along double-stranded sequences can affect miRNA expression and gene silencing [25]. Genetic variants in miRNA precursors, miR-196a-2 (rs11614913 C>T) and miR-146a (rs2910164 G>C) have been associated with cancer and tuberculosis [26]–[30]. Here, we conducted a case-control and a family-based study to test these miRNA SNPs with leprosy susceptibility. Further, we performed functional studies using cell cultures and biopsies from skin and nerves to investigate miRNA mature expression form to define a genotype-phenotype correlation.
The case-control study includes a total of 1,098 individuals from Rio de Janeiro; of these, the 491 patients were recruited from the Souza Araújo outpatient unit, located at Fundação Oswaldo Cruz (FIOCRUZ), Rio de Janeiro, Brazil. The data for 607 controls was obtained from a bone marrow donors' bank in Rio de Janeiro comprising of samples from local healthy individuals. A detailed presentation of this population has been described in Table S1 and elsewhere [16], [31]. A replication population was also tested. Subjects for the family-based study were enrolled from Duque de Caxias, a hyper endemic city from the Rio de Janeiro state (Table S2). This population exhibited 97 nuclear families (426 subjects) [31]. All patients were routinely diagnosed according to Ridley and Jopling criteria (1966). Also, we adopted the World Health Organization (WHO) classification for treatment purposes, and patients were classified as paucibacillary/PB (including TT and borderline-tuberculoid) and multibacillary/MB (including LL, borderline-lepromatous and borderline-borderline). Population characteristics according to the WHO classification and reactional status are summarized in Table S1 and S2. All patients signed an informed consent and this project was approved by the institutional ethics committees from the involved institutions.
Nerve biopsy samples were obtained at Souza Araújo outpatient unit. A detailed description of nerve samples and clinical forms was previously published [32]. To perform the correlation of TNF mRNA expression with miR-146a genotype we used 33 nerve samples (19 diagnosed with leprosy and 14 with other neuropathies). Among these specimens, we were able to determine miR-146a expression in 12 samples from leprosy patients and 7 from other peripheral neuropathies. In the group of neuropathies other than leprosy, our clinicians were able to accurately diagnose three out of 7 patients. Among those there was: chronic inflammatory demyelinating polyneuropathy (CIDP, n = 2); and one case of systemic lupus erythematous. All undiagnosed patients returned to their neurological clinic for follow-up.
Skin biopsies were obtained from patients who live in Rondonópolis (Mato Grosso State, Brazil), enrolled and diagnosed by professionals from Instituto Lauro de Souza Lima (Bauru city, São Paulo State, Brazil). These specimens comprise 54 skin samples, amongst which 17 patients were diagnosed as MB [borderline borderline (BB) = 10, borderline lepromatous (BL) = 2, LL = 5; distributed as 3 women and 14 men; mean age: 41.6±9.6]. Thirty seven patients were classified as PB [borderline tuberculoid (BT) = 21 and TT = 16; distributed as17 women and 20 men; mean age: 43.9±16.9].
The sample collection and procedures described in this work were approved by the Oswaldo Cruz Foundation (FIOCRUZ) and Instituto Lauro de Souza Lima (ILSL) ethics committees. All patients or their parents/guardians signed a written informed consent (IRB protocol - Fiocruz 151/01 and ILSL 172/09).
M. bovis BCG Moreau strain (obtained from Fundação Ataulpho Paiva, Rio de Janeiro, Brazil) was cultured, for about 2 weeks, in Middlebrook 7H9 (Invitrogen, Carlsbad, CA) containing 0.02% glycerol and enriched with 10% ADC Middlebrook and 0.5% Tween-80 at 37°C as described elsewhere [32]. Live M. leprae from Instituto Lauro de Souza Lima (Bauru, São Paulo) was aseptically cultured in footpads of athymic NU/NU mice, purified and enumerated using methods described previously [33]–[35]. All infections experiments with live M. leprae were conduct at 33°C. A portion of live M. leprae was irradiated with ionizing radiation 10 kiloGray (Acelétrica Ltda).
THP-1 cells were purchased from American Type Culture Collection (ATCC, Rockville, EUA) and cultivated with RPMI-1640 (LGC Biotecnologia, Brazil) supplemented with 2 mM L-Glutamine, 100 U/mL penicillin, 100 µg/mL streptomycin and 10% heat-inactivated FBS (HyClone Laboratories, Canada) at 37°C, 5% CO2. Before infection, cells (5×105/well) were differentiated into macrophage-like cells (mTHP-1) using 80 nM phorbol 12-myristate 13-acetate (PMA, Sigma-Aldrich) for 24 h. Then, mTHP-1 were washed with PBS (1×), which was replaced by fresh antibiotics-free medium. Subsequently, stimulation (3, 24, and 48 h) was performed with irradiated or live M. leprae (Multiplicity of Infection - MOI 10∶1, 100∶1) at 33°C. After infection, total RNA was extracted as described below.
PBMC from healthy donors were collected in K3EDTA-tube (Labor Import Com. Imp. Exp. Ltda, Brazil), and isolated by Ficoll-Hypaque density gradient. After centrifugation (2,500 rpm, 30 min, 25°C), the interface containing mononuclear cells monolayer was collected, washed twice with PBS 1×, and cultivated in RPMI 1640 (LGC Biotecnologia, Brazil) supplemented with 10% human A/B RH+ serum (Sigma-Aldrich).
After isolation, PBMC were carefully selected according to miR-146a genotype (rs2910164 G>C), as described below. Cells were infected with BCG Moreau strain (MOI 10∶1) for 24 hours at 37°C. BCG was used as a surrogate model for M. leprae infections since it is a better inducer of TNF and also because this mycobacteria is able to induce miR-146a expression in vitro (data not shown). The supernatant was collected and detection to TNF levels was evaluated by Enzyme-linked immunosorbent assay kit DuoSet (R&D Systems, EUA) according to the manufacturer's protocol. Samples optical density (OD) was taken at 450 nm and estimated based on a standard curve, ranging (15.6–1000 pg/mL). Measurements were performed in duplicate.
Genomic DNA for the genetic study was extracted from peripheral blood or directly from PBMCs aliquots according to salting-out method as described [36]. RNA, and then DNA, from skin and nerve biopsy specimens were extracted using Trizol (Invitrogen) according to the manufacturer's instructions [37]. After DNA extraction, all samples were genotyped for SNPs as described below. Total amount of nucleic acids and purity were measured at NanoDrop ND-1000 (Thermo Scientific) instrument. The quality inspection of RNA was tested using agarose gel electrophoresis (1.2%) of 200 ng of SYBR Green II-stained RNA visualized at a transiluminator system (L-Pix Touch, Loccus Biotecnologia).
Allelic discrimination was performed using TaqMan Genotyping Assay (Applied Biosystems, CA, USA) for miR-196a-2 (rs11614913 C>T), miR-146a (rs2910164 G>C), miR-125a (rs12975333 G>T) and miR-223 (rs34952329 *>T) SNP. DNA (10–50 ng) amplification was performed in a final volume of 5 µL (2.5 µL of the TaqMan Genotyping Master Mix (Applied Biosystems), 0.125 µL of the TaqMan primers and probes). For miRNA expression analysis we performed a pooled-RT by using a set of specific stem-loop primers for each target (miR-146a, RNU44, RNU48) as indicated by the manufactures (TaqMan, Applied Biosystems). Briefly, non-denatured total RNA (200 ng) was incubated with RT primer pool (0.02×), dNTP (2 mM), Superscript III (10 U/µL, Invitrogen), RNase inhibitor (0.253 U/µL, Invitrogen) and first strand buffer (1×) in a final volume of 15 µL. The cDNA obtained was diluted (1∶6) and 2 µL were subjected to real-time PCR reaction, at a final volume of 10 µL. For RPL13a and TNF expression, total RNA (500 ng) was reversed transcribed following Superscript III manufacturer's instruction (Invitrogen). Then, 5 µL of diluted cDNA (1∶5) was amplified by real time PCR using SYBR Green PCR Master Mix (1×) and primers (0.5 µM) at the final volume of 20 µL. Both genotyping and miRNA expression were run on a StepOne Plus thermocycler detection system (Applied Biosystems). Specifically for TNF mRNA expression in nerve biopsies, data was retrieved from previous experiment and reanalyzed stratifying patients according to genotypes. Quantitative RT-PCR was performed using Biomark multiplex assays (Fluidigm, CA) as previously described [32].
After genotyping, we performed the Hardy-Weinberg equilibrium (HWE) analysis by chi-square tests. Then, we determined the genotypic, allelic and minor allele carriers frequencies, in order to perform comparisons between case and control groups. Genotypes and alleles with higher frequency were taken as baseline. The measure of allelic and genotypic association with leprosy was estimated by the Odds Ratio (OR) values generated after the application of a logistic regression model as described in detail elsewhere [31]. We also assessed an OR value adjusted for sex, ethnicity and age-at-onset for all comparisons evaluated (leprosy per se, subgroup PB–MB, reaction per se, and types of leprosy reactions). We also performed a case-control analysis using age as a categorical variable, for that analysis we adjusted for sex and ethnicity. Case-control study statistical analysis was performed using the packages genetics and coin from open source software R version 2.12.2 (available at http://www.R-project.org/). MiR-146a allele-dose effect (GC/CC) was determined by the Cochran–Armitage trend test. The analysis of the family-based transmission/disequilibrium test (TDT) was performed with FBAT software, version 2.0.2c. The TDT allows exploring if miR-146a C-allele is transmitted, from heterozygous parents to its affected child [38]. The amount of transmitted alleles was determined in the software Haploview [39]. We first performed the TDT analysis considering all affected child, regardless of their clinical form. A second analysis was conducted to verify the allele transmission to affected child according to their PB/MB status. To test for statistical significance among subgroup analysis we performed heterogeneity testing determined by Cochran's Q statistic.
Gene expression statistical analyses were done using Prism 5 (GraphPad software). Two-tailed Mann-Whitney t-test was applied for two sample group comparisons. For multiple testing, Kruskal-Wallis test was used followed by Dunn's post-test. Data is presented as mean ± SEM, except for ELISA (median). The value of p<0.05 was taken as statistically significant.
Allelic, genotypic and carrier frequencies were determined in both cases and controls and they did not deviate from HWE for SNPs miR-146a and miR-196a-2 (Table 1). The other two tested SNPs (miR-125a and miR-223) proved not polymorphic in our population. No association was detected for miRSNP-196a-2 genotypic or allelic frequencies with leprosy per se. Nevertheless, the polymorphisms of miR-146a gene have a susceptibility effect to leprosy per se for genotypes (GC ORadjusted = 1.44; p = 0.04 and CC ORadjusted = 2.18; p = 0.0091). This effect was also observed for allelic frequencies and C-allele carriers (ORadjusted = 1.47; p = 0.03 and ORadjusted = 1.56; p = 0.008; respectively). The genotypic OR values prompted us to investigate if it was directly proportional to the C-allele presence in each genotype (allele-dose effect), which was confirmed by applying the Cochran–Armitage trend test (χ2 = 96.6, p = 2.2×10−16). Interestingly, when we evaluated the influence of age-at-leprosy diagnosis in the association effect of miR-146a, we found a stronger effect in the subgroup correspondent from 25 to 34 years/old (Figure S1 and Table S3). Furthermore, we performed a comparison between controls and clinical forms, as stratified as PB and MB (Table S4). Once again, we could not find any association with miR-196a-2. Nonetheless, for both comparisons (PB vs. control or MB vs. control) we observed the risk association of C-allele for miR-146a, which was more prominent in the PB group. Significance levels were maintained in different comparison levels in PB subgroup versus controls rather than MB subgroup (Table S4). But, the heterogeneity test demonstrated no statistical significance (p-value = 0.40)
A confirmation of the genetic finding for miR-146a was observed after TDT analysis (Table 2). Considering leprosy per se, twenty-eight over 41-affected informative patients, received the C-allele indicating its over-transmission (p = 0.003). When we stratified affected individuals according to their PB or MB classification, the TDT for PB patients revealed sixteen over twenty one-affected transmissions (p = 0.01). The number of transmissions of the C-allele for MB affected patients were not significant (p = 0.23) (Table S5). These data confirm that miR-146a, in our replication sample, is associated with leprosy per se, however we could not provide sufficient evidence to infer subtype specificity.
In order to evaluate if those miRSNPs could be associated with leprosy reaction episodes, we subdivided only our patient group in (1) controls, patients without occurrence of reactional episodes and (2) cases, patients who exhibited only one type of leprosy reactional (LR) episodes (erythema nodosum leprosum, ENL or reverse reaction, RR). Those who have experienced both episodes were excluded from analysis. We could not observe any association between miRSNP-196a-2 and leprosy reactions (data not shown). Although we found that CC genotypes showed borderline association with protection to leprosy reactions (Table S6), and ENL as outcome, no statistical significance was found after adjustment for the covariates gender, ethnicity and age (Table S7).
Previous results have shown that specific pathways associated with pro-mycobacterial profiles that reprogram cellular environment to establish a suitable niche for bacterial survival are dependent on M. leprae viability [40]. So, we asked whether the functional role of miR-146a was dependent on live M. leprae. Irradiated M. leprae did not induce miR-146a expression (data not shown), on the other hand, live M. leprae infection for 3, 24 and 48 h at two different MOIs (10∶1 and 100∶1) induces miR-146a expression at 100 bacilli per cell (Figure 1). MiR-146a expression started at 24 h and was sustained until 48 h of infection.
So far, we observed an associated SNP in a gene that was being up-regulated by M. leprae infection. Then, we decided to explore miR-146a expression in skin and nerve biopsy samples trying to correlate genetic, clinical and biological findings. Initially, we determined miR-146a levels in nerve leprosy patients (L) and performed a comparison with biopsies from patients with non-leprous (NL) neuropathies. We found that miR-146a is more expressed in leprosy nerve biopsies group than in NL biopsies (Figure 2A). The examination of miR-146a expression according to host genotype revealed that carriers of C-allele were able to produce high levels of the mature miRNA (Figure 2B) in nerves. We determined miR-146a levels in skin biopsies from leprosy patients and performed a comparison between the MB and PB groups. Our results showed that biopsies from patients in both clinical forms express moderate-to-high levels of this miRNA, but no difference between MB and PB was detected (Figure 2C). Stratification according to Ridley and Jopling (R&J) clinical forms (LL, BL, BB, BT and TT) did not show any differences in miR-146a levels (data not shown). Furthermore, stratification by the risk allele (C-carriers) showed a tendency of augmented miR-146a expression in skin biopsies, although not significant (Figure 2D). Comparisons indicate that miR-146a C-allele seems to induce higher levels of miR-146a, which was also increased in leprosy patients, although no clustering was observed between clinical forms.
It was previously reported that miR-146a negatively regulates cytokines in primary peritoneal macrophages of mice, such as TNF [41]. Therefore, at this point, our hypothesis was whether the polymorphism associated with risk (C-allele) correlates to lower levels of TNF. For this, we analyzed the expression of TNF in nerve biopsies [32] and stratified according to the genotypes. As shown in Figure 3A, the presence of C-allele is associated with a reduction of TNF expression (p = 0.045), irrespective the disease type (L or NL). Then, we tested if miR-146a genotypes could influence TNF secretion. For that purpose, the cells were either left uninfected or infected with BCG Moreau and we compared infected groups with different genotypes. As shown in Figure 3B, the presence of the C-allele was related with less TNF secretion when comparing it to its control, GG-infected genotype (p = 0.0352).
In this study we showed that a SNP in miR-146a (rs2910164G>C), located in a cytokine cluster (5q31) associated with autoimmunity [42], [43] and Crohn's disease [44], was associated with leprosy susceptibility. Interestingly, live M. leprae up regulates this miRNA and carriers of the risk allele were also expressing more miR-146a. Furthermore, we were able to correlate lower levels of TNF with the presence of the risk allele. We also selected other two candidate miRNA SNPs (miR-125a, miR-223) previously identified as associated with regulation of immune responses [45], [46], but neither were polymorphic in Brazilians. The miR-196a-2 were chosen based on their involvement in Crohn's disease [47]. Nonetheless, we could not find any association between miR-196a-2 and leprosy although a previous report provided evidence to a common genetic fingerprint in Leprosy and Crohn's disease [48].
It was reported that miR-146a (rs2910164) GC polymorphisms plays an important role in papillary thyroid carcinoma while CC genotype are linked with risk and the reduction of survival in patients with glioma [49]. Controversial studies concerning susceptibility to cancer were investigated by a meta-analysis. They could not find a pattern between the SNP and the tumor type, conversely, the study pinpointed that there is an association between GG variant genotypes and increased risk of cancer among Asians [50], maybe reflecting the heterogeneity of the disease. Considering mycobacteria infections, it was demonstrated that the G-allele has an association with pulmonary tuberculosis in different directions in Han (protection) and Tibetan (risk) populations [51]. Here, we provide consistent evidence of G>C miRSNP-146a associated with leprosy among Brazilians. Using two different study designs, case-control and family-based, we found that the C-allele was strongly associated with susceptibility to leprosy per se and age-at-diagnosis was an important adjustment for the association, which was also suggested previously in leprosy [52], [53]. For all case-control comparisons, we tested the miRSNP-146a association considering sex and ethnicity with or without (data not shown) age as covariate, but the results remain unaltered after the inclusion of age-at-diagnosis correction. However, considering age subsets independently, a stronger association in the early-onset leprosy (25–34 years/old) was detected. This last observation is consistent with the idea that the early-onset may reflect a stronger genetic effect [17], [52], [54]. Curiously, the genetic design using leprosy reactions as outcome suggested an association between CC genotypes and LR protection towards protection, although not statistically significant after correction considering or not the covariate age.
MiR-146a mature form contributes to the reduction of TNF synthesis by down-regulation of adapter molecules IRAK1/TRAF6 through 3′UTR matching [55]. In THP-1 ectopically super-expressing miR-146a, Boldin and coworkers described that the exacerbated immune response was down-regulated by the reduction of TNF and IL-6 levels. Also, they found an uncontrolled autoimmune profile in miR-146a−/− knockout mice, as the animals were hyper-responsive to LPS challenge, producing high levels of those pro-inflammatory cytokines TNF, IL-6 [41] that was also in agreement with previous reports [56], [57] and our results here. It was recently shown that M. bovis BCG induces miR-146a expression and regulates TNF levels [58]. In our model, only live M. leprae was able to stimulate miR-146a expression. A recent paper from Siddle and colleagues, identified some SNPs in miRNA genes as markers of expression of quantitative trait loci (eQTL) in dendritic cells infected with M. tuberculosis [54], [59]. In fact, it has been proposed that SNPs along the strands that generate miRNAs can have great impact on both biogenesis of mature miRNA as well as the gain or loss of function of a particular miRNA [25], [60]. The miRSNP-146a is localized in the precursor strand and involves a shift of G∶U pair to C∶U mismatch. Jazdzewski showed that miR-146a expression was lower in the presence of the C-allele when compared to the G-allele [27], [28]; confirmed by others reports on cancer [61], [62]. In our analysis, miR-146a expression in nerve biopsies from leprosy patients revealed a different pattern: C-allele carriers are related with the high levels of the mature miR-146a. In agreement with our findings, it was shown by Kogo and colleagues that miR-146a was highly expressed in carriers of CC genotype than GG in both healthy and tumor tissues from patients with gastric cancer [63]. Also, in lupus a study showed that the presence of the C-allele correlates with increased expression of the mature miR-146a [64]. Nevertheless, we did not observe differences in skin biopsies from PB and MB patients. Perhaps, in this case, it seems that rs2910164 G>C SNP might impact miR-146a expression very early in the progression from latent infection to active disease, since pure neural form could be considered an earlier stage of the leprosy development. We could hypothesize that in the early stages of progression towards active disease, miR-146a expression in the macrophages/dendritic cells may be differentially regulating cytokine secretion and the emergence of T cell specific subpopulations precipitating disease outcome [65]. This controversy in the literature suggests that the presence of this SNP G>C (rs2910164), and maybe others, may govern the expression of mature miRNAs.
Recently it has been shown that miR-21 targeting CYP27b1, an enzyme that convert the vitamin D pro-hormone to its active form, inhibits the microbicidal vitamin D dependent-pathway [66]. Also, the same study showed that miR-146a was the second most differentially expressed miRNA in lepromatous leprosy skin biopsies [66], although in our hands, with a high number of samples we detected no difference between the clinical forms of the disease.
In summary, we demonstrated the genetic association between miR-146a C-allele with leprosy susceptibility. Our data also suggest that miR-146a was overexpressed in leprosy biopsies and also produced by mTHP-1 infected with live M. leprae. Subjects carrying the risk allele also express high levels of miR-146a which correlates with lowest levels of TNF as readout of the inflammatory responses.
|
10.1371/journal.pntd.0002459 | Southernmost Asia Is the Source of Japanese Encephalitis Virus (Genotype 1) Diversity from which the Viruses Disperse and Evolve throughout Asia | Although a previous study predicted that Japanese encephalitis virus (JEV) originated in the Malaysia/Indonesia region, the virus is known to circulate mainly on the Asian continent. However, there are no reported systematic studies that adequately define how JEV then dispersed throughout Asia.
In order to understand the mode of JEV dispersal throughout the entire Asian continent and the factors that determine the dispersal characteristics of JEV, a phylogenetic analysis using Bayesian Markov chain Monte Carlo simulations was conducted on all available JEV E gene sequences in GenBank, plus strains recently isolated in China. Here we demonstrate for the first time that JEV lineages can be divided into four endemic cycles, comprising southern Asia, eastern coastal Asia, western Asia, and central Asia. The isolation places of the viruses in each endemic cycle were geographically independent regardless of years, vectors, and hosts of isolation. Following further analysis, we propose that the southernmost region (Thailand, Vietnam, and Yunnan Province, China) was the source of JEV transmission to the Asian continent following its emergence. Three independent transmission routes from the south to north appear to define subsequent dispersal of JEV. Analysis of JEV population dynamics further supports these concepts.
These results and their interpretation provide new insights into our understanding of JEV evolution and dispersal and highlight its potential for introduction into non-endemic areas.
| Japanese encephalitis virus (JEV) probably originated in the Malaysia/Indonesia region. Currently, there are no systematic studies that adequately define how it subsequently dispersed throughout Asia. In this study, we demonstrate that JEV lineages can be divided into four endemic cycles, comprising southern Asia, eastern coastal Asia, western Asia, and central Asia. In each endemic cycle the source of virus was geographically independent regardless of year, vector, and host of isolation. The southernmost region (Thailand, Vietnam, and Yunnan Province, China) was identified as the most likely source of JEV transmission from its origin to the Asian continent. Based on the evidence, we identified three probable JEV dispersal routes from south to north. Analysis of JEV population dynamics further supports this view. Our results provide new insights into the understanding of JEV evolution and dispersal and highlight its potential for introduction into non-endemic areas.
| Japanese encephalitis (JE) is arguably one of the most serious viral encephalitic diseases worldwide [1], [2]. According to the latest report of the World Health Organization, JE is endemic in 24 Asian and Oceanian countries, with an estimated 67,900 JE cases annually (the total morbidity rate is 1.8/100,000 population). An estimated 3 billion people live in countries where JE is endemic. Additionally, with increased international travel to JE endemic areas, more people are at risk of JE infection. Therefore, JE is not only an endemic disease in Asian and Oceanian countries, it could also potentially cause significant public health issues in non-endemic countries or regions and has the realistic possibility of becoming a serious global public health problem [1]–[4].
JEV is the prototype member of the JEV serogroup within the genus Flavivirus, family Flaviviridae. The viral genome is a positive-sense, single-stranded RNA that is approximately 11 kb in size. The genome carries a single open reading frame (ORF) encoding a polyprotein that is processed into three structural proteins [capsid (C), membrane (M), and envelope (E)] and seven nonstructural proteins (NS1, NS2A, NS2B, NS3, NS4A, NS4B, and NS5) [5], [6]. Phylogenetic analysis of JEV has shown that based on the E gene or the complete genome, JEV can be divided into five genotypes (G1–G5) [7], [8], [9].
JEV is maintained in nature in a cycle involving vertebrate hosts (including pigs and waterbirds) and Culex mosquitoes. Culex tritaeniorhynchus is the primary vector [10]. Pigs are important reservoir hosts of JEV. Migrating birds are thought to be an important factor in the dispersion of JEV to new geographical areas [11]. During the virus transmission cycle, mosquitoes become infected with JEV when they feed on infected pigs and birds. They replicate the virus and subsequently feed again, in some cases transmitting the virus to humans or horses which are incidental hosts of JEV.
JEV was predicted to have originated from the tropical Indonesia/Malaysia region because there is evidence that this region had all genotypes of JEV circulating, whereas only the more recent genotypes circulate in other areas [12]. However, the virus is currently known to circulate throughout Asia. These observations raise several questions. Firstly, what is the pattern and direction of JEV dispersal from the Indonesia/Malaysia region to the entire Asian continent? Secondly, what are the primary factors that determine the dispersal characteristics of JEV? Thirdly, what is the contribution to virus dispersal of migratory birds, seasonal winds, mosquitoes and other factors, such as temperature and rainfall? Resolution of these intriguing questions will not only inform science but will also provide guidance for public health authorities in the development of prevention and control strategies for JE. Previous reports on transmission patterns showed that JEVs in East Asia were introduced from South East Asia [13], and JEVs circulating in Japan were introduced from South East Asia and continental East Asia [14]. However, these studies were limited to local areas. Therefore, in the present study, together with the new JEV strains isolated in China, we analyzed all available sequences of the currently predominant genotype (G1) of JEV isolates that are widely dispersed over the Asian continent.
The E protein gene of 22 JEV strains newly isolated in China from 2005 to 2010 was sequenced. Among these 22 strains, two were isolated from Yunnan Province in 2005 and 2006; four were isolated from Guangxi Province in 2006, Henan Province in 2006, Shanxi Province in 2006, and Jiangxi Province in 2009, respectively; two were isolated from Liaoning Province in 2006 and 2007; three were isolated from Shandong Province in 2008 and 2009; three were isolated from Chongqing Municipality in 2008 and 2009; and eight were isolated from Hubei Province during 2008–2010. The isolation protocols have been described elsewhere [8]. Briefly, the viruses were amplified once by infecting Aedes albopictus C6/36 mosquito cells. After development of cytopathic effects (CPE), culture supernatants were harvested, and viral RNA was extracted using the QIAamp Viral RNA Mini Kit (Qiagen, Hilden, Germany). The purified RNA was used as the template for cDNA synthesis using Ready-to-Go You-Prime First-Strand beads (Amersham Biosciences, Piscataway, NJ, USA). The complete E gene was amplified using the following primers: JEV-Ef [5′-TGYTGGTCGCTCCGGCTTA-3′ (955–973)] and JEV-Er [5′-AAGATGCCACTTCCACAYCTC-3′ (2516–2536)]. Amplified products were examined by agarose gel electrophoresis (1%), purified using a QIAquick Gel Extraction Kit (Qiagen) and then sequenced directly.
The envelope sequences of these 22 newly isolated JEVs determined in the present study were deposited in GenBank under Accession Numbers JQ937333 and JQ937336–JQ937356 (Table S1).
The most recent study showed that in evolutionary terms, the G1 JEV genotype is the youngest of five genotypes and is the dominant genotype circulating in Asia [15]. For studying the dispersing patterns of G1 JEV, totally 656 E sequences of JEV, with information regarding the isolation time and place, were downloaded from GenBank (as of May 1, 2012). ClustalX version 2.0.9 [16] was used to generate sequence alignments of 678 E gene sequences (including 22 newly contributed sequences). The dataset was screened for recombination using RDP3 (Recombination Detection Program3) and GARD (genetic algorithm for recombination detection) [17], [18]. No recombination events were identified (data not shown). Subsequently, in order to differentiate G1 from the other four genotypes, the Neighbor-joining method in Mega Version 5.05 [19] was applied for phylogenetic analysis. Finally, 359 E sequences of the G1 JEV genotype were obtained and comprised a sequence database for phylogenetic analysis (Table S1).
The sequence database constructed above was analyzed using Bayesian Markov chain Monte Carlo (MCMC) method. The GTR+I+G was selected as the optimal nucleotide substitution model by MrModelTest [20]. The nucleotide substitution rate and divergence time of the most recent common ancestor (TMRCA) were estimated using the relaxed (uncorrelated lognormal) molecular clock model under a coalescent model of constant population size in the BEAST software package [21]. Demographic histories were inferred by Bayesian skyline reconstruction. The analysis was run through 600,000,000 generations to ensure sufficient mixing. Convergence of parameters was checked using TRACER and was indicated as effective sample size (ESS>200), and the maximum clade credibility (MCC) tree was built using TreeAnnotator with 10% burn-in (http://beast.bio.ed.ac.uk/). Statistical uncertainty was expressed for nodal support by 95% confidence intervals of the highest posterior density (HPD).
In order to infer the history of geographical dispersion of JEV, the Bayesian stochastic search variable selection (BSSVS) was used to provide evidence for statistically supported diffusion between state variables under BEAST v1.7.5 [22]. This method estimates the most probable state at each node in the MCC trees, allowing us to reconstruct ancestral positions for ancestral viral lineages along the tree. For phylogeographic reconstructions, each region was coded as a discrete trait. BSSVS output and surfaces representing uncertainty for continuous diffusion processes were formatted as KML using the SPREAD utility [23]. Determination of each locality was coordinated and performed using Google Earth v.6.2.2. Mapinfo was finally used to display the dispersal pattern of JEV based on the phylogeographic analysis.
Based on the isolation sources of 359 G1 JEV strains, by 2010, the distribution of the G1 JEV genotype had shifted northward to a latitude of 45° (Japan) and westward to a longitude of 75° (India) and had covered almost all JEV endemic areas, including Australia, Thailand, Vietnam, Cambodia, India, Japan, Korea, Taiwan, and most regions of China (15 provinces). The Malaysian G1 JEV sequence was not included in the present study because only the PrM sequence was available. The 22 newly isolated JEVs contributed to knowledge regarding the geographical distribution of the G1 genotype in central Asia, a highly endemic area for JEV and added to strains isolated in China after 2005, but especially after 2009. A time span of 44 years was found for the isolation time of the 359 strains studied, from 1967, when the first G1 JEV genotype was isolated, to 2010. Additionally, the viral strains used were from various sources, including insect vectors (various mosquito species and midges) and host animals (pigs and human patients).
The maximum clade credibility (MCC) tree of the 359 E sequences of JEV is shown in Figure 1. The tree showed that the G1 JEV genotype isolates can be divided into seven clusters according to their geographic isolation sites (designated clusters 1–7). They were further grouped into four lineages, lineage I (clusters 1 and 2), II (clusters 3–5), III (cluster 6) and IV (cluster 7). According to the geographic locations of the principal JEV strains in each lineage, four lineages were also designated as the southern Asia endemic cycle, the eastern coastal Asia endemic cycle, the western Asia endemic cycle, and the central Asia endemic cycle, respectively. Virus strains in the southern Asia endemic cycle were mostly derived from Vietnam, Thailand, Cambodia, Australia, and Yunnan Province in China. Strains in the eastern coastal Asia endemic cycle were mainly derived from Shanghai, Zhejiang, Liaoning, Shandong, Taiwan, Japan, and Korea; strains from southernmost Yunnan Province and those isolated after 2008 in Chongqing, Hubei, and Jiangxi of central China (such as JX0939, SZ18, JL18, ES57, HBZG0907, and HBZG0809) are also included in this endemic cycle. Most of the strains in the western Asia endemic cycle were obtained from Tibet and India in western Asia, and other strains were isolated in southernmost regions such as Thailand and Vietnam or Japan. The dominant strains in the central Asia endemic cycle were from the inland provinces in China, including Guizhou, Sichuan, Chongqing, Henan, Hubei, Gansu, and Shanxi. This latter cycle also contains a few strains from the southernmost part of Asia such as Vietnam and Yunnan Province in China, and the eastern coastal regions of Asia such as Shandong, Taiwan, and Guangxi in China, Japan, and South Korea.
The most recent common ancestor (TMRCA) for the G1 JEV genotype is estimated to have diverged approximately 78 years ago based on the Bayesian MCMC approach using E gene sequences. The resultant G1 JEV genotype first appeared in southernmost Asian regions, such as Thailand, Vietnam, and Yunnan Province in China and established endemic cycles in those regions.
Detailed analysis of the geographical locations of strains in the four endemic cycles showed that at least one strain of JEV isolated from southernmost regions of Asia, including Thailand, Vietnam, and Yunnan Province in China was present in eastern coastal Asia, western Asia, and central Asia endemic cycles (Figure 1, Table 1). For example, isolates from Yunnan Province in China were included in the eastern coastal Asia endemic cycle, isolates from Thailand and Vietnam were included in the western Asia endemic cycle, and isolates from Vietnam and Yunnan Province in China were included in the central Asia endemic cycle. Moreover, based on the chronological order of evolution, the strains isolated from the southernmost regions mostly occurred earlier than others and rooted those in each endemic cycle. However, those strains from other regions of Asia were found in endemic cycles consistent with their places of isolation. Furthermore, homology analysis revealed that nucleotide homology of strains isolated from the southern Asia endemic region was about 94%. However, strains from endemic regions of the eastern Asian coast and central Asia shared more than 96% and 97% homology. Thus, all data above indicate that viruses isolated from southernmost regions of Asia maintained the diversity of virus populations.
In addition, strains isolated from Thailand, Vietnam, and Yunnan Province in China in the same endemic cycle, have a wide span of isolation time. For example, a virus strain isolated in 1979 in Thailand and strains isolated later in 2005 in Thailand were included in the southern Asia endemic cycle. Moreover, the eastern coastal Asia endemic cycle contained strains isolated in 1982 in Yunnan Province in China and others isolated later in 2005 in Yunnan Province; also the western Asia endemic cycle contained strains isolated in Thailand in 1992 and later in 2005; and the central Asia endemic cycle contained strains isolated in Vietnam in 2001 and others isolated in 2007. Thus, in addition to the diversity of virus populations, viruses isolated from the southernmost regions of Asia also maintained the stable genetic characteristics of JEV. This implies that the southernmost region of the Asian continent plays a key role in transmission of JEV from its origin to the Asian continent, providing a source for continental dispersal of JEV strains.
The estimated history of JEV dispersal in endemic regions is shown in detail in Figure 2. The maps in Figure 2 display dispersal characteristics over time. According to our reconstructions, G1 JEV was initially introduced to Thailand and Vietnam located in southernmost Asia during the 1970s after originating in Malaysia/Indonesia (Figure 2, 1970). It then dispersed to Japan and Shanghai, located in east coastal Asia around 1980 (Figure 2, 1978, 1981). Subsequently, the virus was introduced to Sichuan province located in Central China from east coastal Asia around 1990 (Figure 2, 1990). Simultaneously, it spread to India located in the West part of Asia. In 2000, two lineages were dispersed to east coastal Asia (Zhejiang, Liaoning, Taiwan, South Korea) and to the Chinese inland provinces (Henan, Hubei, Gansu and Guizhou), respectively (Figure 2, 2000). After 2000, in addition to continuing its dispersion in east coastal Asia and in Chinese inland provinces, a lineage from Yunnan was introduced to the eastern coastal areas (Figure 2, 2003, 2006). On the other hand, a lineage from Vietnam was introduced to inland China and also from Thailand to Yunnan (Figure 2, 2006). Around 2009, the G1 genotype appears to have dispersed from India to Xizang (Figure 2, 2010). This virus dispersal history supports our contention that the southernmost regions of the Asian continent acted as the source of JEV prior to its transmission throughout the Asian continent.
Furthermore, by combining the divergence time and geographical distribution characteristics of each cluster (Table 1) within the MCC tree, three different routes from southern to northern Asia were postulated for dispersal of G1 JEV (Figure 3) Through the eastern route, JEV dispersed from Thailand and Yunnan Province in China, to Japan, South Korea, and Shanghai, Zhejiang, and Liaoning in China generating the eastern coastal Asia endemic cycle. Through the western route, JEV dispersed from Thailand and Vietnam to the western regions of Asia, reaching India and Tibet, establishing the western Asia endemic cycle. Through the central route, JEV was transmitted from southern countries such as Vietnam, to central Asia, and reached inland provinces of China including Sichuan, Chongqing, Guizhou, Hubei, Henan, Gansu, and Shanxi. In addition, JEV was introduced to inland China provinces from bordering east coastal regions and established the largest central Asia endemic cycle.
A skyline plot of the G1 JEV genotype population dynamics is shown in Figure 4. There was minimal fluctuation during the first half of the plot. This was followed by a major population increase from 1980 to 1990, a relatively stable period from 1991 to 2003, a marked decrease during 2004–2007, and then a relatively stable period after 2008.
According to the time nodes of change in population dynamics identified in the skyline plot, the isolation times of JEV in the four endemic cycles were analyzed (Table 2). Viruses were all isolated from the southernmost regions of Asia (Thailand, Cambodia, and Yunnan Province in China) before 1990 during the major period of population increase. Subsequently, from 1991 to 2003, a period of population stability, G1 JEV was found in eastern coastal and southernmost regions of Asia, such as Shanghai and Liaoning in China, Japan, and South Korea. Although there was a virus population decline from 2004 to 2007, the dominant G1 JEV genotype continued its expansion to the central Asian areas. From 2008, the G1 JEV genotype dispersed to all endemic regions in the entire Asian continent and maintained a relatively stable population. JEV continued to be isolated in the southernmost regions of Asia during the entire fluctuation process of virus populations from the first occurrence. In conclusion, all the data are consistent with the concept that the southernmost regions of the Asian continent played a key role as the source for evolution and dispersal of the G1 JEV genotype.
Based on phylogenetic and phylogeographic analysis of the envelope gene of G1 JEV, the following conclusions were drawn from this study. 1) Southernmost Asia, particularly Thailand, Vietnam, and Yunnan Province in China, appear to represent the source for the continental dispersion of JEV which appears to have originated from the Indonesia/Malaysia region; 2) During the dispersal of G1 JEV, limited introductions were observed among difference geographic locations; and 3) Reverse dispersion from north to south occurred during the more recent years. The rationale behind the concept that southernmost Asia might be the source of JEV G1 can be explained as follows. Lying in the tropics and subtropics, these southern regions have a high annual average temperature and heavy annual rainfall both of which are particularly suitable conditions for high population density breeding of a wide range of mosquitoes [24]. C. tritaeniorhynchus, the predominant transmission vector for JEV, is widely distributed in Thailand and Vietnam with estimated coverage of 80.9% and 60%, respectively [10]. Additionally, following the remarkable increase in the acreage of irrigated rice in recent years, the distribution of C. tritaeniorhynchus has further expanded in Thailand and Vietnam which are traditional rice exporting countries [1]. Furthermore, swine husbandry in these regions has developed rapidly with the number of farmed pigs increasing by 100% from 1990 to 2005 [1]. In conclusion, the combination of a suitable climate year-round, wide distribution of the primary vector and abundant pig farming, provide the ideal prerequisite for the emergence, maintenance and reproduction of JEV in these southernmost regions of Asia.
In addition to these factors, examination of the recognized flight paths of migratory birds reveals a remarkable coincidence between the eastern, central, and western routes of JEV dispersal patterns and the recognized eastern, central, and western flight paths of migratory birds in Asia (Figure 3). In previous studies [11], [12], migratory birds were shown to be important hosts for introducing JEV to new territories. The black-crowned night heron (Nycticorax nycticorax), plumed egret (Egretta intermedia), and little Egret (Egretta garzetta) are the main birds that carry and spread JEV [11], [25]. These observations are consistent with the opinion that migratory birds carrying JEV migrate from south to north annually whereas non-migratory or resting birds are fed upon by mosquitoes in their local habitats. JEV is then further dispersed by local mosquitoes acquiring JEV from the infected birds and transmitting it to domestic pigs which then amplify the virus and provide a local source of infection when local mosquitoes feed on the infected pigs. After a period of evolution and dispersal, dominant JEV populations establish endemic cycles in mosquitoes, pigs and other hosts providing the opportunity for JEV epidemics in local areas. However, taking into account current expert opinion, global climate change could influence the migration patterns or routes of birds, resulting in JEV dispersal into new local areas or even new continents. It is therefore essential to extend JEV surveillance in migrating birds, mosquitoes and pigs in areas currently considered to be free of JEV, including Europe and the New World, where closely related viruses such as West Nile virus (Europe and the Americas) and Usutu virus (Europe) are already known to have established endemic cycles [25], [26]. It was also reported that mosquitoes can be carried very long distances on the wind, especially during the typhoon season [27], [28]. It is therefore conceivable that wind-blown mosquitoes may play a significant role in the dispersal of JEV.
Also, based on studies of the G1 JEV genotype, it is evident that although changes in JEV population dynamics throughout the Asian continent were recorded, a relatively stable state of JEV populations, and wide distribution throughout the Asian continent, was also observed. It therefore seems reasonable to propose that in the absence of adequate control strategies for JEV, the maintenance and widespread distribution of current dominant JEV populations provides a foundation for further expansion of JEV into traditionally non-endemic areas.
In the early and mid-20th century, JEV has caused many pandemics in Asian countries, such as Japan, China and India [2], [29], [30], resulting in huge health, social and economic burdens. With global warming and the increase in acreage of irrigated rice, more Asian regions provide habitats ideal for breeding of Culex tritaeniorhynchus, the primary vector of JEV [9]. Additionally, with the changes in pig farming practices and increasing international trade and personnel exchanges, JEV is being provided with numerous opportunities to disperse northwards and westwards to regions outside Asia [31]. For example, a range of Culex species is distributed in Europe, such as Culex pipiens pallens and Culex bitaeniorhynchus. These are recognized transmission vectors of JEV [32]. Moreover, pigs are raised in large areas of north-western European countries [33]. Therefore, if JEV is constantly being introduced to these regions, via infected migratory birds or transportation of infected mosquitoes, during the summer season, the possibility exists for JEV to become established and to form an endemic cycle among the local Culex mosquitoes and pigs. Additionally, whilst many viral encephalitic cases in Europe and Asia are known to have been caused by tick-borne encephalitis virus in regions adjacent to Europe [34], JE incidence is currently considered to be extremely low or non-existent in these regions [3]. However, JEV has been detected in specimens collected in the 1990s in the Wooded Steppe region of Northern Eurasia [35] and in avian tissues collected in Russia (GenBank Accession Number AF501313–15). Therefore, it is worth considering the possibility that at least in some cases of encephalitis in Russia and western and central Europe, the aetiological agent may be JEV rather than tick-borne encephalitis. Whilst such a proposal would possibly have been ridiculed a few years ago, several pathogenic mosquito-borne flaviviruses have emerged in Europe during the past decade and recent reports suggest the possibility of the presence of a G3 JEV genotype in birds in Italy [36], [37]. In conclusion, there is clearly a need for a coordinated system of surveillance throughout Europe and Asia in the hope of identifying potential threats of emergence of JEV in new territories.
It is worth noting that, although questions on the transmission pattern of G1 JEV were answered to a certain degree through this study, other questions still require answers. For example, what factors were primarily responsible for the dispersion from Malaysia/Indonesia to Thailand/Yunnan? Why is the JEV G1 genotype no longer detectable in Malaysia/Indonesia? What is the situation regarding the other JEV genotypes? Do they have similar transmission patterns? Further studies will be needed to answer these questions. Our study was based on the available strains. It remains to be seen if, when more data are available, subsequent studies support or revise our conclusions.
|
10.1371/journal.pcbi.1000116 | Similarity Queries for Temporal Toxicogenomic Expression Profiles | We present an approach for answering similarity queries about gene expression time series that is motivated by the task of characterizing the potential toxicity of various chemicals. Our approach involves two key aspects. First, our method employs a novel alignment algorithm based on time warping. Our time warping algorithm has several advantages over previous approaches. It allows the user to impose fairly strong biases on the form that the alignments can take, and it permits a type of local alignment in which the entirety of only one series has to be aligned. Second, our method employs a relaxed spline interpolation to predict expression responses for unmeasured time points, such that the spline does not necessarily exactly fit every observed point. We evaluate our approach using expression time series from the Edge toxicology database. Our experiments show the value of using spline representations for sparse time series. More significantly, they show that our time warping method provides more accurate alignments and classifications than previous standard alignment methods for time series.
| We are developing an approach to characterize chemicals and environmental conditions by comparing their effects on gene expression with those of well characterized treatments. We evaluate our approach in the context of the Edge (Environment, Drugs, and Gene Expression) database, which contains microarray observations collected from mouse liver tissue over the days following exposure to a variety of treatments. Our approach takes as input an unknown query series, consisting of several gene-expression measurements over time. It then picks out treatments from a database of known treatments that exhibit the most similar expression responses. This task is difficult because the data tends to be noisy, sparse in time, and measured at irregular intervals. We start by reconstructing the unobserved parts of the series using splines. We then align the given query to each database series so that the similarities in their expression responses are maximized. Our approach uses dynamic programming to find the best alignment of each pair of series. Unlike other methods, our approach allows alignments in which the end of one of the two series remains unaligned, if it appears that one series shows more of the expression response than the other. We finally return the best match(es) and alignment(s), in the hope that they will help with the query's eventual characterization and addition to the database.
| Characterizing and comparing temporal gene expression responses is an important computational task for answering a variety of questions in biological studies. We present an approach for answering similarity queries about gene expression time series that is motivated by the task of characterizing the potential toxicity of various chemicals. Our approach is designed to handle the plethora of problems that arise in comparing gene expression time series, including sparsity, high-dimensionality, noise in the measurements, and the local distortions that can occur in similar time series.
The task that we consider is motivated by the need for faster, more cost-efficient protocols for characterizing the potential toxicity of industrial chemicals. More than 80,000 chemicals are used commercially, and approximately 2,000 new ones are added each year. This number makes it impossible to properly assess the toxicity of each compound in a timely manner using conventional methods. However, the effects of toxic chemicals may often be predicted by how they influence global gene expression over time. By using microarrays, it is possible to measure the expression of thousands of genes simultaneously. It is likely that transcriptional profiles will soon become a standard component of toxicology assessment and government regulation of drugs and other chemicals.
One resource for toxicology-related gene expression information is the Edge (Environment, Drugs, and Gene Expression) database [1]. Edge contains expression profiles from mouse liver tissue following exposure to a variety of chemicals and physiological changes, which we refer to as treatments. Some of the treatments in Edge have been assayed as time series. Figure 1A provides a simplified illustration of the type of data with which we are concerned. The small database in this figure contains time series data for four different treatments, each of which includes measurements for three genes. The true, underlying expression response is not known, but instead the database contains sampled observations which may be noisy. We use the term observation to refer to the expression measurements made at a single time point in a treatment.
The computational task that we consider is illustrated in Figure 1B. Given an expression profile as a query, we want to identify the treatment in the database that has the expression profile most similar to the query. In the general case, the query and/or some of the database treatments are time series. In this case, we want to also determine the temporal correspondence between queries and putatively similar treatments in the database. In the toxicology domain, we are interested in answering this type of query in order to characterize poorly understood chemicals.
There are several properties of the expression time series at hand that are important considerations for our work.
These properties of the data result in several additional challenges for the task we consider.
To address these challenges, we have developed a generative model that approaches the problem from a probabilistic perspective. In order to temporally align gene-expression time series using our model, we employ a novel method for dynamic time warping. Dynamic time warping [3],[4] is an approach for aligning pairs of time series that was originally developed for speech recognition problems. It employs dynamic programming to find an optimal alignment with respect to a given scoring function. We also use spline interpolation as a preprocessing step to predict expression responses for unmeasured time points, in order to reconstruct a more complete time series.
Our time warping approach differs in several substantial ways from the standard dynamic programming method. Unlike the standard approach, our method does not force the two series to be globally aligned. Instead, it permits a type of local alignment in which the end of one series is unaligned. We refer to this case as shorting the alignment. This aspect of the approach is motivated by the consideration that one of the series may show more of the temporal response than the other. For example, one series may not have been measured for as long as the other. Another significant way in which our approach differs from standard time warping is that it is based on an explicit, generative model. This model allows the user to explicitly encode costs/probabilities that characterize the likelihood of various types of differences in closely related time series. The most significant way in which our approach differs from standard time warping is that it enables the user to impose fairly strong biases on the form that the alignments can take. In particular, it allows alignments that partition the given time series into a small number of segments in which the changes from one time series to the other (e.g., in terms of amplitude) are fairly uniform. This is important given the sparsity, high-dimensionality, and noisiness of the time series being aligned.
We also investigate variations on spline interpolation in order to find an approach that results in accurate reconstructions of sparsely sampled time series. We find that we achieve more accurate interpolations when using higher order splines. Further, our experiments indicate that it is helpful to relax the splines' fit to the observed data, rather than potentially overfitting by exactly intercepting each observed data point.
In earlier work, our group [5] and others [6] have developed systems for classifying chemicals according to the expression profiles they induce. The approach that we present here differs in that it takes into account the temporal aspects of expression profiles, and it is able to answer similarity queries. The latter property is important because some classes may be very sparsely populated in the database, and class labels may not be available or readily defined for some treatments.
Lamb et al. [7] consider the task of finding expression profiles that are similar to a given query profile, such as one induced by a particular drug. Their approach does not represent time series, however. Moreover, it assumes that the query includes a specified set of genes which are known to be correlated with some state of interest, such as the expression activity induced by the drug. Our approach does not require that such a gene set be provided.
Aach and Church [8] were the first to apply the method of dynamic time warping [3] to gene expression profiles, and other groups have followed [9],[10]. The method we present differs in several key respects. First, our method is able to not only align a pair of time series, but it is also able to pick out the known time series most similar to an unknown one for purposes of classification. Second, we use nonlinear spline models in conjunction with time warping in order to interpolate to unseen time points. Third, we consider local alignments of time series in which one of the series is shorted.
Bar-Joseph et al. [11] have investigated splines and warping in the context of clustering and aligning time series. Our work differs primarily in the task being considered and the use of a more expressive warping model. They restrict their attention to linear warping, whereas we use a “multisegment” model that warps different regions of the series by different amounts.
Listgarten et al. [12] have developed a method for multiple alignment of time series data that has some similarities to our approach. The task they consider—multiple alignment—is different than ours, and their method does not employ splines.
A related approach to aligning time series is proposed by Gaffney and Smyth [13]. They use an expectation-maximization method in concert with a mixture model in order to simultaneously align and cluster time series. Our work, however, is not concerned with clustering known time series. Rather our aim is to use a database of previously seen time series to answer similarity queries about a new one. Further the biases they allow are not appropriate to our task. They allow only linear scaling in time and measurement space whereas we need more complicated warpings, and they allow translation in both these dimensions as well which is unnecessary for us.
Another similar approach is correlation optimized warping (COW), devised by Nielsen et al. [14]. They compare time series by dividing them into several roughly equal segments and summing the Pearson's correlations of corresponding segments. The segments may vary in length by up to a slack factor provided by the user, and dynamic programming is used to find the segments with the maximum sum of correlations. Unlike our approach, their method assumes that the series will be globally aligned, without any shorting. Further, the use of correlation can be limiting as COW is unable to distinguish between two series that are proportional to one another.
Our approach is also related to various probabilistic sequence models, such as generalized hidden Markov models, that directly evaluate the likelihood of segments of a sequence, instead of incrementally computing these likelihoods one sequence element at a time. Models of this type have been used for tasks such as gene finding [15] and secondary structure prediction [16].
In this section we detail our generative model for classifying and aligning time series, and present a dynamic programming algorithm that is able to find optimal alignments under this model. We also present a review of B-spline interpolation and discuss some useful variations of the method. We use spline interpolation to reconstruct unobserved microarray observations.
Our approach to answering similarity queries involves three basic steps: (i) we use interpolation methods as a preprocessing step to reconstruct unobserved expression values from our sparse time series; (ii) we use our alignment method to find the highest scoring alignment of the query series to each treatment series in the database; (iii) we return the treatment from the database that is most similar to the query, and the calculated alignment between the two series.
We have implemented all our algorithms in Java. The source code is available for download at http://www.biostat.wisc.edu/aasmith/catcode/.
One challenge that arises when aligning a pair of expression time series is that the series may have been sampled at different time points. Moreover, the sampling may be sparse and occur at irregular intervals. To address these issues, we first use an interpolation method to reconstruct the unobserved parts of the time series before trying to align them. This interpolation step allows us to represent each time series by regularly spaced observations. We refer to the “observations” which come from the interpolation, as opposed to measurement, as pseudo-observations.
Although linear interpolation is a natural first approximation, other work has explored the use of B-splines to better reconstruct missing expression data [11]. A B-spline is a piecewise polynomial function that is a generalization of a Bézier curve. We present a brief review here, although for depth we refer the reader elsewhere [17].
As shown in Figure 2, a B-spline is the weighted sum of a set of basis splines. The basis splines are determined by the desired order k of the splines, and the points of discontinuity which are called knots. There are n bases, where:(1)and they are defined via the Cox-de Boor regression formulas:(2)(3)where bi,k is the ith base of order k.
It follows that the segments of the kth-order basis splines have degree of k−1, so a second-order B-spline consists of line segments, a third-order spline consists of quadratic segments, etc. The splines are also continuous down to the (k−2)th derivative. The actual interpolating B-spline s inherits these properties. It is formally defined as:(4)
The weights Ci are known as control points, and solving for them is a simple matter of solving linear equations. With n points (ti,xi) to interpolate:(5)
With fewer than n points, the problem is underconstrained and cannot be solved with such a large k. With more than n points, the problem is overconstrained and can only be solved in a least-squares sense. This is easy to do with standard linear algebra techniques. However, one must make sure that every base overlaps with at least one observation, or the matrix will be rank-deficient and the equations unsolvable.
Unfortunately, B-splines have a tendency to overfit curves in data-impoverished conditions. Such reconstructions can show large oscillations in an attempt to exactly intercept every observed data point. This can be especially problematic with microarray data, which are already inherently noisy. The solution we use is to solve for the control points of a low-order spline, and then use those control points for a higher-order one. Such a spline will tend to fall within the convex hull created by the lower-order spline [17]. We refer to such splines as smoothing splines, and refer to B-splines solved with conventional methods as intercepting splines.
Each possible alignment we consider for two given time series (the query and the database series) partitions the series into m segments, where the ith segments of the series correspond to one another. Our dynamic programming method tries to find a partitioning of the series that reveals the maximal similarity between them. As discussed earlier, we want to take into account that the nature of the relationship between the two series may vary in different segments. For example, it may be the case that the first part of the expression response occurs more slowly in one treatment than in a similar treatment. Recall also that the segments do not have to cover the entirety of both series—one of the series may be “shorted.”
Figure 3 illustrates the type of alignment we want to consider. This figure shows the optimal alignment between a query treatment and a given treatment in the database. (For simplicity, the figure shows each treatment as consisting of only a single gene.) This alignment involves three different segments, and in each segment the amplitude and stretching relationships between the two series are somewhat different. We use the term stretching to refer to distortions in the rate of some response, and the term amplitude to refer to distortions in the magnitude of the response. In addition, the alignment has shorted so that the full query is aligned with only a partial database series.
To determine the similarity between a query time series q and a particular database series d, we can calculate how likely it is that q is a somewhat distorted exemplar of the same process that resulted in d. In particular, we can think of a generative process that uses d to generate similar expression profiles. We can then ask how probable q looks under this generative process.
Given this generative process idea, we calculate the probability of a particular alignment of query q given a database series d as follows:(6)where m is the number of segments in the alignment, qi and di refer to the expression measurements for the ith query and database segments respectively, and si is the stretching value and ai is the amplitude value for the ith segment. The location of each segment pair is assumed to be given here.
Pm represents a probability distribution over the number of segments in an alignment, up to some maximum number M of allowed segments. Ps represents a probability distribution over possible stretching values for a pair of segments, Pa represents a probability distribution over possible amplitude values, and Pe represents a probability distribution over expression observations in the query series, given the database series and the stretching and amplitude parameters.
To represent Ps, we use a discretized version of the following distribution:(7)
We choose this distributional form because it is a variation of the log normal distribution that is symmetric around one, such that P(x) = P(1/x). Thus for example, stretching some expression response by a factor of two is equiprobable to compressing it by a factor of two. This symmetry property means that it does not matter which series we consider to be the query and which we consider to be from the database. As we discuss in the next section, our dynamic programming algorithm only allows segments to begin and end at a limited number of points. Thus, our distribution is actually discretized so that probability mass is allocated only to possible stretching values, and then renormalized.
We use a similar distribution to represent Pa, the distribution of amplitude values, since we also want to have P(x) = P(1/x) symmetry with these values. Thus a twofold increase in an expression response is treated as equiprobable to a twofold decrease.
To calculate Pe(qi|di,si,ai), we transform our representation of di using the given stretching and amplitude values, and then ask how probable qi appears when we use this transformed di series as a model. Let us first consider a simple case in which our time series have only one gene, and we are mapping only one point from the query segment qi to the database segment di. Let t represent a time coordinate in the segment qi, and let qil and qir denote the leftmost and rightmost time coordinates in the ith query segment. Let dil and dir denote the corresponding bounding time coordinates for the ith database segment. Then we can map a time coordinate from segment qi into the corresponding coordinate in di as follows:(8)where the stretching value si is defined by:(9)
Our model for “generating” points in the query series from a point in the database series is a Gaussian centered at the database point. Let p(x,μ,σe) represent the probability density function of this Gaussian, where μ is the mean and σe is the standard deviation of the Gaussian. We can then compute the probability of generating a query point qi(t) located at time t as:(10)
In other words, we center a Gaussian on the expression level at the mapped time coordinate in the database series, and ask how probable the scaled expression value from the query looks at that time coordinate.
To generalize this calculation to multiple observations in the query series, we make the simplifying assumption that the observations are independent, and we have:(11)where ni is the number of query observations in segment i.
Each of our observations represents measurements for hundreds of genes. We therefore generalize the description above by having p(x,μ,σe) be a multidimensional Gaussian, with one dimension for each gene measured. In our current work, we treat the genes as independent of one another given the time point. Thus the covariance matrix for this Gaussian is zero on all of the off-diagonal terms.
We assume that σe represents variation in expression measurements that are due to technical and biological variability. Thus, we estimate the standard deviation for each gene by considering the variance in a sample that consists of all the replicated experiments in the database.
In addition to considering the likelihood of the query series under the assumption that it exhibits a similar response to the given database series, we also consider its likelihood under a null model. The notion of a null model here is one that generates alignments by randomly picking observations from the database to align with the query sequence. The rationale for using such a null model is analogous to the use of a model of unrelated sequences in the derivation of substitution matrices for protein sequence alignment [18],[19]. In the case of protein sequence alignment, we want to know the relative likelihood of two cases: one case in which the correspondence between the sequences is explained by their relatedness through evolution, and the alternative in which the sequences are unrelated. In our task, we similarly want to compare the probability of an alignment given a model of relatedness (described above), and an alternative that asks how probable the query would look if we aligned it to an unrelated series.
The value of a null model for our application is that it enables alignments of differing lengths, including shorted alignments, to be compared on an equal footing. Under our scoring function which incorporates the null model, segments have a positive score only if the database series in that segment explains the corresponding segment from the query series better than the null model does.
Let p(x,μDB,σe) represent the probability density function of a multidimensional Gaussian whose mean μDB is the average expression level of the observations in the database, and whose standard deviation is σe as before.
We then estimate the probability of the ith segment of the query series under the null model as:(12)
Since our null model assumes that there is only a single segment with no amplitude change or stretching, we can compute the probability of the entire query series q as follows:(13)
Putting together the terms above, we can score a given alignment based on the log of the likelihood ratio of the query series under the “database series” model versus the query series under the null model as:(14)
Up to now we have described this process in terms of using a database series to generate the query series. However, we want our alignment method to be symmetric so that it does not matter which series we consider to be the query and which we consider to be from the database. Due to the last two terms, this will not necessarily be the case using the scoring function defined above. Therefore, we modify the scoring function so that it also considers using the query series to generate the database series:(15)
Here Pe(di|qi,1/si,1/ai) is calculated in an analogous manner to Pe(qi|di,si,ai) but the inverses of si and ai are used to generate observations in the database series.
Given a pair of time series, we do not know a priori which alignment (i.e., placement of corresponding segments) is optimal. However we can find the optimal alignment using dynamic programming. The following algorithm takes as input two time series, termed q and d, both of which are represented by regularly spaced observations (or interpolated pseudo-observations) of the gene expression values.
In particular, given a segment pair (qi,di), we can calculate its score as follows:(16)
The arguments to this scoring function define the leftmost and rightmost time coordinates of the segments being aligned from the query series and the database series. These points are selected from the set of regularly spaced observations mentioned above. The stretching parameter, si is defined by the relative lengths of the two segments. We find the amplitude coefficient ai via a least-squares method. Although this least-squares method is not guaranteed to find the optimal value of ai, we have found that, in practice, it provides solutions comparable to a dense grid search of the parameter, and it is much faster than the latter.
The core of the dynamic program involves filling in a three-dimensional matrix Г in which each element γ(i,x,y) represents the best score found with i segments that align the query subseries from time 0 to x with the database subseries from time 0 to y. As above, x and y must be selected from the given observations in the two series. The basic idea is that in order to determine γ(i,x,y), we look through all γ(i−1,a,b) where a<x and b<y. We then add the score of the segment from (a,b) to (x,y) to the value γ(i−1,a,b), assigning the best such sum to γ(i,x,y).
We define γ(i,x,y) with the following recurrence relation:(17)where the base case is:(18)
Here, q.r and d.r refer to the rightmost (last) time coordinates in the query series and the database series, respectively. The first condition in each recurrence relation ensures that the distribution over the number of segments Pm is taken into account when we consider the last pair of segments in a candidate alignment.
Recall that we are interested in possibly shorting the alignment, thus finding a local alignment rather than a global one. Allowed alignments are those that explain the entire extent of at least one of the two given time series. In order to recover the optimal alignment, we use a traceback procedure that involves scanning the elements of Г that represent alignments that include the entirety of the query series, the entirety of the database series, or both. The procedure returns the alignment corresponding to the highest-scoring entry among these. More formally, we find the score of the best alignment as follows, and start the traceback from the identified element:(19)
This dynamic program can be thought of as having three key “penalty terms” that determine the relative scores of alignments. These penalty terms correspond to the probability distributions that govern (i) the number of segments, (ii) the stretching values, and (iii) the amplitude values used in an alignment.
Preferences for the number of segments to be used in alignments are expressed by providing a distribution for Pm. In our work to date, we have assumed a uniform distribution up to the allowed number of segment pairs. It might be valuable to use a distribution that favors fewer segment pairs, however. Preferences for stretching and amplitude values are controlled via the standard deviation σ parameter in the distributions over these values. For example, as σa for the amplitude distribution is made smaller, a difference in amplitude between the series is penalized more in the scoring scheme.
In this section we present experiments that evaluate the utility of our novel time warping method and spline models for the task of answering similarity queries with expression profiles.
The data we use in our experiments comes from the Edge toxicology database [1], and can be downloaded from http://edge.oncology.wisc.edu/. Our data set consists of 216 unique observations of microarray data, each of which represents the expression values for 1,600 different genes. Each of these expression values is calculated by taking the average expression level from four treated animals, divided by the average level measured in four control animals. The data are then converted to a logarithmic scale, so that an expression of 0.0 corresponds to the average basal level observed in the control animals.
Each observation is associated with a treatment and a time point. The treatment refers to the chemical to which the animals were exposed and its dosage. The time point indicates the number of hours elapsed since exposure occurred. Times range from 6 hours up to 96 hours. The data used in our computational experiments span 11 different treatments, and for each treatment there are observations taken from at least three different time points.
We can assume that for all treatments there exists an implicit observation at time zero. This is the time at which the treatment was applied, so all expression values are assumed to be at base level. Therefore every query automatically includes at least two observations: the actual query time(s) and the zero point. Thus earlier points in time can be interpolated, even when there seems to be only a single query observation.
Figure 4 illustrates the evolution of four genes over time for 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD), for four different dosages.
Before we evaluate our generative alignment method, we wish to determine which type of spline (including simple linear interpolation) is the best to use in our preprocessing step. We do this by running a leave-one-out experiment in which we classify each observation in our data set in turn, using the remaining observations as the database. However, we exclude from the database any observation with the same treatment (i.e., chemical and dosage) and time as the query observation. We exclude from the queries observations from the last observed time of each treatment because we cannot interpolate pseudo-observations at these times when they are removed from the database series. We reconstruct hourly pseudo-observations for every treatment, using the different methods of interpolation. We search the reconstructed database for the pseudo-observation that is most like the query. We predict the query's treatment and time to be the same as this nearest neighbor. Notice that by excluding replicates of the query from the database, we are forcing our classifier to use interpolation in order to find the correct answer. We wish to know how accurately we are able to (i) identify the treatment from which each point was extracted, and (ii) align each query point to its actual time in the time series for the treatment. We refer to the former as treatment accuracy and the latter as alignment accuracy.
We note that this task is only a surrogate for the actual task with which we are concerned—classifying uncharacterized chemicals and aligning them with the most similar treatment in the database. It is a useful surrogate, however, because it is a task in which we know the most similar treatment and the correct alignment of the query to this treatment.
The metric we use to measure distance between the query observation and the database pseudo-observation being considered is a scale-independent Euclidean distance. The expression values of each database observation are all multiplied by a scalar, which is chosen via a least-squares method in order to minimize its distance to the query observation.
We consider seven different interpolation methods in all. We look at both intercepting and smoothing splines as explained in the Methods section, with orders three, four, and five. The control points for the smoothing splines are based on those for second-order interpolation. We also perform linear interpolation as a control. We use the observed times themselves as our knots (points of discontinuity). If there are too few observation times for a particular order, we use the highest possible order. (For example, if there is only a single observation, we interpolate linearly between it and the implicit zero point, regardless of the overall order used.) To allow for smoothing splines, we must keep the number of bases n constant. By Equation 1, the number of knots must decrease when the order k increases. We do this by resampling them down to the proper number.
There are several advantages to using the observed times as the knots for our interpolating splines. First, it allows easy comparison to the basic linear interpolation control. Second, we assume that the data was taken at those times because interesting behavior was anticipated. Using them as knots allows our splines more flexibility there. Third, it keeps the linear equations from being rank-deficient as explained earlier. With uniformly spaced knots (as used by Bar-Joseph et al. [11]) it is possible to be unable to solve for some control points.
The results of this experiment are shown in Figure 5. The top line shows classification accuracy, while the lower lines show alignment accuracy—where a case is considered “correct” if in addition to the proper treatment, the predicted time is correct to within 24 or 12 hours respectively. We test the significance of the differences in accuracy (from the linear interpolation control) using McNemar's χ2 test. Highlighted points are those deemed significant, with p<0.05. For all three accuracy measures we see improvement when using smoothing splines, while intercepting splines perform similarly or worse than the linear interpolation control. The fifth-order smoothing spline has a significantly higher classification accuracy (p≈0.025), and also appears to have better alignment accuracy (p≈0.132 for Δt≤24 and p≈0.180 for Δt≤12). By contrast the more traditional intercepting spline is likely overfitting its interpolation to the limited number of observed times. Although the fifth-order intercepting spline is not significantly different from the linear one for classification accuracy (p≈0.739) and alignment accuracy to within 24 hours (p≈0.705), there is a noticeable hit in the stricter alignment accuracy (p≈0.021). The p-values for the lower-ordered splines are qualitatively similar.
Based on these results, we restrict our attention to smoothing splines in subsequent experiments.
We now turn our attention to evaluating our multisegment time series alignment algorithm. For all of the experiments reported in this section, we set the parameters of this method as follows. We set the probability that the model has one, two, or three segments at each, and 0 beyond that. We estimate σe (the deviation of the expression Gaussian) to be the standard deviation of the known observations as described previously. We set both σs (the stretching deviation) and σa (the amplitude deviation) to be 10×(# genes)−1. Thus the three main components of the model have roughly similar influence.
We assemble queries by randomly subsampling time series in our data set. We assemble ten such queries from each treatment. We build each query by first selecting the number of observations in it, then choosing which time points will be represented, and finally picking an observation for each of these time points. The query sizes are chosen from a uniform distribution that ranges from one up to the number of observed times in the given treatment. The maximum size of a query is eight, although most consist of four or fewer observations. The time points are chosen uniformly as are the observations for each chosen time.
We then classify and align the query using all the other observations as the database. We preprocess both the query and the eleven database treatments using smoothing splines to reconstruct pseudo-observations at every four hours (starting at time zero, when all expression values are at the basal level). As before, we use the highest interpolation order possible in cases where there are too few observations for the prescribed one. We then align the query against all eleven treatments using our method. We return the database treatment with the highest scoring alignment, as defined by Equation 14. Because the alignment also maps each query time to a database treatment time, we can find the temporal error for any query time point. We thus calculate the average temporal error for the times in the original query in order to assess alignment error.
We consider several other alignment methods as baselines. We term the first baseline one-segment generative. This method is essentially the same as our multisegment generative alignment method, except that its alignments consist of only a single segment. It allows amplitude scaling and stretching, but only within its one segment pair.
The second control is traditional Euclidean dynamic time warping [3],[4]. Briefly, this method computes alignments by creating a matrix Г with elements defined recursively as(20)where D(di,qj) is the Euclidean distance between points di and qj in the two series and predecesssors(γ,(i,j)) refers to the matrix elements adjacent to γ(i,j) with both indices less than or equal to i and j respectively. The first element γ(0,0) is just the Euclidean distance at time 0, and each other element γ(i,j) is the score of warping d from times 0 to i and q from 0 to j. We then create a normalized score matrix Γ̅ where(21)
This makes it easy to compare warpings to different treatments, where one or the other dimension has been shorted.
Another control we consider is linear parametric warping. This is similar to the method explored by Bar-Joseph et al. [11], except that we make the assumption that the series are aligned at time zero. To find an alignment, we search possible slopes of the alignment line, and return the slope that results in the least average Euclidean distance between the query and the given database treatment.
Finally, we consider correlation optimized warping (COW) [14] as another baseline. This method takes as input two parameters: the number of warping segments m and a slack factor s. Both the query series and the database series are split into m segments. However while the segments of the query series are of equal length, the segment lengths of the database series may be up to s longer or shorter than an equal division would warrant. It is assumed that the starting and ending points of both series are aligned. The Pearson's correlation of each segment pair is calculated, and these are summed to score a given alignment. Dynamic programming is used to find the exact lengths of the database segments that maximize this value. We tried all values for m from one to ten together with all the values for s from zero to five. We report results for those (m = 10 and s = 5) that resulted in the highest accuracies.
The results of these experiments are shown in Figure 6A. For each method the top line represents classification accuracy with different orders of splines, the middle line represents alignment accuracy by adding the criterion that the average time error in the mapping is less than or equal to 24 hours, and the bottom line shows alignment accuracy where this tolerance is decreased to 12 hours. Points highlighted with a small square are significantly different from the corresponding point using our one-segment generative model (p≤0.05) according to McNemar's χ2 test. Likewise, the large square indicates a significant difference from the three-segment generative model.
The one-segment and three-segment models are only significantly different from each other in a handful of cases. Because we have added no distortion to the queries, the one-segment model should be sufficient to explain them. We might expect to see some degradation when using the three-segment model, as it is allowed much more freedom in where it places its segments. However, it seems that this is not the case; the three-segment model results in slightly higher accuracies. One explanation for this result is that the spline preprocessing does not create perfect reconstructions of the missing data, and the more expressive three-segment model is better at compensating for this error. Of the control methods, only COW is competitive with our generative method. There is no significant difference between its accuracy and that of our method. Euclidean dynamic time warping classifies fewer queries correctly than our method, although those it does tend to be aligned correctly. This is probably because it has a strong bias toward performing little warping.
To better test the utility of the multisegment model, we next consider distorting the query time series temporally. We use three different distortions. The first one doubles all times in the first 48 hours (i.e., it stretches the first part of the series), and then halves all times (plus an offset for the doubling) for the next 24 hours. The second distortion halves for the first 36 hours and then doubles for 60 hours. The third one triples for the first 60 hours and then thirds for another 20. It should be noted that not all the treatment observations extend this long in time. The short ones (e.g., those for which we only have measurements up to 24 or 48 hours) will thus not be distorted as much as the long ones.
Aside from the distortion, we perform the same experiment as before. We show the results in Figure 6B. In this experiment, the three-segment model results in more accurate classifications and alignments than the simpler one-segment model. Both DTW and the linear method appear brittle when confronted with distortions. Although our three-segment method significantly beats COW only when the strictest correctness criteria are used, the results shown are the best COW returned for a wide variety of parameters. We did not perform a similar parameter search for our own method.
One concern is that by adding distortion we could be changing the best classification of a given treatment. For example, maybe we would distort 10 µg/kg of TCDD in exactly the right way to make it look like 64 µg/kg. To address this concern, we have performed similar distortion experiments in which we align a distorted query series only to the database series that was used to generate it. The results of this experiment are qualitatively the same as those reported in Figure 6.
We conduct further experiments to evaluate the importance of the stretching and amplitude components of our model. First, we conduct an experiment in which we effectively remove the amplitude component of our model by fixing the value of ai to 1.0 for all segments. With all of the probability mass on this single value, the log Pa(ai) term in Equation 14 becomes zero. In a separate experiment, we set σa = ∞, which makes all amplitude changes equally likely. Similarly, we perform experiments in which we force si to 1.0 and set σs = ∞. The results of these experiments are shown in Figure 7.
Totally disallowing either stretching or amplitude changes has an overall deleterious effect on the accuracy of the alignments. However there seems to be little negative effect in allowing stretching and amplitude changes but not penalizing for greater values. These results imply that the stretching and amplitude components of the model are valuable, but that the accuracy of the alignments is relatively insensitive to the actual penalties selected.
We next consider a set of experiments in which we assess the accuracy of computed alignments as a function of the amount of data in the query. We restrict our experiments to a single treatment (41 observations of 1 µg/kg TCDD at eight time points), although other treatments yielded qualitatively similar results. We randomly pick out n observations from different times in the treatment to form each query. We use all the remaining observations in the treatment as the database. We interpolate both query and database series as before (every four hours), compute the best alignment using the one-segment and three-segment methods, and then assess alignment error. We do this 100 times for each value of n, which we vary from one to eight. We also vary the spline order from two to five, and repeat the experiment with the query times distorted (as in the last section) and not distorted. We perform paired, two-tailed t-tests on the alignment errors from the two methods in order to determine significant differences.
We expect the alignment error to generally decrease as we increase the query size. We also expect the one-segment method to perform slightly better when there is no distortion, and the three-segment method to be preferable when there is. However this latter behavior could be confounded for small query sizes, where the three-segment model may not have enough data to determine the segment parameters.
The results when we interpolate with third-order splines are shown in Figure 8. (The other orders of spline yield substantially similar results.) For queries of size two or less, the one-segment model performs slightly better. Its average error is less than that of the three-segment model, by less than one hour. However as the query size grows larger, the expected results become more apparent. When there is no distortion, the one-segment model is adequate. When there is distortion, a multisegment model is clearly preferable.
We next consider the sensitivity of the accuracy of the multisegment method to the number of segments it is allowed to use in its alignments. We would like to know to what extent the alignment accuracy degrades as the method is allowed to use more segments than the optimal alignment requires. We conduct an experiment in which we vary the number of segments from one to five, with query sizes of only one, four, and eight. The results of this experiment are shown in Figure 9 for the third-order spline case. Here each line represents one of the query sizes, from one at the top to eight at the bottom. A highlighted part of a line shows a significant change in alignment accuracy when going from an m-segment model to an (m+1)-segment model.
Again, we see that in the data-rich situation, the best models are those that closely approximate the number of segments needed to simulate the temporal distortion (or lack thereof) applied to the query. In data-poor situations, the alignments of the one-segment method are as accurate as multisegment alignments. Significantly, the accuracy of the multisegment method is quite robust when it is allowed to use more segments than necessary. This is important, as in practice we will not generally know the correct number of segments in order to find the best alignment of a query and its best matching series in the database.
Finally, we consider calculating the alignments for four treatments that we know are closely related. Figure 10 illustrates the alignments computed by our method for a 10 µg/kg dose of TCDD to itself and three other dosages of the same chemical.
These alignments illustrate several interesting phenomena. First, they indicate that the overall amplitude of the response increases along with the dose. This effect is illustrated by the boxed numbers on the segments in Figure 10. Second, the 10 µg/kg and 64 µg/kg dosages induce similar responses, both in their amplitude and temporal evolution. Third, the alignment to the 100 µg/kg dosage suggests that the response induced by this treatment initially progresses more slowly than the responses caused by the lower doses. This somewhat surprising result and the abovementioned effects are consistent with the expression profiles for the highly expressed genes shown in Figure 4.
We have presented an approach for answering similarity queries among gene expression time series, and aligning those queries in time. Our approach employs spline models to interpolate sparse time series, and a novel method for time warping. We have investigated our approach in the context of a toxicogenomics application in which we would like to know which treatments in a database of well characterized chemicals are most similar to a given query treatment.
The work we have presented features several novel aspects and contributions.
There are several avenues of future work we plan to pursue. One is to address the time complexity of our multisegment algorithm, which is O(n5), where n is the length of the series. Alignment to all eleven database series and subsequent classification currently take about a half hour to execute. By contrast, the time complexity of ordinary dynamic time warping is only O(n2). When the calculations are restricted to the so-called Sakoe-Chiba band, a narrow band centered on the diagonal of the warping matrix, the time complexity approaches O(n) [20]. We would like to devise heuristics to speed up our multisegment method. For example, although shorting complicates the use of a Sakoe-Chiba band, it might be possible to restrict calculations in the warping space to some other shape, such as a cone. Alternatively, we could perform a first pass with the faster one-segment model, and then restrict the multisegment model to an area near it in warping space.
In addition, we have made two independence assumptions that we plan to revisit in future research. First, we have assumed that each gene is independent of all the others given the model. We expect that representing some gene dependencies would lead to more accurate classifications and alignments. Second, we assume that the measurements at each time point are independent of each other time point. We plan to investigate a Markov-model like approach that represents dependencies between neighboring time points. |
10.1371/journal.pgen.1000908 | Admixture Mapping Scans Identify a Locus Affecting Retinal Vascular Caliber in Hypertensive African Americans: the Atherosclerosis Risk in Communities (ARIC) Study | Retinal vascular caliber provides information about the structure and health of the microvascular system and is associated with cardiovascular and cerebrovascular diseases. Compared to European Americans, African Americans tend to have wider retinal arteriolar and venular caliber, even after controlling for cardiovascular risk factors. This has suggested the hypothesis that differences in genetic background may contribute to racial/ethnic differences in retinal vascular caliber. Using 1,365 ancestry-informative SNPs, we estimated the percentage of African ancestry (PAA) and conducted genome-wide admixture mapping scans in 1,737 African Americans from the Atherosclerosis Risk in Communities (ARIC) study. Central retinal artery equivalent (CRAE) and central retinal vein equivalent (CRVE) representing summary measures of retinal arteriolar and venular caliber, respectively, were measured from retinal photographs. PAA was significantly correlated with CRVE (ρ = 0.071, P = 0.003), but not CRAE (ρ = 0.032, P = 0.182). Using admixture mapping, we did not detect significant admixture association with either CRAE (genome-wide score = −0.73) or CRVE (genome-wide score = −0.69). An a priori subgroup analysis among hypertensive individuals detected a genome-wide significant association of CRVE with greater African ancestry at chromosome 6p21.1 (genome-wide score = 2.31, locus-specific LOD = 5.47). Each additional copy of an African ancestral allele at the 6p21.1 peak was associated with an average increase in CRVE of 6.14 µm in the hypertensives, but had no significant effects in the non-hypertensives (P for heterogeneity <0.001). Further mapping in the 6p21.1 region may uncover novel genetic variants affecting retinal vascular caliber and further insights into the interaction between genetic effects of the microvascular system and hypertension.
| Retinal vessels provide a window to microvascular systems elsewhere in the body. The diameter of retinal vessels varies between racial/ethnic groups, being generally wider in African Americans compared to European Americans. To determine whether genetic background may contribute to this observed difference, we scanned the entire genomes of 1,737 African Americans, searching for genomic regions where individuals with either wider retinal venular or narrower retinal arteriolar caliber have a difference from the average percentage of African ancestry. We find that the percentage of African ancestry is positively correlated with retinal venular caliber, particularly in the hypertensive individuals. We detect substantive evidence of association between excess African ancestry and wider retinal venular caliber on chromosome 6p21.1 in the hypertensives, but not in the non-hypertensives. The 6p21.1 region contains genes that are known to be involved in development and modulation and of retinal vessels. Our results suggest that genetic factors may contribute to the observed difference in retinal vascular caliber between African Americans and European Americans. Further fine-mapping studies of the genomic region may identify variants affecting retinal vascular caliber.
| Changes in retinal vascular caliber provide unique information regarding the structure and state of the microvasculature in the eye, possibly reflecting pathophysiological processes in the microvascular systems elsewhere in the body. Narrowed retinal arteriolar caliber has been known to be predictive of hypertension [1],[2] and coronary heart disease [3], while wider retinal venular caliber is associated with higher blood pressure [4],[5], impaired fasting glucose, diabetes, dyslipidemia [6]–[8], and risk of coronary heart disease [9]. In particular, because retinal vessels share embryological, anatomical and physiologic characteristics with cerebral vessels [10], wider retinal venular caliber bas been closely linked to both subclinical and clinical cerebrovascular diseases, including lacunar infarction, white matter lesions, clinical stroke [9],[11],[12] and cerebral hypoxia [13].
Retinal vascular caliber has been observed to vary between racial/ethnic groups. In the Multi-Ethnic Study of Atherosclerosis, African Americans and Hispanics had wider retinal arteriolar and venular caliber compared to Whites and Asian Americans, even after controlling for cardiovascular risk factors [8]. In the Atherosclerosis Risk in Communities (ARIC) Study and Cardiovascular Health Study, African Americans had larger retinal arteriolar and venular calibers than European Americans while controlling for age, gender and mean arterial blood pressure [Wong TY, unpublished data, 2009]. The underlying reasons for this racial/ethnic difference are unclear and might be related to systemic, environmental, and measurement factors [14],[15]. However, several lines of evidence provide support for genetic factors also being involved in the regulation of retinal vascular caliber. The heritability of retinal arteriolar and venular caliber was estimated to be 0.48 and 0.54, respectively, in the Beaver Dam Eye Study [16]. Results from two twin studies also showed retinal vascular caliber may be primarily determined by genetic influence with the heritability of 0.57–0.70 for arteriolar caliber and 0.62–0.83 for venular caliber [17],[18]. The observed racial/ethnic differences in retinal vascular caliber could not be fully explained by systemic and environmental factors alone, which prompted our hypothesis that differences in genetic background may partially account for differences in retinal vascular caliber across racial/ethnic populations.
To identify chromosomal regions which may harbor genes that modulate retinal vascular caliber, we utilized admixture mapping, a technique that scans the genomes in recently admixed populations, such as African Americans, for regions which may contain variants that not only differ in frequencies between the two genetically diverse ancestral populations (Europeans and West Africans in this case), but can also partially explain differences in phenotypes between populations [19]–[24]. Since the identification of ancestry-informative markers and the development of appropriate statistical methods for admixture analysis in African Americans [23], admixture mapping and subsequent fine-mapping studies have has been successful in identifying determinant genetic variants for prostate cancer [25],[26], end stage renal disease (ESRD) [27], white blood cell count [28],[29], and the circulating levels of interleukin 6 and interleukin 6 soluble receptor [30]. Differences in retinal vascular caliber between African Americans and European Americans make this an ideal phenotype to study with the admixture mapping approach.
The main hypothesis of the present study was that some alleles affecting retinal vascular caliber are present at higher frequency in Africans than Europeans. We thus conducted a genome-wide admixture mapping scan for retinal arteriolar and venular caliber using approximately 1,365 ancestry informative markers in self-identified blacks from the ARIC study. In addition, hypertension is known as one of the most important risk factors for cardiovascular and cerebrovascular diseases. Previous studies indicated the importance of interactive effects between genes and vascular risk factors, in particular hypertension, on the occurrence of cardiovascular diseases [31]–[33] and cerebrovascular disorders [34]–[36]. It has been suggested that studies investigating the role of genetic components in vascular diseases would be more effective by analyzing interaction of genetic effects with conventional risk factors [35]. Therefore, a secondary purpose of our study was to test the hypothesis that the effect of genetic variants that conferred differences in retinal vascular caliber in African Americans was modified by hypertension.
Central retinal artery equivalent (CRAE) and central retinal vein equivalent (CRVE) were measured from retinal photographs of study subjects in the ARIC study to represent the average retinal arteriolar and venular caliber, respectively (see Methods). Table 1 shows demographic and phenotypic characteristics of the 1,737 African Americans included in the present study by hypertensive status. Of them, 1,001 (57.9%) had hypertension. The estimated percentage of African ancestry (PAA) in the African-American subjects was 82.4±9.8%.
The correlations between retinal vascular caliber and PAA are shown on Figure 1. CRVE was significantly correlated with PAA (correlation coefficient, ρ = 0.071, P = 0.003), while the correlation between PAA and CRAE was weaker and not statistically significant (ρ = 0.032, P = 0.182). To assess the potential interaction between genetic effects and hypertension, we examined the correlations limited to hypertensive individuals as an a priori subgroup analysis. We found that among hypertensive individuals, the correlations with PAA became stronger for both CRAE (ρ = 0.084, P = 0.008) and CRVE (ρ = 0.094, P = 0.003).
We carried out genome-wide admixture mapping scans for both CRAE and CRVE using case-only and case-control approaches, using up to 1,365 ancestry-informative SNP markers. Cases (N = 261) and controls (N = 260) for CRAE were defined as the extreme 15% of samples with narrowest caliber and the 15% with widest caliber, respectively, after adjustment for age, sex, study site, 6-year mean blood pressure, and fasting glucose level. For CRVE, the extreme 15% of samples with widest caliber was defined as cases (N = 260), while the extreme 15% with narrowest caliber as controls (N = 261). The mean CRAE was 46.2 µm (±15.3) narrower in cases compared to controls. On the other hand, the mean CRVE was 51.0 µm (±16.1) wider in cases compared to controls. Using 18 pre-specified European ancestry risk models, we did not detect significant admixture association with either CRAE or CRVE (Table S1). The genome-wide score, derived by averaging the evidence of association across all loci examined in the genome, was −0.73 for CRAE and −0.69 for CRVE (Table 2), which did not meet the thresholds of >2 for genome-wide significance [37].
To examine whether hypertension modified the effects of genetic ancestry on retinal vascular caliber, we performed admixture mapping scans by hypertension status. Cases and controls were defined as the extreme 15% (after adjustment for the covariates as described above) in the hypertensive subset and included about 150 subjects in each group. On average, CRAE was 46.5 µm (±15.5) narrower and CRVE was 51.6 µm (±16.7) wider in the cases, compared to the controls. We found genome-wide significant evidence of associations with CRVE in the hypertensive subset (Table 2). The genome-wide score in the case-only analysis was 2.31, which meets our threshold of >2 for significance. The strongest admixture association for CRVE was observed at 6p21.1 (42.5 Mb on chromosome 6 in build 35 of the human genome reference sequence; Figure 2), with the peak locus-specific LOD of 5.47, again reaching our priori defined thresholds of >5 for significance [37]. To further correct for the multiple hypothesis testing in two subgroups, we divided our test statistic, which is the likelihood ratio that compares the model under a risk model to the null model by 2. The likelihood ratio was 105.47 = 295120.9. The likelihood ratio after corrected for two hypothesis testing was 295120.9/2 = 147560.5, corresponding to a LOD score of 5.17, which still exceeds the threshold of >5 for significance. No other locus exceeded a LOD score of 5.
To evaluate the stability of our results, we carried out a longer analysis with 10-times more iterations in our Markov Chain Monte Carlo run. We obtained a similar strength of signal with a genome-wide score of 2.28 and a peak locus-specific LOD of 5.43 at the same location. The risk model with the strongest score corresponded to a risk of 0.5 due to one copy of an European ancestral allele with the inverse risk for carrying zero copies (see Methods for the set of risk models). Further refining the risk models, we obtain a genome-wide score of 3.66 and a locus-specific LOD of 6.85 in this region.
The admixture-generated signal for CRVE in the hypertensive subset was further supported by the case-control analysis. At the 6p21.1 peak, the cases had a highly statistically significant increase in African ancestry compared to controls (case-control Z score = −5.26, P = 1.44×10−7; Figure 2). The association was nominally genome-wide significant (P = 2.88×10−4) after conservatively correcting for multiple hypothesis testing (by multiplying by 2,000 because we tested 1,000 independent chromosomal chunks in two subgroups).
We did not find any significant associations with CRAE in the hypertensive subset (genome-wide score = 0.18). The highest locus-specific LOD was 2.37, arising from chromosome 5, followed by the second highest LOD of 2.04 on chromosome 6 (Figure S1). Admixture scans were also carried out in the non-hypertensive subset, but we did not observe any evidence of association with either CRAE (genome-wide score = −0.40) or CRVE (genome-wide score = −0.21; Table 2).
To further assess the robustness of the admixture-generated signal, we extracted the local estimate of African ancestry at the 6p21.1peak, and tested for association with continuous CRVE by hypertension status. This enabled us to increase power by including all samples in a quantitative analysis. We carried out a series of linear regression analysis, with the normal-quantile transformed CRVE (after adjusted for covariates as described above) as a dependent variable and the local ancestry as an independent variable. As shown in Table 3, local African ancestry alone was strongly associated with the transformed CRVE in the hypertensive subset (P = 2.9×10−8; Model 1). To assess whether continuous CRVE was associated with local ancestry at 6p21.1 due to their associations with global ancestry (i.e., PAA), we modeled the transformed CRVE as a function of local, global and regional ancestry. We found that the residual association of local ancestry with the transformed CRVE after adjustment for both global and regional ancestry remained significant (P = 3.9×10−6), indicating that there may be a gene in the region of 6p21.1 that is associated with CRVE above and beyond the fact that variants in this locus are highly differentiated between ancestral populations and thus correlated with global ancestry. The association was nominally genome-wide significant (P = 7.8×10−3) after correcting for multiple hypotheses tested (by multiplying by 2,000). The results were similar when CRVE was additionally adjusted for other covariates (Model 2), including high-density lipoprotein (HDL) cholesterol, low-density lipoprotein (LDL) cholesterol and plasma triglyceride levels, body mass index (BMI), smoking, and alcohol consumption. We estimated that each additional copy of an African ancestral allele at the 6p21.1 peak was associated with a CRVE increase of 0.37 Z-score units on average (equivalent to ∼6.14 µm).
In contrast, in the non-hypertensive subset, the local ancestry effect at the 6p21.1 peak on CRVE was weak and did not reach significance (Table 3). The effect of local ancestry showed significant heterogeneity (P<0.001) between the hypertensive and non-hypertensive groups, which was in line with the results of the above dichotomous admixture scans.
Given the significant statistical evidence for peak at 6p21.1, we constructed 95% credible interval (CI) for the loci identified. The 95% CI spanned from 40.8 to 43.9 Mb on build 35 of the human genome reference sequence (Figure S2). The locus-specific LOD score and the association of the CRVE to local ancestry for the SNPs located near and within the 95% CI are presented in Table S2.
We used admixture mapping to search for genomic regions that may account for inter-individual variations in retinal vascular caliber. We found evidence for association with retinal venular caliber at 6p21.1 in hypertensive African Americans and observed concordant results when venular caliber was examined as a continuous variable, with higher levels of African ancestry being significantly associated with wider retinal venular caliber. The significant evidence of association with the local ancestry at the 6p21.1 peak was above and beyond the contribution of both global and regional ancestry. Methodologically, these results are interesting in that subset analysis was required in order to detect the association. We note that subsets analysis has previously been very successful in admixture mapping. For prostate cancer, the chromosome 8q24 locus was not detected until the analysis was limited to a subset of individuals with a younger age at diagnosis [25]. For ESRD, the admixture signal to be much stronger among non-diabetic ESRD (mainly hypertensive ESRD) only, compared to diabetic ESRD [27]. Subsequent fine mapping identified genetic variants strongly associated with both prostate cancer [26] and non-diabetic ESRD [27]. We hope to follow up the present analysis with successful fine-mapping as well.
Our findings in persons with hypertension in the subgroup analysis imply that genes associated with hypertension may have exerted their effects on retinal vascular caliber. While we cannot exclude the possibility of a chance finding, we were able to show consistent results in the local ancestry analysis, which included the total population. Furthermore, from a physiological perspective, our finding that there is a genetic association with retinal venular caliber specifically in people with hypertension is sensible. Hypertension is known to have profound effects on the retinal microcirculation [38]–[40], and may induce gene expression relevant to the modulation of retinal vessels (see further discussion below) [41]. Our findings are also in line with previous studies in other vascular diseases that indicated hypertension exaggerates the effects from genetic factors [31]–[36].
In hypertensive persons, we detected significant genetic association for retinal venular caliber, but not arteriolar caliber. Because retinal arterioles and venules likely possess distinct genetic determinants [42], there may be no common genetic variants with a strong effect accounting for differences in retinal arteriolar caliber between European and African populations. Moreover, a diminished capability of retinal arterioles to remodel because of progressive sclerosis and rigidity of arteriolar vessel walls with age [43], may have precluded a degree of change in arteriolar caliber equal to that observed in venular caliber.
To our knowledge, there have been only two prior studies examining the genetic basis of retinal vascular caliber with a genome-wide approach, and both of them used linkage analysis [18],[45]. By genotyping 385 microsatellite markers in the Beaver Dam Eye Study, Xing et al. [45] found several loci for CRAE and CRVE, with the most significant loci at 3q28 (empirical P = 1.2×10−4) and 8q21 (empirical P = 2.9×10−3), respectively. A recent linkage analysis in the Australian Twins Eye Study identified 8p23.1 (LOD = 2.24) and 2p14 (LOD = 2.69) as suggestive loci for CRAE and CRVE, respectively [18]. Although findings of the two linkage analyses were not replicated in our study, possibly due to differences in study design and populations, all three studies provide lines of evidence that structural changes in the microvasculature of retina may have genetic determinants.
One major concern of the present study is a potential measurement error on retinal vascular caliber itself, which has been suggested to account for some of the observed racial/ethnic difference in retinal caliber [15]. A recent report suggested that retinal pigmentation could be a source of error in the computer-assisted methods to measure retinal vascular caliber from photographs [15]. The study reported that arteriolar and venular calibers were significantly wider in East Asian children than their white counterparts. However, when the analysis was confined to children with dark brown iris (a surrogate of retinal pigmentation) only, the differences between racial groups were less pronounced. Nevertheless, measurement errors are less likely to bias results in the present study for the following reasons. First, in the ARIC study, the retinal vessel edge was not detected based on computer-generated pixel density curve, but located manually by graders [46]. We believe the manual grading of retinal vascular caliber would be less prone to bias than computerized grading schemes, yet this remains to be proved. Second, genetic ancestry was shown to be significantly correlated with human pigmentation [47]. Although we did not measure skin or retinal pigmentation in the ARIC study, the inclusion of global ancestry as a covariate in our local ancestry analysis (Table 3) may thus provide an alternative way to adjust for the differences in retinal pigmentation.
The 95% CI for the 6p21.1 locus, a ∼3.1 Mb region, contains genes that may have biological relevance to the development and modulation and of retinal vessels (Figure S2). One such gene is the vascular endothelial growth factor (VEGF) gene. VEGF is an endothelial-cell selective mitogen intimately associated with vasculogenesis, angiogenesis and vascular permeability. In the retina, VEGF plays a crucial role in the induction of retinal vasculature and its expression is regulated by hypoxia during embryonic development [48]. Moreover, animal experiments showed that VEGF acts as a survival factor for newly formed retinal vessels [49], and it continues to be produced by retinal astrocytes and pericytes in the vicinity of retinal vessels in adults [50]. Interestingly, mechanical stretch on retinal vessel endothelium induced by systemic hypertension could increase the expression of VEGF and its receptor [41]. Further mapping work is needed to determine whether variants in the VEGF gene indeed contribute to variations in retinal vascular caliber. If proven, this may help explain why we detected the association at 6p21.1 only in hypertensive persons.
In addition to VEGF, there are two other genes in the 95% CI that may potentially be associated with retinal phenotype: peripherin 2 (PRPH2/RDS) and guanylate cyclase activator 1A (GUCA1A), both of which are also expressed in retina [51],[52]. GUCA1A plays a role in the recovery of retinal photoreceptors from photobleaching [52]. PRPH2/RDS is mainly located in the outer segment of both rod and cone, and defects in this gene are associated with retinal degenerations [51]. VEGF appears to be the strongest candidate gene based on its known function. However, it does not exclude the possibility that the admixture-generated signal is due to other genes.
The human leukocyte antigen (HLA) loci, located on chromosome 6p21.3 and about 10 Mb away from the 95% CI, are gene-rich and highly polymorphic [53]. The HLA region has been shown to play an important role in multiple disease susceptibility, particularly in autoimmune and infectious diseases [54]. HLA alleles are strongly associated with many neighboring SNPs, sometimes located at a considerable distance from the HLA allele with the linkage disequilibrium (LD) extending several Mb [55]. It remains to be determined whether the HLA alleles and the alleles in the 95% CI are in LD and may thus be associated with retinal venular caliber.
In summary, using a genome-wide admixture mapping scan in 1,737 African Americans, we detected a risk locus influencing retinal vascular caliber in hypertensive individuals at 6p21.1, where the association between local ancestry and retinal venular caliber was strong, suggesting the presence of a genetic effect beyond the effects of global ancestry. Follow-up fine mapping or haplotype tagging across the peak will be necessary to determine whether this region harbors genetic variations that may interact with hypertension to modulate retinal venular caliber. Understanding the genetic basis of retinal vascular caliber may provide novel insight into the development and remodeling of the microvasculature in the brain and elsewhere in the body.
This study was conducted according to the principles expressed in the Declaration of Helsinki. All sample collections were carried out according to institutionally approved protocols for study of human subjects and written informed consent was obtained from all subjects.
Subjects of the present study were from the 2,997 African-American participants of the ARIC study at the third examination. The ARIC study is a prospective epidemiologic study that examines clinical and subclinical atherosclerotic disease in a cohort of 15,792 persons (including 4,266 African-Americans), aged 45 to 64 years at their baseline examination. Participants were selected by probability sampling from four U.S. communities: Forsyth County, NC (12% African American); Jackson, MS (100% African American); the northwest suburbs of Minneapolis, MN (<1% African American); and Washington County, MD (<1% African American). The sampling procedure and methods used in ARIC have been described in detail elsewhere [56]. Participants self-reported their ethnicity. The baseline examination (visit 1) took place from 1987 to 1989, a second examination (visit 2) from 1990 to 1992, and a third examination (visit 3) from 1993 to 1995. Retinal vascular calibers were measured at visit 3. Data from visit 3 were used for the present analysis.
The final sample for the present analysis included 1,737 African Americans after excluding the following samples (N = 1,267): 1) African-American subjects who lived in Minneapolis, MN, or Washington County, MD, or 2) did not consent to genetic studies or did not have DNA samples available, 3) samples that were not genotyped successfully or that failed to pass quality control (see “Elimination of problematic samples”), 4) subjects who had no retinal photographs, ungradable photographs or retinal vascular occlusions, and 5) subjects who had missing data on the covariates used in the main admixture mapping scans (see “Admixture mapping”).
The retinal photography procedure and its assessment have been described in detail elsewhere [46]. Briefly, a 45-degree retinal photograph of one randomly selected eye of each participant was taken at visit 3 following 5 minutes of dark adaptation. This photograph was centered on the region of the optic disc and the macula and was taken using an autofocus camera.
Trained graders who were masked to participant measured the calibers of all arterioles and venules coursing through a specified area surrounding the optic disc according to a standardized protocol [46]. Individual vessel measurements were combined into summary indices: CRAE and CRVE. These indices represents average retinal arteriolar and venular caliber of the eye after taking into account the branching patterns. These measurements of retinal vascular calibers are reliable, with generally high intragrader and intergrader reliability coefficients (0.84 and 0.79, respectively) [46],[57].
Current blood pressure was defined as measurements at the time of retinal photography at visit 3, and 3- and 6- year past blood pressure was defined as measurement taken at visit 2 and visit 1, respectively. Mean arterial pressure was defined as two thirds of diastolic plus one third of systolic blood pressure. Mean arterial pressure across the three visits was averaged to obtain the 6-year mean arterial pressure. BMI was calculated as weight (in kg)/height (in meters) squared. Blood collection and processing followed a standard protocol [58]. Fasting glucose was assessed by a modified hexokinase/glucose-6-phosphate dehydrogenase procedure. Total plasma cholesterol and triglyceride were measured using enzymatic methods [59]. LDL cholesterol was calculated [60], and HDL cholesterol was measured after dextran-magnesium precipitation of non-HDL lipoproteins [58]. Cigarette smoking and alcohol consumption were ascertained from a questionnaire interview. Hypertension was defined as current systolic blood pressure ≥140 mm Hg, diastolic pressure ≥90 mm Hg, or self-reported use of medications for high blood pressure during the 2 weeks preceding the clinic examination at visit 3.
We genotyped a total of 1,536 SNPs included in the phase 3 admixture panel [28,61]. This panel was constructed by using the panel of ancestry informative markers previously published by Smith et al. (phase 1 panel) [24], improving this panel by mining new ancestry informative markers from the data sets of Hinds et al. [62] and the Phase 2 International Haplotype Map [63], and then validating them to confirm that they were indeed ancestry informative. Genotyping was performed by the Center for Inherited Disease Research (CIDR, Johns Hopkins University, Baltimore), using the Illumina BeadLab platform [64]. The ARIC study has a rigorous quality control program, including blind duplicates. Many genotypes in duplicates were obtained using the Illumina BeadLab technologies in ARIC African-American participants, and CEPH and Yoruban samples. The mismatch rate among duplicate genotypes was 0.1%.
We used previously published genotyping data to estimate the frequency of each SNP in West Africans and Europeans [24],[25],[64]. We used only those SNPs for which we were able to obtain data from both West African (Yoruba) and European American (CEU) populations from the International Haplotype Map. For SNPs in the phase 1 panel, we also added additional genotyping data from African and European samples, which was the same as the data collected in Smith et al. 2004 [24].
After genotyping, samples were eliminated based on the following criteria: 1) samples with low (<94%) call rate, 2) samples showing gender mismatch between self-reported data and genetically estimated gender based on 50 markers on the X chromosome, and 3) duplicate samples (defined as >75% match in the genotypes between two samples. Moreover, we used built-in data checking programs in the ANCESTRYMAP [23] software to exclude samples with an apparent excess or deficiency of heterozygous genotypes compared with the expectation from the individuals' global ancestry. An apparent excess of heterozygous genotypes (defined as the Z-score >10) usually indicates the individuals have parents with divergent ancestries (for example, one parent who is entirely of European ancestry) and such individuals nearly always have estimated European ancestry close to 0.5 [23].
To decrease the likelihood of false-positives in our admixture scans, we applied a series of filters that had the goal of detecting and removing any SNPs with problematic genotyping, as described previously [24,30,61,66]. First, SNPs were dropped if there were atypical clustering patterns, ill-defined clusters, or relatively low genotyping success rate (95%). This left us with 1,416 SNPs (all with genotyping call rate >97%). We then applied three previously described filters to further eliminate SNPs from the analysis [66]. 1) We eliminated SNPs (N = 15) if they did not meet the requirement for Hardy-Weinberg equilibrium (P>0.01) in both ancestral West African and European populations. 2) We applied a “freqcheck” filter that examined whether the observed frequency of a SNP in African Americans was statistically consistent with being a mixture of the frequencies observed in the West Africans and European American samples that we used to represent the ancestral populations [23]. 3) Lastly, we applied a “ldcheck” filter that for each sample, iteratively eliminated SNPs that were less informative (in terms of the information content about ancestry) until none were within 200 Kbs of each other or in detectable LD with each other in the ancestral West African or European populations [23]. After imposing these requirements, 1,365 SNPs were left for analysis.
Using the ANCESTRYMAP software [23], we estimated a global ancestry for each individual, as indicated by PAA. ANCESTRYMAP uses a Markov Chain Monte Carlo approach to account for uncertainty in the unknown parameters (including SNP allele frequencies in the West African and European ancestral populations, the number of generations since mixture, and the average proportion of ancestry inherited from ancestral populations) that emerge from the Hidden Markov Model analysis. All Markov Chain Monte Carlo runs used 100 burn-in and 200 follow-on iterations, as recommended [23], except for one longer run of 1,000 burn-in and 2,000 follow-on iterations, which we used to check the stability of our results.
We used the ANCESTRYMAP software [23] to search for genomic regions associated with an increased percentage of either European or African ancestry. The main dichotomous admixture scans used the values of retinal vascular caliber adjusted using the following covariates: age, sex, study sites (Forsyth County or Jackson), 6-year mean arterial blood pressure and fasting glucose level, because the latter two systemic factors were both significantly correlated with PAA (both P<0.01) and known to be associated with retinal vascular calibers [4],[7],[8]. For the tests for associations of the local ancestry, we additionally adjusted for other cardiovascular risk factors.
For the purpose of this dichotomous admixture analysis, study participants were ranked by the adjusted values for each trait. For CRAE, the 15% of individuals with the lowest values were defined as cases, and the 15% with the highest values as controls. For CRVE, conversely, the 15% with the highest values for were defined as cases, and the 15% with the lowest values as controls. Because ANCESTRYMAP uses Bayesian statistics, a prior distribution of risk models is required [23]. We tested 18 pre-specified European ancestry risk models to assess overall evidence of association by averaging across all models. The first 6 models used 0.4, 0.5, 0.67, 1.5, 2.0 and 2.5-fold risks of being a case due to inheritance of one copy of an European ancestral allele, with a risk of 1 for carrying zero copies of an European ancestral allele. The next 6 models used the same risk set as the first for carrying one copy of an European ancestral allele, whereas the risk of carrying zero copies were set to the reciprocal of the risks for carrying one copy. The last 6 models specified that inheritance of either one or two copies of an European ancestral allele had a risk of 1, but carrying zero copies had risks of 0.4, 0.5, 0.67, 1.5, 2.0 and 2.5. By convention used in the manuscript, a risk <1.0 for inheritance of one copy of an European ancestral allele at a given locus represents a risk model where European ancestry decreases risk relative to African ancestry. This set of models reflects the hypothesis that European ancestral alleles are less likely to confer risks but also tests for the alternative possibilities [23].
ANCESTRYMAP provided two scores to assess statistical significance: a locus-specific score and a case-control score [23]. A locus-specific score was obtained in cases (case-only analysis) by calculating the likelihood of the genotyping data at the SNPs at the locus under the risk model and comparing it to the likelihood of the genotyping data at the SNPs at the locus assuming that the locus is unassociated with the phenotype. The ratio of these two likelihoods is the “likelihood ratio”, and the log-base-10 of this quantity is the “LOD” score. A locus-specific LOD score of >5 has been recommended as criterion for genome-wide significance [37]. To obtain an assessment of the evidence for a risk locus anywhere in the genome, we averaged the likelihood ratio for association across all loci in the genome, and took the log10 to obtain a “genome-wide score”. We interpreted a genome-wide score >2 as significant [37].
A case-control score was calculated by comparing locus-specific deviations in European ancestry in cases versus controls at each locus across the genome. This score tests whether any deviation in ancestry from the genome-wide average is significantly different comparing cases with controls [23]. If there is no locus associated with phenotype, the case-control score is expected to be distributed approximately according to a standard normal distribution. For loci identified by the case-control score, the level of genome-wide significance was defined as a Z score >4.06 or <−4.06, corresponding to an uncorrected nominal P<5×10−5, or a corrected nominal P<0.05 after conservatively correcting for 1,000 hypotheses tested (approximately equals the number of independent chromosomal chunks assigned to either African or European ancestry).
The main admixture scans were based on a dichotomous phenotype (i.e., cases and controls) in a subset of our samples. To check whether the results were consistent in our entire sample, we used the ANCESTRYMAP software to obtain local estimates of African ancestry at the admixture peak [23], and then assessed the association of the local ancestry with retinal venular caliber as a continuous trait using linear regression models. Using ANCESTRYMAP, we also obtained regional estimates of ancestry based on the SNPs on chromosome 6p in the admixture panel. In addition to the five covariates used in the main admixture scans, CRVE was further adjusted for HDL cholesterol, LDL cholesterol and plasma triglyceride level, BMI, smoking, and alcohol consumption, all of which covariates have been shown to affect retinal vascular calibers [6]–[8],[14]. We then performed a normal-quantile transformation for CRVE to ensure normality. In the linear regression models, the transformed CRVE was used as a dependent variable and the local estimates of ancestry as an independent variable. To determine whether there was evidence of residual association with local ancestry after adjustment for global and regional ancestry, we included each individual's PAA and estimated regional ancestry on chromosome 6p as covariates in the regression models. This enabled us to increase power by including all samples in a quantitative analysis, rather than using only a subset of samples with the highest 15% and lowest 15% values in the dichotomous admixture scans described above. To determine whether the association between the local ancestry and CRVE differed significantly by hypertension status, Z tests were used to compare the difference in the regression coefficients obtained from the hypertensive and non-hypertensive groups.
To determine a 95% CI for the position of a trait locus, we obtained the likelihood ratio for association at each marker on the chromosome where we identified an association. This provided a Bayesian posterior probability for the position of the underlying causal variant assuming a flat prior distribution across the region for the position of the trait locus. We defined the CI as the central region of this peak containing 95% of the area.
|
10.1371/journal.pcbi.1005304 | Suboptimal Criterion Learning in Static and Dynamic Environments | Humans often make decisions based on uncertain sensory information. Signal detection theory (SDT) describes detection and discrimination decisions as a comparison of stimulus “strength” to a fixed decision criterion. However, recent research suggests that current responses depend on the recent history of stimuli and previous responses, suggesting that the decision criterion is updated trial-by-trial. The mechanisms underpinning criterion setting remain unknown. Here, we examine how observers learn to set a decision criterion in an orientation-discrimination task under both static and dynamic conditions. To investigate mechanisms underlying trial-by-trial criterion placement, we introduce a novel task in which participants explicitly set the criterion, and compare it to a more traditional discrimination task, allowing us to model this explicit indication of criterion dynamics. In each task, stimuli were ellipses with principal orientations drawn from two categories: Gaussian distributions with different means and equal variance. In the covert-criterion task, observers categorized a displayed ellipse. In the overt-criterion task, observers adjusted the orientation of a line that served as the discrimination criterion for a subsequently presented ellipse. We compared performance to the ideal Bayesian learner and several suboptimal models that varied in both computational and memory demands. Under static and dynamic conditions, we found that, in both tasks, observers used suboptimal learning rules. In most conditions, a model in which the recent history of past samples determines a belief about category means fit the data best for most observers and on average. Our results reveal dynamic adjustment of discrimination criterion, even after prolonged training, and indicate how decision criteria are updated over time.
| Understanding how humans make decisions based on uncertain sensory information is crucial to understanding how humans interpret and act on the world. Signal detection theory models discrimination and detection decisions as a comparison of “stimulus strength” to a fixed criterion. In a world that is constantly changing a static criterion makes little sense. We investigate this as a problem of learning: How is the decision criterion set when various aspects of the context are unknown (e.g., category means and variances)? We examine criterion learning in both static and dynamic environments. In addition to a more traditional discrimination task in which the criterion is a theoretical construct and unobservable, we use a novel task in which participants must explicitly set the criterion before being shown the stimulus. We show that independent of environment and task, observers dynamically update the decision criterion, even after prolonged training in a static environment. Our results provide evidence against an assumption of stability and have implications for how psychophysical data are analyzed and interpreted and how humans make discrimination decisions under uncertainty.
| Understanding how humans make decisions based on uncertain sensory information is crucial to understanding how humans interpret and act on the world. For over 60 years, signal detection theory has been used to analyze detection and discrimination tasks [1]. Typically, sensory data are assumed to be Gaussian with equal variances but different means for signal-absent and signal-present trials. To decide, the observer compares the noisy sensory data to a fixed decision criterion. Performance is summarized by d′ (discriminability) and c (decision criterion) based on measured hit and false-alarm rates. Standard analysis assumes stable performance (all parameters fixed) and observer knowledge of the means, variance, prior probabilities and payoff matrix [1,2,3,4,5,6,7].
The assumption of stable performance is problematic for two reasons. (1) Observers may learn about the environment and use that information to set the decision criterion. (2) The environment may not be stable or the observer may not believe that the environment is stable. To circumvent these problems, researchers include training sessions, fix the environmental parameters (e.g., priors, payoffs) within blocks, and treat learning effects as additional noise (i.e., its “variance” can simply be added to those of internal and/or external noise in the experiment). However, research investigating history effects in psychophysical tasks has shown that an observer’s current decision is affected by multiple aspects of the stimulus history (e.g., recent decisions, stimulus intervals, trial type, etc.). These effects occur even when the environment is stable, the stimulus presentation is random, and observers are well trained [8,9,10,11,12,13]. Observers behave as if the environment is dynamic and, as a result, measures of discriminability and sensitivity are biased and the confidence intervals computed for the best fitting parameters of the psychometric function are too narrow [14]. While assuming instability in a static world is suboptimal, in a world that is constantly changing a fixed criterion makes little sense.
To optimize decisions in dynamic environments, observers must update decision criteria in response to changes in the world by adapting to the value and uncertainty of sensory information. Humans respond appropriately to changes in visual and motor uncertainty [15,16,17,18,19,20]. Observers adjust the decision criterion when uncertainty is varied randomly from trial to trial [18]. If the location of visual feedback for a reach is perturbed dynamically over trials, participants track this random walk near-optimally [15]. Landy and colleagues [17] demonstrated that participants tracked discrete changes in the variance of a visual perturbation. Summerfield and colleagues [19] investigated a visual discrimination task in which participants categorized gratings with orientations drawn from two overlapping distributions. Means and variances were updated randomly with different levels of volatility. Participants’ performance changed as a function of volatility.
While the above studies examine the dynamics of decision-making, they only provide indirect evidence of criterion shifts. Many of these studies observed changes in decisions and response time, but few studies have examined how trial history specifically affects decision criteria and what is the underlying mechanism responsible for learning and updating the decision criterion. Lages and Treisman [13] describe the dynamics of criterion setting and updates of priors based on previous stimulus samples and responses applied to tasks with no experimenter feedback, so that the criterion drifts to the mean of previously experienced stimuli. Summerfield and colleagues [19] consider a discrimination task with feedback in which the categories and their associated uncertainty can change several times per block of trials. They compare several suboptimal models, all of which predict the choice probability by probability matching.
Here, we investigate how humans learn to set and update criteria for perceptual decisions in both static and dynamic environments. To examine the underlying mechanisms of criterion learning, we take a quantitative approach and compare models of how a decision criterion is set as a function of recently experienced stimuli and feedback. Observers completed two different experimental tasks. One task was the typical discrimination task, in which the observer’s criterion is unobservable. We introduce a novel overt-criterion task, in which the decision criterion is set explicitly by the observer. This allows us to measure and model the setting of the decision criterion directly. We used the overt-criterion task, which has greater statistical power due to the richer dataset, to develop and test models of how the criterion is updated in standard discrimination experiments under uncertainty. In contrast to the models investigated by Summerfield and colleagues [19], we directly measure the criterion, and include parameters for sensory noise and predict a specific response based on the noisy stimulus information and a model of criterion update. While observers converged to the optimal criterion over many trials when conditions were static and followed dynamic changes in the category means, we found that, in both tasks, the majority of observers used suboptimal learning rules. Our results reveal dynamic adjustment of a discrimination criterion, even after prolonged training in a static environment.
All observers completed three tasks: (1) An orientation-discrimination task in which discrimination thresholds were measured and used to equate the difficulty of the covert- and overt-criterion tasks across observers (Fig 1A), (2) A covert-criterion task in which observers categorized an ellipse as belonging to category A or B (Fig 1C), and (3) An overt-criterion task in which observers explicitly indicated their criterion on each trial prior to the presentation of a category A or B ellipse (Fig 1D). Additionally, 8 out of 10 observers completed an orientation-matching task in which adjustment noise was measured (Fig 1B). Categories in the covert- and overt-criterion tasks were Gaussian distributions with different mean orientations and equal variance (Fig 1E; see Methods).
In Expt. 2, observers completed the same tasks in Expt. 1. However, the environment was dynamic and only 3 out of 10 observers completed the orientation-matching task. Category distributions were Gaussian distributions with different mean orientations and equal variance, but means of the category distributions changed gradually over time via a random walk (see Methods).
The present study examined the strategies observers used to learn and update their decision criterion in an orientation-discrimination task under both static and dynamic uncertainty. Under static conditions in which category means were constant, we showed that while observers converged to the optimal criterion over many trials, their trial-by-trial behavior was better described by suboptimal learning rules than by the optimal rule. Thus, even though conditions were static, the criterion continued to systematically drift with changes in stimulus statistics throughout the experiment. Under dynamic conditions in which category means changed slowly over time, observers followed changes in the means of the category distributions closely with a 1–4 trial lag. Specifically, we found that at the group level a model in which the recent history of past samples determines a belief about category means, the exponentially weighted moving-average rule, was more likely than the alternative models across most tasks and conditions with the exception of the overt-criterion task under dynamic conditions in which the reinforcement learning model was more likely. Our results suggest that the decision criterion is not fixed, but is dynamic, even after prolonged training. Finally, we provided a novel technique, the overt-criterion task, which can be used to explicitly measure criterion placement and a computational framework for decision-making under uncertainty in both static and dynamic environments.
Based on findings in the visuo-motor and reinforcement-learning literature, in which feedback is gradually or discretely updated [15,16,17,20,21], we would expect a model in which recent samples are given more weight to better explain performance under dynamic conditions. However, this is a suboptimal strategy under static conditions. Nevertheless, research on history effects in psychophysical tasks suggests that observers behave as if the environment is dynamic, which is consistent with our results [8,9,10,11,12,13,14]. Furthermore, the regression analysis we performed in Expt. 1 revealed beta weights for the covert- and overt-criterion tasks that suggest an exponentially weighted average of the previous stimuli. Overall, our analysis provides additional evidence against the ideal-observer model and the assumption of a stable criterion, even in static environments. Intuitively, in a world that is constantly changing, it makes sense to continually update your decision criterion, weighting your most recent experiences more heavily.
The previous study that is closest in spirit to the current work is that of Summerfield and colleagues [19]. Their experiment was similar to ours; in their case category means changed suddenly at every 10 or 20 trials, and category variances could also change. However, they used a traditional discrimination task without explicit measurement of the criterion, and their primary analysis used the predictions of each of three models in a decidedly suboptimal manner: probability matching. They found two extremely different models, a limited-memory model and a Bayesian model (that uses probability matching rather than the optimal decision) both accounted for significant amounts of variance in their data. In our analysis, we are interested in the entire sequence of computations from estimating the stimulus parameter of interest (orientation, perturbed by sensory noise) through the binary category decision, and compare a wider array of models that include the ideal observer. We found that suboptimal models that use the recent history of past samples best accounted for both covert and explicit criterion setting [19].
While a dynamic decision criterion might be useful in the real-world, by using such a criterion (especially in Expt. 1) in our experiments, observers are making suboptimal inferences about the category membership of an ellipse. These results are consistent with the idea that suboptimal inference is more than just internal noise [22]. This is also consistent with the overestimates of the noise parameters that we find in our model fits, which suggests that there is additional noise beyond sensory and adjustment. Acuña and Schrater [23] suggest that seemingly suboptimal decisions in sequential decision-making tasks can be accounted for by uncertainty in learning the structure of the task. Uncertainty about the structure of the environment could affect observers’ criterion placement (i.e., observers might be uncertain as to whether the category parameters are changing and/or the rate of change). In novel situations, one must learn the task structure and the parameters of the environment to perform optimally.
For the purpose of increasing statistical power for our model comparison, we introduced a novel task, in which we made the decision criterion explicit. Previous research suggests that participants change strategies when implicit tasks are made explicit [24,25,26]. Specifically, participants who perform optimally during an implicit task are not optimal when the task is made explicit. This is thought to be a result of higher-level strategic adjustments interfering with lower-level processing. While strategies were fairly consistent under static conditions, we found a clear difference in preferred strategy under dynamic conditions. Specifically, we found the exponentially weighted moving-average model fit best in the covert-criterion task and the reinforcement learning model fit best in the overt-criterion task. Additionally, we observed a difference in the exponentially weighted moving-average model’s decay rate and the reinforcement learning model’s learning rate. Across experiments, the decay and learning rates under static conditions were slower than decay and learning rates under dynamic conditions. However, there was also a difference across tasks. In both experiments, the decay rate was slower in the overt- than the covert-criterion task and the learning rate was faster in the overt- than covert-criterion task. Since a slower decay rate is beneficial under static conditions but disadvantageous under dynamic conditions and a faster learning rate is beneficial under dynamic conditions but disadvantageous under static conditions, the parameter differences we observed might explain the differences we see in the preferred strategies used across tasks. In particular, this may explain why the reinforcement learning model performed better than the exponentially weight moving-average model in the overt-criterion task under dynamic conditions. The differences in decay and learning rate between the covert- and overt-criterion tasks might be due to a difference in time-scale that results from the temporal dynamics of the two tasks (the overt-criterion task took twice as long to complete the same number of trials) or due to the different levels of processing (e.g., sensory vs. motor) required for each task. In the future, it might be interesting to see how the decay and learning rates trade off as a function of the rate of change (i.e., the random-walk variance) in the experiment.
Previous research shows that participants update the decision criterion when changes to the prior probabilities and payoff matrix occur [1,2,3,4,5]. There is a systematic bias in these shifts: Humans exhibit conservatism, that is, a bias towards the neutral criterion when the optimal criterion is shifted away from neutral. While several hypotheses have been proposed as to why conservatism occurs, most recently Ackermann and Landy [2] have suggested that conservatism can be explained by distorted probability and utility functions. Our results do not explain this bias, but it is likely that conservatism is present and contributes to the dynamics of trial-by-trial criterion shifts under the conditions of Expt. 2. To provide a better understanding of this bias, further research should aim to examine criterion learning in situations in which conservatism is known to exist.
Finally, psychophysical studies rely heavily on accurate estimates of d′. By calculating d′ from hit rates and false alarms in the usual way, a fixed criterion is assumed. However, if the observer’s criterion varies over trials, performance will be a mixture of multiple points on the ROC curve, resulting in a biased (too-low) estimate of d′. We have shown here that decision criteria are adjusted dynamically. Examining the dynamics of trial-by-trial criterion placement provides us with a richer understanding of participants’ behavior when making decisions in the presence of uncertainty. Our results suggest that typical estimates of d′ are biased, and that by using a model that accounts for a dynamic criterion we can compute a more accurate measure of discriminability and in turn, obtain a more comprehensive understanding of discrimination under uncertainty.
This research involved the participation of human subjects. The Institutional Review Board at New York University approved the experimental procedure and observers gave informed consent prior to participation.
Ten observers participated in Expt. 1 (mean age 25.4, range 20–33, 5 females) and Expt. 2 (mean age 23.4, range 20–28, 4 females). Five observers provided data for both experiments, three of whom completed Expt. 1 prior to completing Expt. 2. All observers had normal or corrected-to-normal vision. One of the observers (EHN) was also an author.
Stimuli were presented on a gamma-corrected Dell Trinitron P780 CRT monitor with a 31.3 x 23.8 deg display, a resolution of 1024 x 768 pixels, a refresh rate of 85 Hz, and a mean luminance of 40 cd/m2. Observers viewed the display from a distance of 54.6 cm. The experiment was programmed in MATLAB (MathWorks) using Psychophysics Toolbox [27,28].
Stimuli were 10 x 2° ellipses presented at the center of the display on a mid-gray background (Fig 1). In the orientation-matching and overt-criterion tasks, a yellow line was presented at the center of the display (10 x .35°). In all tasks except the overt-criterion task, trials began with a central yellow fixation cross (1.2°).
Ten observers participated in one, 1.5-hour session consisting of an orientation-discrimination task (~10 min), a covert- and an overt-criterion practice block (~5 min combined), one block of the covert-criterion task (~20 min), and one block of the overt-criterion task (~40 minutes). The order of the covert- and overt-criterion tasks was randomized across subjects. Eight out of ten observers returned for a second session in which they completed an orientation-matching task (~20 minutes). Before starting the experiment observers were given detailed instructions regarding the tasks they would be asked to complete. The two short (20 trial) practice blocks were used to ensure that observers understood the experimental tasks. Before each block, a condensed version of the instructions and the name of the task were shown to remind observers of the procedure and inform them of the task they would be completing on that block.
Expt. 2 was similar to Expt. 1 except that observers did not complete the orientation-matching task and the category distribution means in the covert- and overt-criterion tasks were not constant throughout the block. Rather, category means were updated on every trial following a random walk. The category A mean on trial n+1 was μA,n+1 = μA,n+ε, where ε ~N(0,σrandom) and σrandom = 5°. The relative position (μA < μB) and the distance between the means remained constant.
In the covert-criterion task, the statistical structure of the task involves three variables: category C, stimulus orientation S, and measurement X. On each trial, C is drawn randomly and determines whether S is drawn from category A (N(μA,σ)) or category B (N(μB,σ)). We assume that on each trial, the true orientation is corrupted by sensory noise (with standard deviation σv) to give rise to the observer’s measurement of orientation (X~N(S,σv)). The observer uses this measurement to infer the category.
In the overt-criterion task, the statistical structure of the task involves five variables: criterion orientation θc, criterion placement z, category C, stimulus orientation S, and measurement X. On each trial, criterion orientation is inferred from the previous trials. We assume that criterion orientation is corrupted by adjustment noise (z~N(θc,σa)). After the criterion is set, C is drawn randomly and determines whether S is drawn from category A (N(μA,σ)) or category B (N(μB,σ)). As in the covert-criterion task, we assume the true orientation of the stimulus is corrupted by sensory noise (X~N(S,σv)). Finally, the observer uses this measurement and the feedback about its category membership to update the criterion orientation for the next trial. We found that model fits for the overt case could not discriminate adjustment noise (σa) from sensory noise (σv), and so for this case, sensory noise was fixed and only an adjustment noise parameter was fit. Sensory noise was set to each observer’s measured sensory uncertainty. Below we describe both optimal and suboptimal models of criterion learning that vary in computational and memory demands. The selection of the following models was partially inspired by the models used in Summerfield and colleagues’ research [19] investigating perceptual classification strategies in rapidly changing environments, in which they compared a Bayesian observer model to a Q-learning model and a heuristic model that is similar to our limited-memory model. In their models, sensory noise is omitted, and in its place, a fixed degree of trial-trial choice variability is introduced by a probability-matching rule. In contrast, we compare a more extensive set of models that include parameters controlling the level of sensory noise and predict a specific response based on the noisy stimulus measurement and a model of criterion update.
|
10.1371/journal.pcbi.1003001 | Cooperative Adaptive Responses in Gene Regulatory Networks with Many Degrees of Freedom | Cells generally adapt to environmental changes by first exhibiting an immediate response and then gradually returning to their original state to achieve homeostasis. Although simple network motifs consisting of a few genes have been shown to exhibit such adaptive dynamics, they do not reflect the complexity of real cells, where the expression of a large number of genes activates or represses other genes, permitting adaptive behaviors. Here, we investigated the responses of gene regulatory networks containing many genes that have undergone numerical evolution to achieve high fitness due to the adaptive response of only a single target gene; this single target gene responds to changes in external inputs and later returns to basal levels. Despite setting a single target, most genes showed adaptive responses after evolution. Such adaptive dynamics were not due to common motifs within a few genes; even without such motifs, almost all genes showed adaptation, albeit sometimes partial adaptation, in the sense that expression levels did not always return to original levels. The genes split into two groups: genes in the first group exhibited an initial increase in expression and then returned to basal levels, while genes in the second group exhibited the opposite changes in expression. From this model, genes in the first group received positive input from other genes within the first group, but negative input from genes in the second group, and vice versa. Thus, the adaptation dynamics of genes from both groups were consolidated. This cooperative adaptive behavior was commonly observed if the number of genes involved was larger than the order of ten. These results have implications in the collective responses of gene expression networks in microarray measurements of yeast Saccharomyces cerevisiae and the significance to the biological homeostasis of systems with many components.
| Homeostasis is an inherent property of biological systems, which have a general tendency to adapt, i.e., to recover their original state following environmental changes. In cells, this adaptation is mediated by changes in protein expression. Initially, cells respond to environmental changes by altered gene/protein expression; subsequently, the expression of most genes returns to basal levels, albeit not completely, as shown by recent experimental analyses of yeast. Although simple mechanisms for adaptation through network motifs, composed of just a few genes, are well understood, how regulatory networks involving many genes that activate or repress each other can generate adaptive behaviors is unclear. Here, by numerically evolving gene regulatory networks, we obtained a class of genes whose expression dynamics showed adaptation over almost all genes, from which we revealed the general logic underlying such adaptive dynamics with many degrees of freedom, which was not reducible to motifs with a few genes. This adaptation was cooperative, i.e., adaptation of one gene mutually relied upon others' adaptive expressions. Moreover, this collective behavior was robust to noise and mutations. The present study sheds a light on the nature of collective gene expression dynamics allowing for biological homeostasis.
| Adaptive responses to environmental changes are fundamental to all living organisms. When environmental conditions change, the cellular concentrations of some chemicals change immediately in response; however, the degree of change is later reduced, returning closer to the basal state. Thus, in general, some variables within a biological system first change in response to environmental changes, but then slowly revert back to pre-stimulus values by adjusting the expression levels of proteins or mediating cellular activity for adaptation to the new conditions. In such an adaptive response, some internal variables change according to the external conditions, while other variables return to the original values, thus realizing both responsiveness and homeostasis.
Recently, simple reaction dynamics models for adaptive response with few degrees of freedom have been studied. For example, Francois and Siggia noted two characteristics in such responses: responsiveness and perfectness of adaptation [1]. They carried out numerical simulations of the evolution of parameter values in simple network motifs consisting of three components to show that both these characteristics are realized. Similarly, Ma et al. studied all possible three-node enzyme network topologies numerically to identify those that exhibit adaptive responses [2]. They found that only two major core topologies can show an adaptive response: a negative feedback loop with a buffering node and an incoherent feed-forward loop with a proportioned node.
In fact, such adaptive responses have been studied with simple chemical reaction models with a few components (e.g., proteins) [3]–[8]. After the immediate response, the expression of one component returns to its original value, and changes in the external conditions are compensated for by adjusting the other components within the system. Such simple chemical reaction dynamics are also abstracted from complex reaction networks as motifs, as mentioned above [9], [10]. However, in real biological reactions, the expression levels of many proteins influence each other through mutual activation and inhibition of gene expression. Adaptive responses stemming from such complex reaction dynamics involve a huge number of chemical species or the expression dynamics of many genes. Indeed, the simple network motifs proposed above may exist as a part of a network but cannot function in isolation [11]. Although simple models could possibly be derived by reducing the degrees of freedom in a complex reaction network, no such reduction scheme is yet available. Therefore, it is important to study adaptive responses within a system consisting of many proteins.
One reported example of such an adaptive response with many degrees of freedom concerns the gene expression patterns in yeast Saccharomyces cerevisiae subjected to diverse environmental changes, including temperature shock, hydrogen peroxide treatment, amino acid starvation, and nitrogen source depletion. Studies using DNA microarrays have shown that certain sets of genes (approximately 900 genes) exhibit similar responses to almost all of these environmental changes, while some genes show unique response patterns to specific conditions only [12]–[14]. For example, after a temperature shift, many genes are either up-regulated or down-regulated shortly after the stimulus and then gradually return to pre-stimulus expression levels. Moreover, many genes that do not specifically respond to heat shock stimulus also show adaptive responses. Such responses are called “stereotyped” responses, involved in protecting and maintaining critical features of the intracellular system. Several other reports have also suggested that a large fraction of genes, i.e., approximately 50%–70% of genes, show adaptive responses. The response is not monotonic; initially, genes expression is altered in response to the stimulus, but this change is later compensated for, at least partially. That is, many gene undergo initial up-regulation followed by down-regulated, or vice versa, returning to (nearly) basal expression [15], [16]. Such adaptation without complete return to the original level is termed as partial adaptation.
This type of multidimensional adaptive behavior should be quite rare in an arbitrary dynamical system with many degrees of freedom. Indeed, it is not common for a system to exhibit changes in a large number of variables in response to input parameters (environmental conditions) and later return these variables to the original values. Furthermore, designing such a system would become increasingly more difficult as the number of involved variables increases. Unveiling the characteristic properties of such an unusual dynamical system of gene expression is the main purpose of the present study.
If such singular behavior is observed ubiquitously in biological systems, one possible origin could be selection through evolution. That is, through the selection of functional networks for higher fitness values, rare networks exhibiting such atypical behavior may evolve. Here, however, we should note that the existence of such selective pressure toward adaptive dynamics over the expression of many proteins is not easy to imagine. For a given environmental change, selection process to achieve the adaptive response of only one or a few specific genes would be naturally expected. There is no need to postulate that many genes exhibit adaptive expression responses for a fitness for selection. Hence, it is important to determine how genes that do not need to be adaptive indeed show adaptive response collectively, as is commonly observed in responses of micro-organisms. Can such adaptive responses over many genes evolve through a numerical evolutionary process of gene regulatory networks by imposing a single, simple fitness condition?
Here, we answer this question by examining the evolution of regulatory networks involving changes in the expression levels of many genes. We numerically evolved these networks by using genetic algorithms with a fitness condition for the adaptive behavior of only a single target gene. Although the specific evolutionary course may not be realistic due to the simplified fitness conditions adopted here, ‘cooperative’ adaptive responses were generally observed. Hence, we expect that these shed a new light on characteristic features of adaptive systems. We also discuss the relevance of such adaptive responses over many genes, cooperative in nature, to biological functions and the possible relationships of these responses with gene expression patterns in yeast Saccharomyces cerevisiae observed by microarray analysis.
We modeled gene expression dynamics using a regulatory network to study adaptive response with many degrees of freedom, following the methods presented in earlier studies [17]–[21]. In a regulatory network, there are nodes corresponding to each gene. The expression level of a gene is represented by the variable . By appropriately normalizing gene expression levels we set , where represents a suppressed state and represents a highly expressed state.
Next, we evolved the regulation matrix by imposing the fitness condition that the target gene shows an adaptive response.
In our model, the adaptive response of the target gene evolved within a few hundred generations to the highest fitness (Figure 3(a)). Here, fitter networks that produce offspring reach almost the highest fitness value after evolution, whereas the minimal fitness value of the networks that are discarded are .
Now, we investigate the response of non-target genes. The response of each gene expression () to a change in is generally classified into the following three cases: (i) perfect or partial adaptation, (ii) monotonic (non-adaptive) relaxation, and (iii) no response (no change at all). Under the fitness conditions set here, the responses of non-target genes do not matter at all, and they can be either adaptive, monotonic, or non-responsive. As seen in Figure 4, however, the responses of an increasingly larger number of genes turn from monotonic to adaptive with continued evolution. In the early generations (Figure 4(a)), almost all non-target genes show monotonic relaxation, but with evolution, more genes with monotonic responses begin to show adaptive responses (Figure 4(c)). To characterize this trend quantitatively, we measured the ‘average adaptiveness’, i.e., the average value of , defined by for and . We did not include the input and target genes for the average, since the input gene always shows monotonic increase, while the target gene always shows adaptive behavior after generations to achieve a high fitness value. The average adaptiveness does not always tightly correspond to the fitness value and there are some networks with high fitness values and small . However, through the course of evolution, as shown in Figure 3(b), we see that an increase in the fitness value is accompanied by an increase in , and networks with larger have higher fitness values in general. By comparing Figure 3(a) with (b), it can be seen that the fitness first increases to a relatively high level, and then the increase in average adaptiveness progresses with generations while the fitness of the target gene keeps on increasing, albeit gradually. After the first increase in fitness, many tiny increases are achieved resulting in large . We term such response with adaptive responses by many non-target genes, i.e. with large value, as a ‘cooperative adaptive response’. Note that in this model, as we prohibit feedback interactions from the target gene, i.e. , the adaptive responses in non-target genes do not originate in the adaptive response of the target gene. Moreover, from the fitness condition itself, there is no selection pressure for the adaptive responses of non-target genes. Perfect adaptation of the target gene is achieved rapidly, and the cooperative adaptive responses of the non-target genes evolve subsequently though there is no request.
Now we take a closer look at the high-dimensional gene expression dynamics involved in realizing this cooperative adaptive response. The dynamics show three stages of evolutionary change in reaching this cooperative adaptive response.
Cooperative adaptive responses require other genes that show adaptive responses and thus represent the collective dynamics through variations in a number of genes. Hence, this type of response can be expected to require a certain minimum number of genes. In Figure 5, it can be easily observed that the average adaptiveness was small for a small network size but increased as the networks became larger. In the region , the average adaptiveness were distributed depending on the evolutionary course. However, for , cooperative adaptive responses with a large values were always attained. Thus, cooperative adaptive responses require a sufficiently large number of genes, and do not emerge at all for .
The fitness values were indeed smaller for , compared with the case for with . The fitness value increases toward the maximal level as increased from 5 to 10, accompanied by an increase in . (Note, however, the fitness for could be increased by fine-tuning the parameters of the gene expression dynamics [1], [2], while such tuning was not necessary for ). Moreover, networks with larger showed higher performance than smaller networks (Figure S1).
Now, we examine the network motifs to achieve the adaptation. According to a previous study [2], only the two types of network motifs with three elements, i.e., the incoherent feed-forward loop ( and in Figure 6) and negative feedback loop [23] can show adaptive responses. For the feed-forward network , as redrawn in Figure 7 (lower right panel), the gene showed a decreasing response by direct negative interaction with input gene right after the imposition of the input , but later increased back to original levels as a result of the positive input of the gene , with a time delay. Thus it showed a downward adaptive response.
In a small network with , which is noncooperative, such motifs and their combined motifs appear with higher frequencies than others. These play a major role in the adaptive response of the target gene. On the contrary, in a network with many degrees of freedom, these motifs are not frequent. Only at the initial stages of evolution adopting the timing difference strategy, they appear frequently, and the network is reduced to this type of incoherent feed-forward loop when coarse-grained. Later, however, with the increase of average adaptiveness , their frequency decreases, and it is much smaller than that expected for a random network.
Moreover, such three-gene motifs for adaptation can be completely deleted while maintaining higher fitness and values. In Figure 7, we show an example of these types of networks, which exhibit adaptive behavior with a high value; these networks did not include any of the three-gene adaptive motifs. Instead, other motifs become more frequent for networks with a high networks (Figure 6 and Figure S2).
We focused on the structures of the input, output, and middle genes that intervene the two. As the negative feedback loop is not so relevant to minimum adaptive motifs in our model, and the frequency of occurrence of such negative-feedback motifs was indeed very small, we focused on the frequencies of the feed-forward loop motifs [24]. We found that the minimal adaptive motifs ( and in Figure 6) appeared frequently in small networks, but were less frequent in large networks. Instead, and , which are not adaptive motifs, become more frequent.
We also studied network motifs with all possible combinations of three genes (Figure S2). We found that minimal network motifs for adaptation with three genes occurred at about the same frequency as random network cases. Moreover, they decreased in frequency as average adaptiveness increased. In contrast, motifs exhibiting mutual inhibition and activation were present in a significantly high fraction of cases and increased in frequency with increases in average adaptiveness (Figure S2).
As already mentioned, in cooperative adaptation, genes differentiate into two groups, activating or inhibiting, with respect to the target gene. The dominant motifs – and in Figure 6 – indeed corresponded to these two groups. For this separation into two groups, genes in different groups often inhibit each other mutually, while genes in the same group activate each other (Figure 8). That is, genes that provide positive inputs to the target gene activate the target as well as other target-activating genes, repress target-inhibiting genes, and vice versa. Thus, motifs exhibiting mutual inhibition and activation increased in frequency as average adaptiveness was increased (Figure S2). In summary, networks showing cooperative adaptive responses with many degrees of freedom have characteristic structures different from those of non-cooperative networks with few degrees of freedom.
Now we examine how some networks show a ‘cooperative adaptive response’, without standard minimal adaptive motifs of three genes. An example of such a network is shown in Figure 7. Although there is no three-gene motif, such networks generally include structures with four or more genes that play a similar role as minimal adaptive motifs, somewhere in the whole network. In Figure 7 (lower left panel), for example, a core structure from Figure 7 (top) was extracted where genes (input gene), , , , and form this type of structure, with a feed forward network. Here, in such extracted motif, we call the gene , , and ‘mediator’ which correspond to the gene in the three-gene motif (lower left panel of Figure 7), and call the gene ‘receiver’ which corresponds to the gene . A receiver gene is not necessarily the target gene in the whole network. In this case, the receiver gene showed a downward adaptive response as a result of direct negative input from gene , and later positive input mediated by three genes (, , and ) with a sufficient time delay. In essence, a rapid direct input from the ‘input gene’ and an opposing delayed input from the mediator gene(s) comprised the adaptive response.
So far the mechanism for the adaptive response of the receiver gene was common with the minimal adaptive motifs. However, there is a critical difference in the cooperative adaptive response, i.e., mediators show adaptive responses rather than monotone responses. In this case, gene showed an adaptive response, as demonstrated in Figure 9. This is a strong contrast with the minimal adaptive motif, where the mediator genes show monotonic response (as in and in Figure 6). When mediators show adaptive responses, the receiver is rapidly activated (or inhibited) by the input gene and is later inhibited (or activated) by mediators. This can result in the adaptive behavior of the receiver gene. However, there is some ‘danger’. When the mediators returned to original levels, then the total input to the receiver would also return to its original state (at ) again. Then the up-down (or down-up) response in the receiver expression would be repeated again. Thus, the expression of the receiver would repeat cyclic (oscillatory) response.
In fact, this cyclic response was avoided in the following way. The receiver gene received inputs not only from its direct mediator (gene in Figure 7) but also from other genes (gene , , and ), such that the expression of the receiver was not driven by the mediator alone. Before the mediator showed adaptation back to its original expression, inputs from other genes showing partial adaptations settled to levels sufficiently different from the original.
An alternative process to avoid the cyclic response, which was often adopted in some evolved networks, is the partial adaptation of the mediator(s). Then the input to the receiver does not come back to the original value, so that the receiver can be settled to show the adaptive behavior. In general, partial adaptive response of a gene can introduce partial adaptation of other genes, and thus, all genes can show adaptive expression, most of which are not perfect but partial. Indeed, never reached , but instead was able to reach at most, while the fitness could be as high as , as seen in Figure 5. These partial adaptive responses ensured the adaptive response of the target. Each receiver showed adaptive behaviors, by appropriate degrees of partial adaptation. The degrees of partial adaptation were determined in a self-consistent manner over all genes to achieve their adaptive response.
Next, we investigated the robustness of adaptive behaviors in the context of noise and mutations. In cooperative networks, most genes showed adaptive responses with plateaus at their peak values. One might expect, then, that for such case, the peak value would not be easily changed by perturbations, implying robustness. However, this was not necessarily true for all conditions; indeed, networks with very large values were not robust to mutations and noise.
To study robustness in the context of mutations, we carried out numerical simulations with increasingly large mutation rates. As shown in the results in Figure 10(a), at mutation rates beyond 0.01, average adaptiveness decreased as increased, while the fitness value after evolution decreased only slightly. Just below this threshold mutation rate, reached a value close to unity and gradually decreased with further decreases in the mutation rate. This suggests existence of an error threshold [25], beyond which a gene expression network with a high value loses robustness to mutations. To confirm this result, we also computed the change in fitness when a single path in the network was removed from the fitted network with a given value. As shown in Figure 11, networks with larger values are less robust to such perturbations. We also carried out numerical simulations with noise in the selection procedure, where networks with lower fitness could leave the offspring with a certain probability. We confirmed that networks with cooperative adaptive responses could evolve when the noise in the selection procedure was sufficiently small, but as this selection noise level was larger, the frequency of cooperative adaptive response decreased.
Since cooperative adaptive response needs a large number of genes, a small mutation rate for the error threshold may be a natural outcome. Such cooperative adaptive response requires mutual cooperative regulation by all other genes. Hence, each gene plays an important role in the regulation of other genes and so their robustness to mutations is weaker. In addition, networks with higher average adaptiveness () are rather rare. We checked the distribution of values in randomly generated networks with large fitness values to study the relationships between fitness and values. Most networks randomly generated showed , and the probability of the appearance of cooperative adaptive networks was too low when sampling over random networks. The distribution of was only slightly shifted towards higher values with an increase in the fitness (Figure S3).
We also examined the robustness of these networks to noise in gene expression dynamics. We added a Gaussian white noise term into eq.(1), with amplitude (i.e., to simulate a Langevin equation), and evolved the network to a high fitness value. As shown in Figure 10(b), increases in the noise level caused decreases in the evolved value to . We also examined the decrease in fitness by adding noise to the evolved networks without noise. The fitness decreased when adding noise (of the level 0.05). Fitness was maximal at around , and the decrease due to noise was larger for networks with values larger or smaller than 0.6 (Figure 11). Networks with moderate cooperativity () showed higher robustness to noise. In summary, such networks with moderate average adaptiveness achieved higher fitness and robustness to mutations or noise.
We also studied the dependence of the cooperative adaptive response on the parameters in our model system. As for the input values, we set and so far. Networks selected with the given fitness under this input condition, indeed could keep cooperative adaptive response for other values of as long as and (Figure S4). We also carried out numerical evolution by using different values in and , and confirmed that as long as , evolved networks achieving large fitness values have large average adaptiveness .
On another front, although the frequency of cooperative adaptive response only slightly depended on the threshold value, , it definitely depended on the parameter , a measure of the sensitivity of gene expression, which corresponds to the Hill coefficient (Figure S5). A high fitness value with cooperative adaptive response (large ) was realized when . Fitness value decreased with the decrease of , while it kept the highest value with . The average adaptiveness remained large for smaller , while it showed a drop as was larger than 20.
Thus, cooperative adaptive response is more important for a system with small values to achieve a high fitness value. In other words, if the on-off expression is sloppier with a smaller Hill coefficient, perfect adaptation for the target gene cannot be achieved with motifs with a small number of genes, but cooperative adaptive response enables perfect adaptation of target gene expression. Cooperative adaptive response overcomes the sloppiness in gene expression dynamics.
On the other hand, the observation that remained at values even after generations of evolution for implies that (highly) cooperative adaptive responses did not evolve there. There are two reasons for this. First, by using the time delay strategy, perfect adaptation of the target gene was realized. Recall that in Figure 3 with , relatively high fitness values were achieved, even with small values, using the time delay strategy for earlier generations, and later, highest fitness values were achieved with increased values. For larger values, however, the fitness value determined by the time delay strategy was close to the highest value, and the gain by the cooperative adaptive strategy was smaller. Second, it was more difficult to show partial adaptations under large values because could hardly take intermediate values between (off) and (on) at the steady state due to the step-function nature of the expression. Hence, partial adaptations with distributed values of final gene expression levels, which are required for the cooperative adaptive response, were more difficult to realize with larger values.
Here, we studied the adaptive response in gene expression dynamics with many degrees of freedom and mutual regulation among genes. We evolved the gene regulatory network using a fitness function for the adaptive response of one target gene. However, we found that evolution led to adaptive expression dynamics in several genes besides the target; this is termed the cooperative adaptive response. Genes that are neither the input nor the target exhibited this adaptive response through mutual regulation.
Furthermore, such networks strengthen the adaptive response of the target gene, even though they are very rare among all possible configurations. However, a sufficient number of genes are required for this cooperative response to occur; small networks with could not evolve at all, but networks with evolved high fitness under all conditions tested. Such singular networks evolved through three steps. First, we obtained an adaptive network utilizing a well-known few-gene network motif, which took advantage of timing differences between target-activating and target-inhibiting signals. Genes that were neither the input nor the target just showed a monotonic change. Next, genes activating the target began to show adaptive responses. At the final stage, genes inhibiting the target also showed adaptive responses. It is worth pointing out that in a network with cooperative responses at this final stage, almost all genes either activate or inhibit the target gene, and the former show adaptive responses with initial up-regulation and subsequent down-regulation, whereas the latter show the opposite adaptive response. Such adaptive responses of most genes are not stipulated by the fitness condition itself, but rather stem from evolution.
As cooperative adaptive responses require a large number of genes, they are not necessarily robust against a large number of mutations. Here, we found that as the mutation rate increased, the number of genes with adaptive responses decreased. Still, for relatively large mutation rates, about half the genes still showed adaptive responses. We also studied the influence of noise and found that networks showing moderate cooperative adaptive responses were able to maintain high fitness values. Conversely, under sufficiently large noise, networks with about half the genes showing adaptive response evolved.
The cooperative adaptive response is more important as the sensitivity in the expression (that corresponds to the Hill coefficient) is lower. In other words, when the on/off expression is sloppier, it is more difficult to achieve higher fitness, i.e., perfect adaptive response, by networks with a few genes. Here, cooperative adaptive response by many genes compensates the sloppiness of each expression dynamics as a collective behavior of many genes, which may be reminiscent of von Neumann's study on reliable computation by unreliable components [26].
As for the adaptive response itself, models with few degrees of freedom have been extensively studied. Ma et al. studied all possible three-node network topologies and found a few minimum adaptive motifs that followed the timing difference observed in the first stage in our simulation [2]. Interestingly, in the cooperative adaptive response observed in the present study, a rather different mechanism was adopted. Indeed, the frequencies of such minimum adaptive motifs decreased as the degree of cooperative response increased, even reaching zero on occasion. Moreover, even when the minimum motifs were included, their behaviors were different from that in isolation; genes that intervene input and output genes showed adaptive responses rather than monotonic responses as in standard motifs. In cooperative adaptive networks, almost all genes except for the input gene showed adaptive responses. Many of them did not show perfect adaptation. Rather, they showed partial adaptation with distributed values of deviation between the initial and final states. Appropriate distribution of such deviations are necessary to achieve adaptive response with mutual activation and inhibition.
Several studies on the responses of cells have shown that many genes exhibit adaptive responses, either initial up-regulation followed by subsequent down-regulation or otherwise, as observed here. Even though the adaptive response is not perfect, many genes (i.e., 50%–70%) show at least partial adaptation, and few genes exhibit monotonic responses or no response [15], [16], [27]. For example, more than half the number of genes in yeast exhibit adaptive responses to several stimuli, as identified by microarray analysis for gene expression patterns [12], [13].
Our study suggested that such responses can generally evolve through gene expression dynamics with mutual regulation of many genes to achieve better adaptive responses of a single gene to environmental changes. It goes without saying that cooperative adaptive responses are robust to replacement of the target gene, because almost all genes already show adaptive responses. We also confirmed that the network can react rapidly to changes in the input gene. Therefore, a network with cooperative responses is advantageous in responding to various types of inputs.
Although our gene expression dynamics and fitness conditions are very simplified relative to the actual biological system, we may expect that the cooperative adaptation dynamics observed in the study can be generalized for systems consisting of a large number of proteins that mutually activate and suppress each other. To confirm this generality, we also simulated models with distributed parameters of the threshold or continuous values of in with , and again found the same cooperative adaptive behavior.
It is often very important and useful to extract motifs with few degrees of freedom from a complicated network by identifying functional roles for such motifs [9], [10]. However, biological networks involve many degrees of freedom. Even if such simple motifs exist, it does not necessarily mean that they function in isolation. Moreover, there may be some other basic mechanism for adaptation inherent in systems with many degrees of freedom. Thus, it is important to study the dynamics and functions of complex networks without decomposing them into motifs with few degrees of freedom. Cooperative adaptive responses are outcomes that emerge only in a system with many degrees of freedom. This may be seen as a kind of cooperative phenomenon, where the adaptive response of one gene relies on that of other genes. Most genes that show up-down adaptive regulation receive positive inputs from up-down adaptive responses and negative inputs from down-up adaptive responses; those that show down-up responses have the opposite interactions. Each gene shows an adaptive response as a result of the adaptive responses of other genes. Thus, adaptive responses are generated in a ‘self-consistent’ manner, through positive and negative adaptive ‘mean-field’ dynamics, generated as a result of the adaptive response of each gene. As discussed in the present model, this self-consistent adaptation is not possible if all genes show perfect adaptation. Instead, most genes show partial adaptation, i.e., final expression levels are not equal to the original levels. Indeed, self-consistent adaptive dynamics over many genes are achieved by suitable distribution of these shifts. Possible condition for such distribution to achieve cooperative adaptive response should be clarified in future, by establishing a proper mean-field analysis. Here, it is interesting to recall that such partial adaptation of gene expression over many components with appropriate distributions is observed in gene expression profiles of yeast Saccharomyces cerevisiae [16] and in recent model simulations [27].
In general, it will be important to explore the cooperative dynamics of a network with many degrees of freedom without decomposing it into functional motifs with few degrees of freedom. According to our numerical study, cooperative behaviors are acquired naturally though the evolutionary process in systems with sufficient degrees of freedom. In a system with many genes, there can be some inherent dynamics that are not reducible to a summation of the dynamics of simple motifs. Living cells involve many degrees of freedom that are not necessarily decomposable, and so the search for cooperative dynamics as explored here will be important.
|
10.1371/journal.pcbi.1000723 | Temperature Control of Fimbriation Circuit Switch in Uropathogenic Escherichia coli: Quantitative Analysis via Automated Model Abstraction | Uropathogenic Escherichia coli (UPEC) represent the predominant cause of urinary tract infections (UTIs). A key UPEC molecular virulence mechanism is type 1 fimbriae, whose expression is controlled by the orientation of an invertible chromosomal DNA element—the fim switch. Temperature has been shown to act as a major regulator of fim switching behavior and is overall an important indicator as well as functional feature of many urologic diseases, including UPEC host-pathogen interaction dynamics. Given this panoptic physiological role of temperature during UTI progression and notable empirical challenges to its direct in vivo studies, in silico modeling of corresponding biochemical and biophysical mechanisms essential to UPEC pathogenicity may significantly aid our understanding of the underlying disease processes. However, rigorous computational analysis of biological systems, such as fim switch temperature control circuit, has hereto presented a notoriously demanding problem due to both the substantial complexity of the gene regulatory networks involved as well as their often characteristically discrete and stochastic dynamics. To address these issues, we have developed an approach that enables automated multiscale abstraction of biological system descriptions based on reaction kinetics. Implemented as a computational tool, this method has allowed us to efficiently analyze the modular organization and behavior of the E. coli fimbriation switch circuit at different temperature settings, thus facilitating new insights into this mode of UPEC molecular virulence regulation. In particular, our results suggest that, with respect to its role in shutting down fimbriae expression, the primary function of FimB recombinase may be to effect a controlled down-regulation (rather than increase) of the ON-to-OFF fim switching rate via temperature-dependent suppression of competing dynamics mediated by recombinase FimE. Our computational analysis further implies that this down-regulation mechanism could be particularly significant inside the host environment, thus potentially contributing further understanding toward the development of novel therapeutic approaches to UPEC-caused UTIs.
| Urinary tract infections (UTIs) represent a major growing threat to global public health. With over 15 million cases a year in the United States alone, UTIs are characterized by very high recurrence/reinfection rates, particularly among women and minority groups [1]. The predominant cause of UTIs is uropathogenic Escherichia coli (UPEC) bacteria, whose wide-spread and increasing antibiotic-resistance has made the development of alternative anti-UPEC treatments progressively more important and urgent. UPEC's foremost virulence factor is hair-like surface structures called type 1 fimbriae. Thus, one such potentially promising therapeutic approach may be to manipulate bacteria's own cellular circuitry toward inducing UPEC to turn off their fimbriae expression—rendering individual microbes benign. This task requires detailed understanding of molecular mechanisms involved, which may be significantly aided by in silico modeling. However, for UPEC fimbriation control circuit and many other systems, low-level all-inclusive quantitative models inevitably become too computationally demanding to remain practical, while high-level qualitative representations frequently prove inadequate owing to the substantial organizational and behavioral complexity of biological processes involved.
We have developed an automated multiscale model abstraction methodology that helps address these problems by systematically generating intermediate-level representations that rigorously balance computational efficiency and modeling accuracy. Here, we use such an approach to examine how different temperature settings quantitatively affect UPEC transitions between fimbriate and afimbriate phases, to gain new understanding of the underlying modular circuit switch control logic, and to suggest further insights into ways this knowledge could potentially be used in therapeutic applications.
| Type 1 fimbriae (pili) represent the foremost virulence factor in lower urinary tract infections (UTIs) by uropathogenic Escherichia coli (UPEC)—the main causative agent that accounts for 80–90 percent of all community-acquired UTIs in the United States [1]–[4]. These adhesive surface organelles have been identified as both the UPEC virulence factor most frequently found in clinical isolates as well as the one that experiences the highest absolute and among the greatest relative increases of component gene expression in vivo during UTIs [5],[6]. Type 1 fimbriae also have been shown to fulfill molecular Koch's postulates [2],[7] and have been further reported as the only major uropathogenic virulence factor that is broadly significant for enteric E. coli strains as well [8],[9]. The hair-like structures involved vary from a few fractions of a micrometer to more than 3 m in length and consist of 7nm-thick right-handed helical rods—largely made up of repeating subunits—with 3nm-wide tips containing the adhesin, which can bind to D-mannose-containing residues found on the surface of epithelial cells and mediate their invasion by UPEC [10]–[13]. Type 1 fimbriae are further thought to aid the UPEC infection process by enhancing the ability of bacteria to form biofilms and to develop intracellular bacterial communities (IBCs) with biofilm-like properties [13]–[18]. The latter allow E. coli to establish quiescent pathogen reservoirs shielded from native host defenses and antibiotic treatments as well as serve to seed subsequent UTIs in a type 1 fimbriae-dependent manner [2], [13], [19]–[21]. This may both contribute to the widespread emergence of multi-drug-resistant UPEC strains (up to 20–50 percent of isolates) as well as help account for the notably high rates of UTI incidence (lifetime risk of over 50 percent for women and nearly 14 percent for men) and recurrence (40 percent in women and 26 percent in men per annum) – along with leading to a number of other significant public health implications (e.g., over 10 million estimated annual physician office visits in the United States alone) [1],[22]. However, while they provide a means for infection, type 1-fimbriated UPEC populations also have lower fitness due to phase-specific mechanisms that directly decrease growth rates through additional costs of fimbriae synthesis and contact-dependent inhibition as well as reduce motility, which allows competitors to more efficiently occupy advantageous colonization sites and take up resources [6], [23]–[25]. Furthermore, type 1 fimbriae-mediated attachment can lead to preferential exfoliation of infected cells as part of the host immune response, which can result in rapid clearance of the infection [13], [20], [26]–[28]. Among other things, this apparent dichotomy between the essential role played by the piliated phase in the establishment of the infection and the noted fitness disadvantages conferred upon individual bacteria by type 1 fimbriae implies that their expression needs to be highly optimized and tightly controlled.
As illustrated in Figure 1, the expression of type 1 fimbriae in E. coli is randomly phase variable, whereby individual cells stochastically switch between fimbriate (ON) and afimbriate (OFF) states with rates regulated by various internal as well as environmental conditions [29]–[33]. With the ongoing advancements in high-resolution single-cell and single-molecule scale experimental methods, such bimodal and bistable mechanisms for generating phenotypic heterogeneity in clonal cell populations have been increasingly often identified and investigated across a broad range of prokaryotic and eukaryotic systems—where they have been shown to influence a diverse spectrum of processes—including organism development, behavior, disease, survival, and memory [34]–[44]. In the case of E. coli type 1 fimbriae, this phase variation is controlled by the fim circuit switch that functions based on the inversion of a 314bp chromosomal region, fimS, bounded by two 9bp inverted repeats left and right (IRL and IRR) [29],[34],[45],[46]. The fimS element contains the promoter for fimA and other genes encoding structural subunits of type 1 fimbriae. As a result, an individual E. coli cell expresses type 1 fimbriae when the fim switch is in the ON position and rapidly becomes afimbriate when the switch flips into the OFF position [34],[47]. This inversion of fimS requires either or site-specific recombinases binding at IRL and IRR [29],[47],[48]. However, whereas mediates recombination with little orientational bias, mediates recombination predominantly in the ON-to-OFF direction [30],[49]. Empirical evidence has further revealed that the inversion of the fim switch is strongly controlled by temperature in a complex manner [30],[31]. In particular, observations at , , and have indicated that wild-type ON-to-OFF switching frequency—dominated by —decreases in an exponential-like fashion as temperature increases, while -mediated switching frequency is higher at than either at or in both defined-rich and minimal media. Experimental results also show that the wild-type ON-to-OFF switching rate is much faster than -mediated switching rate alone, allowing E. coli to rapidly undergo afimbriation under appropriate conditions [30],[50].
This work investigates the logic and behavior of the gene regulatory circuit, which controls the ON/OFF switching of type 1 fimbriae expression, by starting with the reaction-level description of its underlying biochemical and biophysical molecular interaction mechanisms. We are particularly interested in the role of environmental cues in this process and, specifically, of temperature as it is known to control many gene regulatory circuits in bacteria—often those responsible for virulence functions [51]. Temperature variations are also frequently characteristic of host-pathogen interaction dynamics—such as during cytokine response (e.g., through IL-6 as well as IL-8 and IL-1) and the ensuing inflammation that is indicative of the onset and progression of UPEC UTIs—as well as often generally representative of urinary tract pathology [52],[53]. In this context, reaction-level modeling provides a framework for highly accurate description of the underlying biomolecular circuit behavior through application of the corresponding fundamental chemical and physical principles. However, the innate complexity of biological networks involved as well as the key role played by nonlinear, discrete, and stochastic kinetics in regulating the dynamics of cellular pathways driven by molecular-scale mechanisms result in profound computational challenges to their accurate quantitative analysis. The problem becomes particularly acute when dealing with biological systems, such as type 1 fimbriation circuit switch dynamics in UPEC, whose behavior is driven by internal or external discrete-stochastic processes to exhibit qualitative deviations from what might otherwise be expected on the bases of “classical” continuous-deterministic biochemical modeling via mass-action kinetics and reaction rate differential equations [39],[54]. The resulting “deviant” dynamics lead such biological systems to behave in a distinctive but often quite unintuitive manner, which necessitates the use of differential-difference modeling based on the chemical master equation framework (see [54]–[59] and Methods for details). However, while the latter approach is able to accurately account for both the stochastic occurrence as well as the discrete nature of individual molecular interactions that underlie the design, function, and control of most biological circuits—it also tends to produce dramatic increases in the associated analytical and computational demands [60]–[62].
Although these computational limitations may often render any direct implementations of the all-inclusive low-level quantitative models impractical, the use of entirely high-level qualitative representations frequently becomes inadequate as well, owing to the substantial multiscale dynamical and functional complexity that biological systems can manifest. In such cases, in silico analysis can greatly benefit from applications of appropriate intermediate-level system model abstractions—whereby multiple individual biological interactions are aggregated into significantly few(er), but quantitatively analogous functional processes. An optimized model abstraction scheme then looks to accurately capture the target characteristics of biological system behavior, while trading off some tightly controlled degree of precision for significant computational gains. Additionally, the resulting abstracted model of the system may also be useful in helping to uncover any general high-level logical patterns embedded within the biological networks involved, which can otherwise be obscured by the low-level molecular interaction mechanics.
Our method initiates the abstraction procedure with a detailed reaction-level representation of biological processes in question. This enables it to utilize basic biochemical and biophysical principles to rigorously implement many of the existing as well as potentially allow for the development and incorporation of novel abstraction techniques, Table 1, in order to insure the desired degree of modeling accuracy versus computational efficiency for the abstracted representation at the system scale of interest [63],[64].
However, such an approach to model complexity reduction could also lead to a further problem: while most abstractions used in the analysis of biomolecular networks have traditionally been implemented manually and on the mechanism-by-mechanism basis, doing so accurately in a general biological systems setting becomes tedious and time-consuming. The resulting model translation and transformation errors also tend to increase when progressively more intricate organism-scale physiological processes—from cell differentiation and tissue development to cancer, infection, host-pathogen interaction dynamics, etc.—are considered.
The strategy used here is able to substantially overcome these issues by automating the abstraction process via a set of algorithms developed for and implemented in the reb2sac computational tool [63],[64]. Its application has allowed us to generate abstracted representations of detailed reaction-level biological mechanisms—including genetic regulatory networks—which yield results in close correspondence with those obtained by using the underlying low-level models, while also significantly accelerating the required computations and often putting them on par with those of high-level descriptions. For instance, we were previously able to validate the overall robustness and utility of such an automated abstraction approach to biological systems analysis by using it to investigate the lysis/lysogeny developmental decision pathway in E. coli phage [63],[64]. The ensuing abstracted model analysis not only yields results that substantially (and in significantly less time) reproduce those elicited through the examination of the detailed system description reported earlier [65], but is further able to quantitatively investigate and similarly match experimental observations of system properties exhibited under environmental conditions that have been previously shown to cause the detailed model analysis to become so computationally demanding as to make it essentially infeasible [63],[65].
Here, we use such computational analysis aided by automated model abstraction to examine the behavior of the basic genetic regulatory network responsible for the ON/OFF switching of type 1 fimbriae expression in uropathogenic E. coli, Figure 2. We specifically focus on how different temperature settings quantitatively modulate the random switching of the UPEC fimbriation circuit into the transcriptionally silent fim mode through the corresponding ON-to-OFF inversion of fimS. Notably, while the behavior of most molecular processes depends on temperature, in this system global regulatory proteins and play a particularly important role in controlling switch inversion rates not only by directly effecting its internal molecular dynamics, but also by acting as sensors of certain environmental conditions that the fim circuit is subjected to in the physiological range—including those of a host. For instance, acts in a temperature-dependent manner when it binds to DNA regions containing fimB / fimE promoters and represses their expression [31],[66]. Additionally, binds to three sites, which affects switching rates [50],[67],[68]. Since downregulates the expression of lrp [69],[70], also behaves in an effectively temperature-dependent manner. Finally, it has been shown that binds to / regulatory sites and is required for any observable phase variation, in part by playing a structural role in fim switching via its ability to introduce sharp bends into the target DNA [47],[71]. The resulting molecular interactions that involve , , , as well as the fimS DNA element and associated regulatory sites are what largely serves to kinetically effect the ON-to-OFF fim switch circuit dynamics. As the latter physiologically initiates the transition of an individual bacterium from the virulent fimbriate to the largely benign afimbriate phase and given the wide-spread emergence of antibiotic-resistant UPEC, a better understanding of such processes could benefit the development of novel clinical UPEC UTI therapies by, among other things, providing deeper insights into mechanisms potentially able to medically abrogate UPEC virulence by exploiting its internal molecular circuitry responsible for regulating the state of fimS in order to inhibit type 1 fimbriae expression.
Towards this end, the paper begins by considering a detailed reaction-level discrete and stochastic description of the biological molecular network controlling the fim switch. As discussed earlier, we then abstract this detailed representation by utilizing reb2sac, which enables us to successfully circumvent the otherwise significant computational issues involved. The accuracy of our abstracted model analysis with respect to the target system property—i.e., the temperature dependence of the fim switch turn-off rate—is further validated by comparing its results with those computed via the unabstracted detailed model as well as with those derived from empirical observations (where available). This, in turn, serves to explicitly demonstrate how automated model abstractions can be used to help substantially improve the speed and efficiency of biological molecular systems analysis, while also maintaining precision and improving interpretability of results. For instance, the abstracted representation has allowed us to better understand the general circuit-level organization of the regulatory logic behind the UPEC fimbriation switch and to identify the two key subnetworks— recombinase regulation and fim switch configuration—involved in its engineering design. Our conclusions also confirm that temperature has a major and non-trivial role in determining ON/OFF switching of fimbriae expression as well as suggest new insights into the role of in this process and offer novel clues toward its potential translational applications in the host environment. In particular, our results indicate that—when the control circuit behavior is analyzed quantitatively across different temperatures—the primary role of recombinase may not be to increase the total ON-to-OFF switching rate, but rather to reduce it by down-regulating the rate of switching mediated by the competing recombinase . That is, down-regulation of not only reduces the OFF-to-ON switching, but also serves to increase the ON-to-OFF rate in a temperature-sensitive manner, which indicates that this mechanism may provide a powerful regulatory tool for suppressing the fimbriate UPEC phase. Finally, as our analysis implies that the described effect is strongest and the switching rate is most sensitive to the corresponding mode of control in the physiological temperature range of the host environment, it may serve to potentially help identify new biomedical targets in the UPEC molecular virulence circuitry.
Based on the regulatory network diagrammed in Figure 2, we have developed a molecular kinetic reaction-level description of E. coli fimbriation switch system, which has resulted in a detailed model of the fim circuit that comprises 52 reactions and 31 species (Figures 3 and 4). This model is then used to, among other things, quantitatively analyze the effects of temperature on both the total and -mediated ON-to-OFF fim switching probabilities over one cell generation. In particular, starting with the switch in the ON position at various temperature settings—i.e., , , and —where the corresponding empirical observations were available (see Methods and Text S1), the detailed model was simulated 100,000 times by using our implementation of Gillespie's Stochastic Simulation Algorithm (SSA). The ensuing switching behavior of the fim circuit was found to be both qualitatively and quantitatively consistent with that obtained via empirical observations [30] (see Table 2). However, computational demands presented by these detailed model simulations were significant, requiring over 30 hours on a 3GHz Pentium 4 with 1GB of memory (Table 3).
After applying reb2sac automatic abstraction engine with the switch state as the target quantity of interest, the detailed model is transformed into an abstracted model with 10 reactions and 3 species (, , and a conglomerate non-linear stochastic switch – see Figures 5 and 6 as well as Methods for further detail). In order to compare the abstracted and detailed models, we have performed numerical simulations to compute the wild-type and -mediated ON-to-OFF switching probabilities for one cell generation in minimal medium using the same simulator. The results of the abstracted analysis are found to be in close agreement with those obtained using the detailed model and substantially match the empirical observations (see Table 2). However, computational gains from the model abstraction are significant. The abstracted model simulation of 100,000 runs takes less than 2 hours on a 3GHz Pentium 4 with 1GB of memory, which is a speed-up of about 16 times compared with the runtime of detailed model simulations (Table 3).
In addition to allowing for accurate kinetic simulation of circuit-level dynamics, the reaction-level description of biological networks is often useful in helping to reveal their broader structural and functional features, including the innate modular architecture of E. coli fimbriation switch design considered here. Specifically, graph-level analysis carried out as part of the detailed model abstraction process has naturally led us to separate out and identify its two major constitutive subnetworks. These turn out to correspond to the two principal functional units of the fim switch circuit: the module effecting production-degradation of and ; and the module responsible for the configuration dynamics of the fimS element itself (e.g., Figures 5 and 6). Such a view of the internal fim switch circuit organization both makes its logic easier and more intuitive to understand as well as simplifies and provides further basis that serves to facilitate subsequent steps involved in the model abstraction process.
By systematically refining our understanding of the underlying organization logic and improving required computational times, our approach further enhances the ability of in silico analysis to accurately explore various environmental as well as internal conditions and parameter regions of biological systems. This may be particularly useful when certain settings can be deemed physiologically important, yet are not easily amenable to or simply do not presently have sufficient number of experimental measurements available; and which lead to dynamics that are too complex or involve species too numerous to be productively investigated directly at the detailed molecular interaction network level. For example, in the case of the fimS inversion control circuit, probabilities of ON-to-OFF switching at various temperature points (including those outside of the experimental range) can be effectively and efficiently estimated by using the described model abstraction methods. Here, Figure 7 shows both wild-type and -only mediated ON-to-OFF switching probabilities computed via the abstracted fim switch model at – respectively – 7 and 15 additional temperature points, where experimental data are not available (also see Table 2).
Notably, these results not only reaffirm earlier coarser-grained empirical observations of wild-type and -only mediated ON-to-OFF fim circuit switching frequency dependence on temperature [23],[30], but also offer the finer-grained resolution capable—as discussed below in more detail—of providing further insights into this relationship. In particular, while our analysis supports the prior suggestion that the wild-type fim ON-to-OFF rate is overall a decreasing function of temperature that varies by nearly two orders of magnitude in the physiological range, it also appears to indicate that this dependence has a supra-exponential component as well, Figure 7A. Furthermore, when the abstracted model is used to increase the resolution of FimB-mediated switching frequency dependence on temperature, it shows that UPEC may have evolved toward a tightly optimized type 1 fimbriae virulence factor expression control that is designed to sense and differentially respond based on whether the host temperature is within the normal physiological range of or if it is elevated/lowered instead. Whereas the circuit -mediated ON-to-OFF rate appears to be maintained at a relatively elevated but stable level across the entire normal temperature range—it looks to be significantly suppressed immediately outside of this characteristic band, Figure 7B, which may have notable implications for the persistence of the pathogenic UPEC phase and ensuing UTIs (see Discussion).
Since the -mediated switching probability can be orders of magnitude smaller than the wild type ON-to-OFF switching probability (Table 2), the effect of on the temperature control of the fimbriation circuit shutdown rate may also appear minimal. It is, furthermore, not immediately clear why -mediated switching needs to be exquisitely bidirectional rather than simply OFF-to-ON, given that essentially only promotes ON-to-OFF switching and completely dominates the rate in this direction. While various theories have been proposed to explain this feature of the fimbriation regulatory network design (see Discussion), we wanted to generate a quantitative hypothesis regarding the role of in the temperature control of the fim ON-to-OFF circuit switching by using computational analysis methods to perturb the underlying molecular interaction-level network properties and to then explore the behavior of any resulting fimbriation mutants. To do this, we have modified the original fim switch inversion system in silico and generated several detailed mutant models—two of which proved to be of particular interest. One represents a mutant, where fimB has been placed under the control of a strong promoter that leads to overproduction by a factor of two relative to wild-type. The other describes a mutant, such as a knockout or an amino acid substitution, where protein has been rendered nonfunctional in the present context by losing its ON-to-OFF switch-mediating activity. Both mutant models were abstracted using reb2sac and simulated.
Comparing the elucidated mutant and wild-type behaviors at the same 10 temperature points considered earlier (e.g., Figure 7A) now allows us to quantitatively characterize the dependence of this fim switch circuit temperature control on the level of activity in the cell. As illustrated in Figure 8A, the total ON-to-OFF switching probability generally decreases inversely with levels across all temperatures. That is, in the physiological range, the total ON-to-OFF switching probabilities in the fimB− mutant are higher than those in the wild-type, which are—in turn—higher than those in the mutant where is overexpressed. Notably, this not only suggests that the -mediated shutdown of fimbriae expression is efficiently down-regulated by , but that—as shown in Figure 8B—this effect is strongest in the to temperature range, where the total ON-to-OFF switching probability of the fimB− mutant can be over two times higher than that of the wild-type and nearly three times that of the overexpressing mutant. Physiologically, this implies that the presence of at normal or elevated levels greatly enhances the persistence of type 1-fimbriated UPEC phase. Thus, although the -mediated fim switching probability is itself at least an order of magnitude lower than wild-type, may have a key role in regulating and enhancing the control of temperature-dependent functions in the E. coli fim switch circuit by—among other things—also reducing the effect of -mediated ON-to-OFF fim switching. This serves to regulate the type 1 fimbriae-based molecular virulence mechanism and, potentially, may help increase the life-time of the pathogenic fimbriate UPEC phase. The latter result is of particular interest because the effect appears to be most pronounced in the temperature range that corresponds to the intra-host bladder environment, opening up the possibility that it may be directly relevant to UPEC-caused UTIs.
In recent years, rapid advances of experimental biology made it practical to study both molecular- and network-scale organization of many biological and physiological processes in much greater detail than was previously feasible. This, in turn, has made computational analysis not only possible, but also essential to any efforts aimed at understanding the increasingly intricate structures and functions of multiscale biological systems that are being uncovered through empirical means. However, this growing wealth of knowledge about in situ biological processes has also led to the demand for progressively more sophisticated in silico system models. As a result, although accurate molecular-scale biochemical descriptions could be formulated for a large number of experimentally observed systems, their complexity is rapidly exceeding our present as well as near-future computational capabilities—the issue that has become more pronounced with the emerging understanding of the ubiquitous role played by nonlinear and discrete-stochastic (“noisy”) molecular dynamics in gene regulatory, signal transduction, and other biological systems [39]. That is, while their role may often be essential in defining the various design and functional characteristics of biomolecular circuits [72]–[78]—including temperature controls [79]–[82]—the resulting introduction of multiplicative noise and the possibility of ensuing deviant effects [54], [83]–[89] can make computational analysis of such processes particularly demanding [62].
Going forward, these considerations appear to suggest that “model abstractions”—whereby, for instance, multiple biological network interactions comprising individual biomolecular mechanisms are rigorously and systematically aggregated into a few easily tractable, but functionally analogous components—will continue to become an increasingly useful tool in the general context of computational and systems biology. Importantly, model abstractions can serve not only to substantially reduce the computational requirements associated with the analysis of specific multiscale biological processes, but may also lead to identification of functional units that correspond to biologically meaningful modules or motifs (exemplified here by the two functional subnetworks of the fim switch circuit). The latter helps contribute additional insights into the underlying system organization and physiology as well as make their often intricate logic easier to understand.
Yet, given this growing scope and complexity of systems biological models, manual implementation of comprehensive abstractions with accuracy and efficiency becomes a challenge—creating the need for process automation. This work has demonstrated the utility of such an automated model abstraction approach by applying it to the investigation of the role of temperature in controlling the ON/OFF switch state of the fim genetic regulatory circuit that determines the expression of type 1 fimbriae (Figure 1), which is an essential virulence factor in uropathogenic E. coli—the leading cause of urinary tract infections and a major growing public health problem [1]. Insights into this fimbriation process—and, particularly, into the mechanisms that control its shutdown—may be especially useful as the widespread proliferation of antibiotic-resistant and biofilm-forming UPEC strains continues to increase the demands for novel treatment methods. In particular, a thorough understanding of their cellular network function under a range of conditions may allow us to manipulate UPEC's internal molecular virulence circuitry through external means, thus potentially opening up new approaches to modulating their pathogenicity. One such key external regulator is temperature, which not only often acts as an indicator of UTI progression and impacts its course, but may also be amenable to meaningful control in clinical settings. Furthermore, as experimental investigation of these processes in situ may offer a variety of practical challenges, in silico approaches could be very useful in helping to identify how internal molecular virulence machinery is influenced by external temperature variations. However, even in the case of the relatively small biological circuit controlling type 1 UPEC fimbriation switch considered here (Figure 2), its functions are qualitatively affected by the inherently discrete and stochastic as well as the largely nonlinear nature of the underlying biomolecular mechanisms. This necessitates the type of biological systems analysis that is capable of accurately accounting for contributions of molecular-scale reaction-level processes, which typically makes direct in silico studies of such systems highly taxing and investigations of detailed fimbriation circuit switch properties challenging. Here, we were able to substantially circumvent such issues through the use of systematic model abstractions, which allowed us to convert a highly computationally demanding problem of fim circuit switch response to temperature variations into a relatively accessible one by relying upon the automated model abstraction methodology we have developed and implemented in the reb2sac model abstraction tool [63]. We then used this abstracted model to gain deeper insights into the dynamics of this biomedically important system, including the role of in controlling the expression shutdown rates of type 1 fimbriae virulence factor.
To do this, we have first constructed a molecular-scale reaction-based “detailed” model of the regulatory network that controls the orientation of fimS genomic element (Figure 2), which is responsible for ON/OFF switching of type 1 fimbriae expression. This model has allowed us to analyze—with high degree of fidelity, albeit at significant computational costs—the dynamic behavior of UPEC's discrete-stochastic genomic fimbriation circuit, including the ensuing effects of temperature on the wild-type and -mediated ON-to-OFF switching probabilities in minimal medium, which are shown to be quantitatively consistent with those observed empirically (Table 2). We then applied our reb2sac tool to the detailed model of the fim switch circuit. The resulting “abstracted” model substantially reduces the complexity of the problem, enabling us to significantly increase the throughput of our in silico analysis (Table 3), while still maintaining accuracy when compared with the detailed model predictions and available experimental observations (Table 2). This approach has further allowed us to compute the ON-to-OFF switching probabilities at additional temperature points and to investigate the behaviors of characteristic mutants in silico (Figures 7 and 8).
As a result, we have been able to gain a number of insights into the internal dynamics of this clinically relevant system, including into the strong temperature dependence of putative UPEC afimbriation switching rates (e.g., Figure 7), which characterize the intrinsic dynamics that may cause individual bacteria to autonomously transition from pathogenic to benign phase. In particular, while earlier theoretical studies [90],[91] have discussed how the type 1-fimbriation level is regulated by the two recombinases, it has not been entirely clear what role (if any) has in turning off the fim switch, since the ON-to-OFF rate it mediates is at least an order of magnitude lower than that enabled by . This may also seem at odds with the evolutionary selection of the remarkably fair ON/OFF switching probabilities observed. Our analysis (which—it should be emphasized—though based on primary empirical data, is done substantially in silico and so needs further experimental validation) has been able to suggest a possible explanation for this ostensible contradiction by identifying a potentially key regulatory role of in directing UPEC afimbriation. Specifically, while the switching rate it can mediate directly remains low, may competitively modulate the dominant -dependent switching process in excess of three-fold—thus serving to significantly lower wild-type E. coli ON-to-OFF switching rates in the host environment. This process can help to further prolong or abridge the persistence of the fimbriate phase in individual bacteria, which may be crucial for UPEC survival when colonizing bladder and invading urothelium, while trying to escape immune system responses and effects of antibiotic treatments, Figure 8. Furthermore, this -based regulation mechanism may be more robust against small perturbations in level than a simpler fim switch inversion control, which could be of importance in a highly variable and often rapidly fluctuating environment of the urinary tract.
While the extent to which these innate mechanisms are able to curtail or enhance virulence of UPEC in situ could be affected by the various aspects of complex host-pathogen interactions noted previously, it may be worth considering that to date much of the discussion has been framed in the context of such immune response processes as cytokine production, resulting inflammation, and potential subsequent exfoliation of infected bladder epithelial cells that generally lead to the increase in local tissue temperature [27],[52],[92],[93]. However, our results support a further understanding of UPEC adaptation to this aspect of host immune response. Although -mediated fimbriae expression shutdown rate appears elevated but largely insensitive to temperature in the normal range of a host, as temperature increases further—both and ON-to-OFF switching rates are lowered, while E. coli's ability to control this process through variations in becomes optimized. That is, as UTI triggers the onset of an inflammatory response, the resulting increase in temperature tends not only to lock this UPEC control circuit in the pathogenic fimbriate phase, but also to transiently maximize switch sensitivity towards regulation by at several degrees above normal—a range consistent with the corresponding host environment. The potential existence of such sensitized “pathogenic phase lock” (PPL) mechanism and its ensuing effects on UPEC virulence could have direct bearing on some of the clinical challenges in treating UTIs discussed earlier, since many of these characteristics are thought to be associated with type 1 fimbriae-dependent biofilm and IBC formation [15],[16]. The latter structures have been shown to provide persistent pathogen reservoirs in bladder tissue and/or on abiotic surfaces (e.g., those of medical implants, such as catheters) even in cases when antibiotic treatments can effectively sterilize urine [92]. Still, currently recommended treatment strategies include ongoing prophylactic daily or weekly antibiotic therapy in cases of recurrent UTIs (defined as more than 2 episodes in 12 months), even though studies have shown no long-term reduction of UTI recurrence in such patients after prophylaxis cessation as compared with those in placebo groups [94]. Given further risks of various potential side effects—which can range from moderate to severe—and development of drug resistance as well as a number of other undesirable consequences, including growing epidemiological and public health implications [1],[21],[94], presently available basic antibiotics-based UTI treatment strategies cannot be considered satisfactory. In fact, it has been strongly suggested that from a clinical perspective the use of traditional antibiotic therapies cannot be successful against biofilm/IBC-forming bacteria and that other treatment modes, particularly those that target biofilm/IBC/fimbriation-specific processes, need to be developed [95],[96]. Thus, inference of type 1 fimbriae expression regulation circuit logic and elucidation of external intervention strategies able to influence or interfere with its internal dynamics, including via mechanisms that utilize controlled temperature variation to induce PPL relief and subsequent fim switch shutdown as discussed here, could offer promising potential for contributing further understanding towards the development of novel remedial approaches.
Historically, many such original medicinal and other therapeutic methods have had their genesis in traditional or domestic practices [97]—a pattern that has been recently seen to accelerate because of, among other things, growing synergies between Western and Asian medical systems that have already resulted in such notable pharmacological and synthetic biological successes as ephedrine and artemisinin—with more on the way [98],[99]. For instance, while a relatively prolonged exposure to cold has been generally associated with the increased incidence of UTIs [100],[101], a number of complementary therapies have been based around the practice of keeping genitourinary tract area cool or even briefly exposing it to low temperatures as beneficial for the prevention and treatment of various pathological processes, including microbial infections [102],[103]. Yet, while the ongoing research into the effects of cold exposure on differential activation/repression of various adaptive and innate immune system components has now begun to suggest underlying cellular and molecular biological basis for these phenomena observed in clinical applications, their underlying modes of action on the whole remain poorly understood [104],[105]. In this context, the results discussed here provide an example of the quantitative insight that multiscale reaction-based computational modeling brings to such complex processes. Specifically, based on the implications of our study for utilizing alternative temperature-driven approaches in targeting the dependence of UPEC virulence mechanisms on type 1 fimbriae expression—rather than relying solely on antibiotic or other biochemical means—two mechanisms may merit further attention. On the one hand, as host response to UTI includes tissue inflammation and a corresponding local or systemic increase in temperature, our analysis indicates that the adaptive feedback strategy evolved by UPEC tends to bring about PPL conditions, whereby ON-to-OFF type 1 fimbriation circuit switch may become maximally sensitized to . Combined with its central role in mediating the OFF-to-ON switching [47], this implies that lowering activity may lead to a reciprocal decrease in the fraction of virulent fimbriate UPEC phase and subsequent reduction in the associated pathogen load—making the corresponding persistent UTIs more amenable to host immune mechanisms and, potentially, increasing the efficacy of existing medical treatments. However, given the challenges of developing and delivering the required inhibitors as well as further obstacles presented by IBC formation inside epithelial cells, it may not be immediately clear how direct variation of UPEC activity could be meaningfully achieved in vivo. On the other hand, our conclusions also support the notion that decreasing the temperature of UPEC environment may increase shutdown rates of type 1 fimbriation circuit switch (including by indirectly lowering ), thus potentially leading to the up-regulation of afimbriation rates in individual bacteria. This would tend to suppress UPEC pathogenicity by reducing their capability for attaching to and invading urothelial cells as well as by interfering with biofilm/IBC formation and maintenance, which may be expected to decrease their capacity for subsequent re-infection. As in this case only local temperature variations—including those directed by cool/warm intravesical media or such catheter and other device instillation—are principally required in order to elicit the indicated physiological response, the conditions necessary to influence UPEC fimbriation switching in this manner may be practically attainable in biomedical and clinical applications.
It is important to note, however, that this merely suggests the possibility and does not engender any further assessment of potential efficacy such therapies may have in clinical UTI settings. The latter requires a more extensive follow on investigation—particularly in view of additional host-pathogen interaction dynamics, the multicellular nature of the system and commensurably greater complexity of intra-/inter-cellular networks it comprises, the epidemiology of autoinfection processes involved in promoting UTIs from and diversity of the endogenous bacterial flora, etc. as well as any associated difficulties in developing detailed models of the intra-host pathogen environment. Such challenges are often due to our understanding of biomolecular functions involved being insufficiently detailed and/or tissue-specific processes adding further layers of complexity to the overall infection dynamics. For instance, while this work has been able to use modeling and computational analysis in order to explore certain aspects of type 1 fimbriae switch control, the latter are primarily relevant to lower urinary tract infections. In contrast, upper UTIs are predominantly promulgated by P fimbriae—a distinct UPEC adhesive factor, which is regulated by significantly different biomolecular circuitry (see [106],[107] for detailed modeling of the corresponding pap switch) that leads to its own mode of thermoregulation [108]. Still, recent experimental results—from those cited earlier with respect to UPEC and host immune system, to the discovery of TRP channel family of cold and hot sensors in human genitourinary tract [109]—have provided strong evidence that temperature and its variations can have major systemic influence on healthy functions as well as various pathological developments in the urinary tract and surrounding tissues. In fact, basic intravesical cooling or warming with media of desired temperature or via chemical agonists, such as menthol/icilin or capsaicin/resiniferatoxin – respectively, has had a long history of being used to induce nerve desensitization, bladder cooling reflex, and other physiological mechanisms in therapeutic applications ranging from treating patients with detrusor overactivity, bladder pain, and urothelium irritation to diagnosing various urinary tract and neurologic disorders [109]–[111]. This not only directly indicates that patient urinary tract temperature could be practically and therapeutically manipulated in clinical applications, but—as TRP sensors appear specific to animals and fungi [112]—also suggests that thermal regulation of human physiological response processes may be actively effected in a manner that by-and-large does not directly impinge upon prokaryotic pathogens. Conversely, with better empirical understanding and computational modeling of the underlying biological circuits, the same mechanism may allow us to substantively offset the effect on the host of moderate temperature changes by applying compensatory chemical stimuli to appropriate TRP channels and modulating their ensuing activity up to desensitization. This, in turn, opens up the possibility that externally controlled temperature variations may be guided by quantitative systems analysis to specifically target and manipulate the internal dynamics of bacterial or other pathogenic processes in sutu, causing them to either become innately less virulent—for example, as has been discussed here in the context of UPEC fimbriation circuit switching—or making them more susceptible to the immune response as well as antibiotic and other treatments, thus potentially contributing to the ongoing enhancement of existing and the development of novel therapeutic applications.
Taken together, these results broadly serve to further demonstrate the potential utility of computational and systems biological approaches as we are beginning to understand and control many physiological processes in disease and development at the inter-/intra-cellular network and circuit levels [113]–[118], thus enabling greater insights and providing more effective solutions to associated clinical and public health problems. They also highlight the benefits of model abstractions and the need for process automation as tools of in silico biological systems analysis, including their ability to significantly increase the efficiency with which practical multiscale biomolecular and biomedical problems may be addressed in situ. In fact—while this may be directly noted by considering just how much longer it takes to simulate a detailed network model, or how tedious a manual implementation of all constitutive abstractions can be, or significant simplifications in functional logic the corresponding process modularization may be able to achieve—what ultimately makes the automated model abstraction approach compelling is the eventual consideration of how relatively simple the E. coli type 1 fimbriation switch circuit and its temperature controls appear to be as compared to the complexity of many other biological and biomedical processes we may be expected to face in the context of systems and computational biology now or in the near future.
Previous works by Wolf & Arkin, Blomfeld et al., and others have helped elucidate and ascertain the importance of discrete and stochastic mechanisms in the fim system dynamics [23],[30],[45],[47],[71],[90],[91]. For example, it has been shown that fimS inversions are digital (ON/OFF) events that are randomly promoted by or binding to discrete IRL/IRR sites and regulated by the corresponding or occupancies of cis-regulatory genomic elements, which are present in low integer counts. Under these conditions, biomolecular systems can manifest emergent and unintuitive behaviors that may greatly deviate from the predictions of macroscopic continuous and deterministic classical chemical kinetics (CCK – also referred to as reaction rate equations or mass-action kinetics) [54]. Therefore, accurate analysis of the fim switch circuit requires the use of a mesoscopic discrete and stochastic process description based on the chemical master equation (CME) [54],[56],[58],[59],[119],[120].
This approach considers the behavior of biomolecular systems at the individual reaction level by exactly tracking the time-evolution of the discrete number probability distribution for all molecular species present in the system and by correspondingly treating each reaction as a separate random event. An intuitive basis for the (forward) CME can be described as follows: given species at time with the number of molecules each, which are interacting through irreversible chemical reactions with stoichiometric vectors inside a well-stirred tank of constant volume and in thermal equilibrium at constant temperature—the probability that this system is found in the molecular number state at time can be simply expressed as the sum of probabilities that: (i) the system is in the same state at time and does not undergo any transitions; and (ii) the probability that it is in a different state at time and transitions into during . Then, under the Markovian assumption:(1)with at and —the probability that during the system in state undergoes reaction —where is called the propensity function and it is further assumed that is chosen small-enough that almost surely only one reaction occurs during this time increment.
Taking the limit and rearranging Equation 1 gives the expression describing the temporal evolution of :(2)which is the CME form most often used in biological applications [55]–[57],[119].
Unfortunately, solving the CME exactly for most biologically, physiologically, or clinically meaningful systems is typically not feasible either analytically or numerically due to the intrinsic complexity of its differential-difference form. To address this problem, a number of alternative methods—focusing on approximate analytical solutions, general computational techniques, and a range of specific applications—have been developed [62], [121]–[126]. In practice, many of these methods either derive from or have their genesis in the Gillespie's Algorithm (SSA), which enables one to gain insight into possible temporal behaviors of the system by specifying how its sample paths can be exactly drawn from the CME-described probability distribution [62],[127],[128].
Our numerical simulations approach is based on the SSA and, specifically, is implemented as a streamlined version of Gillespie's Direct Method [127]. This is a kinetic Monte Carlo simulation procedure, which—given the system in state at time —determines per iteration: (i) the waiting time to the next reaction, , based on an exponential random variable with mean ; and (ii) the index of the next reaction, , based on an integer random variable with probability . (While the Next Reaction Method [129] is often considered to be the most efficient implementation of the SSA, recent study has discussed how the optimized version of the Direct Method generally performs better for many practical biochemical systems—largely owing to the high computational cost of maintaining extra data structures [130].) Our implementation is similar to other optimized versions of the Direct Method in the sense that it only evaluates propensity functions as necessary to minimize updates. The main difference is that our implementation does not create a dependency graph, but rather utilizes the bipartite graph structure of the reaction-based model to determine which propensity functions must be evaluated (see FimB and FimE Regulation Subnetwork section below for additional detail).
Using this implementation of the SSA in reb2sac, each simulation starts with the switch in the ON position and is run for up to one cell generation of 20 minutes as in [90]. If the switch moves to the OFF position within this time limit, the simulation is then counted as an ON-to-OFF switching event. The ON-to-OFF switching probability is calculated as the number of ON-to-OFF switching events divided by the total number of simulations with the same initial conditions. Alternatively, this could be viewed as computing the total ON-to-OFF switching probability by summing up switching events involved in all possible transition states, while the -mediated events only include transitions carried out due to the binding of —i.e., those going through switch states S4, S7, and S8—see Figure 4.
Our detailed switch inversion model represents a molecular reaction-scale description of the fim circuit (Figure 2), which generally satisfies the Markovian requirement of the SSA. (The discussion of how the individual reactions have been parameterized as well as generally identified from literature can be found below and in Text S1.) Such representations typically constitute the lowest-level (highest-resolution) description of biological systems used in most practical applications, which is one of the reasons why this model is correspondingly referred to as “detailed”.
The reaction network graph examination carried out as part of the motif recognition, data flow, system organization, and abstraction analysis has led us to identify two major modules responsible for dynamically controlling the fimS inversion process as well as integrating external signals provided by global regulator proteins and environmental factors, such as temperature, thus entailing a number of significant analytical and computational simplifications. These subnetworks may be broadly labeled as: (i) the production-degradation processes of and ; and (ii) the processes regulating the configuration of the fim switch itself.
While SSA offers a powerful method for numerically analyzing the behavior of discrete-stochastic biomolecular interaction networks, relying on just one or several simulation runs in order to gain a general understanding of a biological system subject to stochastic decision-making, such as UPEC fimbriation ON/OFF switching, could often be misleading because—similarly to the use of CCK—randomly-simulated individual sample trajectories of the underlying stochastic process are frequently insufficient to characterize its overall probabilistic dynamics [54]. In such settings, it typically requires thousands or more simulations in order to estimate the behavior of a system with reasonable statistical confidence. Yet, because SSA needs every single reaction event to be simulated one-at-a-time, it commonly leads to very high numbers of reaction events per given time step, particularly when the system has large characteristic time-scale separations. This makes computational requirements of exact numerical discrete-stochastic analysis exceedingly demanding for most practical biological and biomedical applications. In addition, the underlying complexity of biological chemical reaction and physical interaction networks as well as their innately differential response to varied environmental conditions generally impede qualitative interpretation of biological system organization and behavior. That is, though detailed reaction-level representations of biomolecular networks allow for very comprehensive descriptions of biological mechanisms, such low-level models can lead to substantial computational costs as well as may, potentially, obscure the overall system structure and dynamics. The problem could be further exacerbated by the particular choices of initial and environmental conditions that biological systems are embedded in. For example, while this paper discussed the behavior of the fim circuit in E. coli growing on minimal liquid medium, the in situ observed switching characteristics may be altered on rich liquid or solid medium [30]. Note that these adjustments in environmental conditions should not be expected to affect the underlying molecular reaction network structure of individual bacteria (since such variations do not determine the presence or absence of constituent elementary biomolecular interactions—only their rates), but rather lead to changes in observations due to effects ranging from heterogeneity in population dynamics among cell colonies on solid medium to input-driven modulations of various process rates comprising the circuit when switching to rich medium. Accurate analysis of the system in the former case requires application of dedicated population modeling schemes that themselves can lead to non-trivial empirical effects [35],[36],[133], thus creating further modeling complexity outside of the present scope. Similarly, in the latter case, changes in empirical settings—such as growing bacteria in a rich medium—tend to produce selective increases of some cellular process rates (e.g., those involved in metabolism/degradation or cell-division) that nevertheless leave many others unchanged. This introduces further time-scale separations into the problem, thus potentially making exact numerical analysis of discrete-stochastic circuit dynamics accessible in a minimal medium, but infeasible in a rich one [63],[64].
One approach toward addressing such challenges is the ongoing development of advanced analytical and numerical approximation methods—whether with respect to time (e.g., tau-leaping [60],[134]), state space (e.g., finite state projection [135],[136]), or other system variable—that are capable of significantly accelerating the analysis of master equation-type models to within a specified level of precision. This potentially makes feasible accurate computational analysis of molecular dynamics behind physiologically-meaningful biological networks that are otherwise too demanding for exact kinetic simulations (as, for example, is the case with bacterial systems grown in rich media or other such initial/external conditions). Thus, derivation and use of quantitatively analogous, but qualitatively and computationally simpler higher-level abstracted representations—which could be efficiently accomplished through systematic and, given the complexity of most biological processes, automatic application of various model approximations and simplifications—becomes essential [60], [62], [63], [134], [135], [137]–[142].
In practice, this could be done by utilizing a variety of techniques. For example, rapid-equilibrium and/or quasi-steady-state approximations [143]–[145] are often used to eliminate the various intermediates without significantly compromising our quantitative understanding of the overall system logic and functionality. Other methods may include: irrelevant node elimination, which removes species and reactions irrelevant with respect to the species of interest by statically analyzing the structure of the model; modifier constant propagation, which replaces a species-state variable in kinetic laws with the corresponding initial value and removes that species if that variable is statically known to be fixed; stoichiometry amplification, which amplifies stoichiometries and reduces the values of propensity functions—making the system and time advancement per reaction larger; and a number of additional approaches—many of which have been implemented in our reb2sac tool (see Table 1) [63],[64],[138]. The key principle behind most of these techniques could be summarized as identifying and abstracting away various redundant or largely irrelevant variables, whose dynamics do not independently influence the behavior of the system under a particular set of conditions—or, equivalently, finding a reduced set of parameters containing sufficient information to indentify system states and transitions between them. Since in the probabilistic context all information about a system is contained within its PDF, this could be viewed as finding a minimal subset of variables or their combinations that span the range of most likely/relevant states and elucidating abstracted laws governing their dynamics from those of the detailed description. (Various methods are available for quantifying the amount of probability distribution thus captured. For instance, information entropy and mutual information could be utilized for identifying the effective complexity of processes involved as well as further used to solve the inverse problem of elucidating system structure based on observations of state occupancies, such as inferring biomolecular network organization from individual species numbers [113], [146]–[149].) Alternatively, having identified the region of state space where most of the system's probability is localized, one may seek to restrict the problem to this lower-dimensional subspace, so as to obtain the corresponding reductions in problem complexity or otherwise coarse-grain its resolution when away from most relevant states and timescales. These approaches can be particularly fruitful when applied to biological molecular systems, whose probability distributions can be described by the CME. The latter offers a well-defined analytical structure for rigorously developing such approximations—which has led to several novel methods being proposed and applied in recent years [136], [137], [150]–[154]. (For example, it has been shown that master equations for switching systems can often be projected to much smaller dimensions with little loss in their accuracy [155].) Notably, since these methods are generally based on deep theoretical understanding of the underlying molecular chemical kinetics and reaction network graph analysis, the resulting abstracted models—such as those generated by reb2sac—on balance could be commensurably expected to accurately capture the overall biological system behaviors as well as to provide rigorous quantification of any potential divergences between the abstracted and detailed descriptions.
Although many approximation and abstraction approaches have been in wide use individually, their traditionally manual implementation grows to be increasingly more tedious and demanding as multiple methods are collectively applied to progressively larger biological systems. This problem is becoming even more acute as advances in systems biology continue to drive rapid increases in the typical size of analyzed networks, eventually rendering them intractable to interaction-level investigation and potentially leading to significant errors in large model transformations required to generate accurate intermediate-level abstractions. Our approach alleviates these problems by using a set of novel and existing algorithms—implemented in the reb2sac abstraction and analysis tool—to automatically survey and test biological networks for patterns and characteristics amenable to various complexity reduction techniques at the given level of accuracy for some specified “target” system property of interest [63],[64]. Among other things, this allows reb2sac to systematically scan through intermediate abstraction levels, to then automatically identify and implement appropriate approximation methods according to user preferences, and—by setting precision thresholds—to ultimately generate abstracted system models optimized for computational efficiency versus accuracy as desired. A high-level flow chart of our automated abstraction methodology is given in Figure 9. Note that the outlined analysis framework is overall quite generic and so could be used not only to generate model abstractions of gene regulatory networks, but also of other biochemical/biophysical reaction systems—including signal transduction pathways, metabolic networks, and other epigenetic processes.
Specifically, as shown in Figure 9, our abstraction engine takes as input a detailed reaction-based model and a set of abstraction properties. The latter help determine which of and how individual abstraction methods should be applied to the input model. These properties can also specify parameters for the conditions used by individual methods, enabling users to control the level of abstraction. The abstraction engine then passes this information through three internal stages: (i) pre-processing; (ii) main abstraction loop; and (iii) post-processing. Pre-processing is used to modify the structure of the input model so that the appropriate abstraction methods in the main loop can be applied more effectively. For example, if a model initially contains irrelevant reactions with respect to a particular species or dynamical property that the user is interested in analyzing—these reactions are removed at the pre-processing step to help speed up the abstraction process. The main loop contains abstraction methods that are applied repeatedly until the structure of the model no longer changes. In the case of gene regulatory networks, abstraction methods such as operator site reduction are typically placed in the main loop. Post-processing is used to transform the model into a form suitable for subsequent application of follow-up analysis methods—e.g., stochastic simulation, Markov chain analysis, etc.
As discussed earlier, transforming a detailed biological system model into an abstracted one can substantially increase the efficiency of its computational analysis as well as potentially improve our understanding of its overall structure and function. In this work, we have used the reb2sac automated abstraction tool to simplify the detailed model by systematically going through the fim switch network and applying various qualifying simplifications and/or approximations as appropriate. The resulting abstracted model is indeed significantly simpler computationally and more understandable logically than the detailed one. For example, the production-degradation reaction scheme of and (Figure 5A) is reduced by first quantitatively identifying the transcriptional regulator binding/unbinding events at the fimB and fimE promoter sites as “rapid” and the corresponding number of the operator sites (one) as “low”—and by then applying the rapid-equilibrium and quasi-steady-state approximations to these processes. The tool then continues to examine the dynamics of other species and finds that the concentrations of and RNA polymerase () do not change over time in our model. Thus, by applying modifier constant propagation, and are replaced with constants whose values are set to the corresponding initial concentrations and species and are removed from the model. This process continues until no further reductions are possible.
Taken together with the constraints imparted by the rates involved and the set target of fim switching probability, these abstractions reduce the detailed subnetwork of and shown in Figure 5A to the one shown in Figure 5B. Similar computational and logical complexity reduction is also achieved for the fim element configuration subnetwork. For instance, the reaction process corresponding to the fim switch inversion through state 6 (see Figure 4) is given in Figure 6A. The corresponding abstracted reaction scheme is shown in Figure 6B. Overall, after applying all of the available and appropriate abstraction techniques listed in Table 1, the detailed model with 52 reactions and 31 species (e.g., two recombinases, global regulatory proteins, and various intermediate complexes given in Figures 3 and 4) is transformed by reb2sac into an abstracted model with 10 reactions and 3 species (, , and switch given in Figures 5B and 6—the latter showing only reactions involved in ON-to-OFF switching events through circuit state 6).
|
10.1371/journal.pcbi.1002638 | High Resolution Genome Wide Binding Event Finding and Motif Discovery Reveals Transcription Factor Spatial Binding Constraints | An essential component of genome function is the syntax of genomic regulatory elements that determine how diverse transcription factors interact to orchestrate a program of regulatory control. A precise characterization of in vivo spacing constraints between key transcription factors would reveal key aspects of this genomic regulatory language. To discover novel transcription factor spatial binding constraints in vivo, we developed a new integrative computational method, genome wide event finding and motif discovery (GEM). GEM resolves ChIP data into explanatory motifs and binding events at high spatial resolution by linking binding event discovery and motif discovery with positional priors in the context of a generative probabilistic model of ChIP data and genome sequence. GEM analysis of 63 transcription factors in 214 ENCODE human ChIP-Seq experiments recovers more known factor motifs than other contemporary methods, and discovers six new motifs for factors with unknown binding specificity. GEM's adaptive learning of binding-event read distributions allows it to further improve upon previous methods for processing ChIP-Seq and ChIP-exo data to yield unsurpassed spatial resolution and discovery of closely spaced binding events of the same factor. In a systematic analysis of in vivo sequence-specific transcription factor binding using GEM, we have found hundreds of spatial binding constraints between factors. GEM found 37 examples of factor binding constraints in mouse ES cells, including strong distance-specific constraints between Klf4 and other key regulatory factors. In human ENCODE data, GEM found 390 examples of spatially constrained pair-wise binding, including such novel pairs as c-Fos:c-Jun/USF1, CTCF/Egr1, and HNF4A/FOXA1. The discovery of new factor-factor spatial constraints in ChIP data is significant because it proposes testable models for regulatory factor interactions that will help elucidate genome function and the implementation of combinatorial control.
| The letters in our genome spell words and phrases that control when each gene is activated. To understand how these words and phrases function in health and disease, we have developed a new computational method to determine what word positions in our genomic text are used by each genome regulatory protein, and how these active words are spaced relative to one another. Our method achieves exceptional spatial accuracy by integrating experimental data with the text of our genome to find the precise words that are regulated by each protein factor. Using this analysis we have discovered novel word spacings in the experimental data that suggest novel genome grammatical control constructs.
| Genomic sequences facilitate both cooperative and competitive regulatory factor-factor interactions that implement cellular transcriptional regulatory logic. The functional syntax of DNA motifs in regulatory elements is thus an essential component of cellular regulatory control. Appropriately spaced motifs can facilitate cooperative homo-dimeric or hetero-dimeric factor binding, while overlapping motifs can implement competitive binding by steric hindrance. Cooperative and competitive binding are an integral part of complex cellular regulatory logic functions [1], [2]. The binding of regulatory proteins to the genome cannot at present be predicted from primary DNA sequence alone as chromatin structure, co-factors, and other mechanisms make the prediction of in vivo binding from sequence empirically unreliable [3]. Thus it is not possible to use primary DNA sequence to determine the aspects of genome syntax that are employed in vivo.
To discover novel pair-wise factor spatial binding constraints in vivo, we have developed a new method called GEM that simultaneously resolves the location of protein-DNA interactions and discovers explanatory DNA sequence motifs with an integrated model of ChIP-Seq or ChIP-exo reads and proximal DNA sequences. We define a binding event location as the single base position at the center of a protein-DNA interaction. GEM reciprocally improves motif detection using binding event locations, and binding event predictions using discovered motifs. In doing so, GEM offers a more principled approach than simply snapping binding event predictions to the closest instance of the motif, and indeed, GEM does not require that all binding events are associated with strong motifs. GEM offers both improved spatial accuracy of binding event predictions and improved motif discovery in ChIP-Seq and ChIP-exo datasets.
GEM's unbiased computational approach has enabled us to discover novel binding constraints between transcription factors from sequenced ChIP experiments. These spatial constraints directly suggest biological regulatory mechanisms that will be useful in future studies. Other methods to resolve binding events in sequenced ChIP data identify statistically enriched regions of ChIP-Seq read density and the peak points of enrichment within those regions [4]–[9], and binding calls can be offset from the bound site by dozens of bases [10]. Recent studies have integrated peak detection and motif discovery by including motif occurrences to score the significance of predicted binding events [11], [12], or by using ChIP-Seq read coverage as a positional prior to improve motif discovery [13], [14]. However, no study has yet used the motif position information to reciprocally improve the spatial accuracy of binding event prediction. SpaMo studied the motif spacing using ChIP-Seq events to infer transcription factor complexes but the predicted motif spacing does not necessarily indicate in vivo binding in the specific cellular conditions [15].
Here we review our GEM derived results, discuss these results in the context of current data production projects, and detail our methods.
We compared GEM's spatial resolution to six well known ChIP-Seq analysis methods, including GPS [8], SISSRs [6], MACS [4], cisGenome [7], QuEST [5] and PeakRanger [9]. We used a human Growth Associated Binding Protein (GABP) ChIP-Seq dataset for our evaluation because GABP ChIP-Seq data were previously reported to contain homotypic events where the reads generated by multiple closely spaced binding events overlap [5]. Thus the GABP dataset offers the opportunity to test if integrating motif information and binding event prediction improves our ability to deconvolve closely spaced binding events with greater accuracy. We also evaluated the methods using ChIP-Seq data from the insulator binding factor CTCF (CCCTC-binding factor) [16], as it binds to a stronger motif than GABP. These two factors are representative of relatively easy (CTCF) and difficult (GABP) cases for ChIP-Seq data analysis. They are also used by other studies as benchmarks allowing for the direct evaluation of our results. GEM performance on other factors may vary.
We found that GEM has the best spatial resolution among tested methods. Spatial resolution is the average absolute value difference between the computationally predicted locations of binding events and the nearest match to a proximal consensus motif. From all observations, spatial resolution is corrected for a fixed offset by subtracting the mean difference before averaging the absolute value differences. To ensure a fair comparison, we used 428 shared GABP binding sites that are predicted by all seven tested methods and which contain an instance of the GABP motif within 100 bp. GEM exactly locates the events at the motif position in 56.5% of these events (Figure 1A). For a dataset with a stronger consensus motif, ChIP-Seq data from CTCF, GEM exactly locates the events at the motif position in more than 90% of the shared events, significantly improving the spatial accuracy of predicted binding events over other methods (Figure 1B). Alternative evaluations with all the binding sites that have a motif at a distance less than 100 bp are also performed for both GABP and CTCF data, and the results (Figure S1) are similar to those above. Thus, GEM's joint model of ChIP-Seq read coverage and sequence is able to more accurately predict the location of binding sites than other approaches, which do not use motif information in their binding event predictions.
GEM is also better at resolving closely spaced binding events [17] in the GABP data than the other methods we tested. For example, GEM uniquely detects two GABP events over proximal GABP motifs that are 32 bp apart on chromosome 2 (Figure 1C). To evaluate binding deconvolution on a genome-wide scale, we identified 477 candidate clusters of closely spaced binding events. Each candidate cluster was detected as bound by all seven tested methods and contained two or more proximal GABP motifs separated by less than 500 bp. GEM identified two or more closely spaced events in 144 of the candidate clusters, significantly more than GPS(108), SISSRs(77), QuEST(77), PeakRanger(36), MACS(4)and cisGenome(5) (Figure 1D).
We tested GEM's ability to discover biologically relevant DNA-binding motifs in data from the ENCODE project [18]. We chose this large collection of experiments because we expected they would be representative of the typical range of ChIP-Seq data noise and sequencing depth. Noise can be caused by low antibody affinity and deviations from ideal experimental procedure. We used a set of 214 ChIP-Seq experiments and associated controls comprising 63 distinct transcription factors that were profiled in one or more cell lines by the ENCODE project and for which validated DNA-binding motifs exist in public databases (Dataset S1). GEM analyzed these ChIP-Seq data, and the most significant GEM-discovered motifs from each analysis (Table S1 and Dataset S2) were compared to corresponding known binding preferences of the same transcription factors using STAMP [19]. A motif alignment with E-value less than 1e-5 was considered a match. For comparison, we also used four popular traditional motif discovery tools covering a range of computational techniques, including MEME [20], Weeder [21], MDScan [22], and AlignACE [23], and three ChIP-Seq oriented tools, POSMO [24], HMS [13] and ChIPMunk [14] on the same data. A set of 100 bp sequences extracted from the 500 most highly ChIP-enriched GPS peaks calls are examined by the motif-finders MEME, Weeder, MDScan, AlignACE, or POSMO. For HMS and ChIPMunk, a set of 100 bp sequences and corresponding read coverage profiles are extracted from the 500 most highly ChIP-enriched GPS peaks calls.
We found GEM outperforms all of the compared motif discovery approaches, even when allowing each method to make multiple motif predictions (Figure 2, Table S2, S3). Therefore, the GEM approach to integrating ChIP-Seq event detection with motif analysis not only improves the spatial resolution of binding events, but also more accurately finds the expected binding motifs present at those events. We note that GEM sometimes failed to find the known motif in datasets where one of the other algorithms succeeds. The complete evaluation is in Table S2, S3.
We then tested GEM on ENCODE ChIP-Seq experiments for 9 distinct transcription factors with no publically described DNA binding motif. For 6 of these transcription factors, GEM discovers novel motifs that are consistent with expected binding sequences based on a small number of binding sites characterized in the literature, or similarity to the known binding preferences of related proteins (Table S4). For example, GEM confirms that BATF has a similar binding preference to other members of the AP1 family of transcription factors. The similar TGAC/G binding preference has previously been supported by EMSA assays on regions upstream potential BATF regulated genes [25].
ChIP-exo aims to improve transcription factor binding spatial resolution by extensively digesting ChIP fragments down to the DNA that is protected by the bound protein complex [26]. While ChIP-exo experiments provide high-resolution binding information, typical peak-finding methodologies may fail to achieve single-base resolution binding event predictions if they do not account for the properties of the ChIP-exo experiment. An example is provided by the published CTCF ChIP-exo experiment [26], where ChIP-exo reads are bimodally distributed around binding sites on both strands because CTCF is cross-linked at two distinct sites of DNA. The published event predictions did not account for this characteristic distribution, and are thus often offset from CTCF binding motif instances. Since GPS and GEM automatically learn a model of sequence reads around binding events, GPS and GEM may be directly applied to ChIP-exo data without modification. We first verified that GEM's model of binding events is able to automatically adapt to the read distribution produced by the ChIP-exo protocol. We compared GEM's final computed read distribution to the expected empirical distribution of ChIP-exo and found that they were consistent (Figure 3B and Figure S2).
GEM improves upon the spatial resolution of binding event detection over other methods for ChIP-exo data (Figure 3A). To investigate the performance of GEM on ChIP-exo data, we compared the binding event predictions of GEM and GPS on ChIP-exo CTCF binding and the “middle of peak-pair” method from the original ChIP-exo study [26]. To ensure a fair comparison, we used 5074 shared binding sites that are predicted by all tested methods and that contain a strong CTCF motif match within 100 bp of the binding positions. The original ChIP-exo study [26] had 5.4% of the binding event calls centered on the motif match position, 40.3% of the calls within 10 bp, and an average spatial resolution of 15.85±15.29 bp. Applying GPS to the ChIP-exo data improved the spatial resolution, with 8.8% calls at 0 bp positions, 59.7% of calls within 10 bp, and average spatial resolution of 10.38±11.26 bp. Applying GEM to the ChIP-exo data located 76.5% calls exactly at the motif match positions, 89.7% of calls within 10 bp, and an average spatial resolution of 3.35±9.71 bp. These results demonstrate that GEM can significantly improve the spatial accuracy of ChIP-exo binding event predictions.
We examined if GEM could detect pairs of transcription factors that bind to the genome with characteristic pair-wise spacing, beginning with the well-known hetero-dimeric pair Sox2-Oct4 [27]. In general, distance-constrained transcription factor binding cannot be predicted based solely on sequence motifs as motif presence does not guarantee binding. Such spatial binding constraints may be caused by combinatorial binding, alternative binding, binding that is orchestrated by multimeric protein complexes, or the spread of constrained enhancer syntax.
We were able to discover Sox2-Oct4 transcription factor spatial binding constraints by combining GEM binding calls from Sox2 and Oct4 ChIP-Seq data. We applied GEM independently to mouse ES cell Sox2 and Oct4 ChIP-Seq data [15] to call the respective binding sites, and then computed the distance between Oct4 sites from Sox2 sites within a 201 bp window. The sequence strand of the GEM binding predictions is oriented using the Sox2 motif when a match to the motif is present. As expected, GEM predicted Oct4 binding sites are predominantly (630 sites out of 2525 in the 201 bp window) located at −6 bp position relative to GEM predicted Sox2 sites (Figure 4A and Figure S3). However, this spacing cannot be observed from the binding calls of GPS or other event discovery methods alone because of their more limited spatial accuracy (Figure 4B). An alternative approach is to snap binding calls to the nearest instance of the transcription factor's binding motif. We tested this approach using GPS binding calls as the starting points and found that the alternate approach captures fewer (277 sites out of 2753) instances of Oct4-Sox2 spatial binding constraints (Figure 4C), presumably because some of the bound motifs do not pass the motif scoring threshold or because some unbound motif instances are located closer to the binding calls than the true motif instances.
We next studied pair-wise binding relationships between 14 sequence-specific transcription factors (Oct4, Sox2, Nanog, Klf4, STAT3, Smad1, Zfx, c-Myc, n-Myc, Esrrb, Nr5a2, Tcfcp2l1, E2f1 and CTCF) and two transcriptional regulators (p300 and Suz12) in mouse ES cells by applying GEM to a large compendium of ChIP-Seq binding data [16], [28]. Binding prediction by GEM enables the detection of 37 pairs of statistically significant spatial binding constraints, involving Oct4, Sox2, Nanog, Klf4, Esrrb, Nr5a2, Tcfcp2I1, E2f1, c-Myc, n-Myc and Zfx (Figure S4, the full list of TF pairs are in Table S6, S7, motifs are in Table S5 and Dataset S3). Interestingly, we found that Klf4, one of the ES cell reprogramming factors, exhibits strong distance-specific binding with many other factors, including Nanog, Sox2, Zfx, c-Myc, n-Myc, E2f1, Esrrb, Nr5a2 and Tcfcp2l1 (Figure S5).
The discovered pair-wise spatial binding constraints reveal complex relationships among the factors. For example, Klf4 exhibits constrained binding with Sox2 but much less significantly with Oct4 (Figure S5). However, we did observe strong distance-specific binding between Oct4-Sox2 (Figure 4A). This raises the question of whether the detected Klf4-Sox2 and Oct4-Sox2 spatial binding constraints are on the same genomic regions. We therefore studied all Sox2 bound regions that are co-bound with Klf4. Out of a total of 5609 Sox2 bound regions with a Sox2 motif instance that can be oriented, 123 regions are co-bound by Klf4 at position +25 bp (Figure 5A). However, only four regions show co-binding of Klf4 at position +25 bp and Oct4 at position −6 bp. More surprisingly, the distance-constrained Sox2/Klf4 regions are co-bound by 6 ES cell factors within a 70 bp window, including Sox2 (at 0 bp), Nanog (at 1 bp), Klf4 (at 25 bp), Esrrb (at 56, 59 bp), Nr5a2 (at 55, 58, 61 bp) and Tcfcp2I1 (at 66, 69 bp). Inspecting the underlying sequences of these regions, we found that the binding motifs of these factors are embedded at the positions consistent with the binding positions (Figure 5B). In addition to the consistent spatial arrangement of motifs, these sequences (spanning from −70 bp to 100 bp) exhibit a high degree of similarity. A subset of the sequences is shifted 3 bases by some insertion/deletions, consistent with the 3 bp shift of some of the factor binding positions. Comparing with p300 and H3K27ac ChIP-Seq datasets [29], we found that almost all (119 out of 123) of these regions are bound by p300, a histone acetyltransferase and transcriptional coactivator that predicts tissue-specific enhancers [30]; the majority of these regions are also marked by H3K27ac, a histone modification associated with active enhancers [29], suggesting that they may be active enhancer regions (Figure S6). These results demonstrated that GEM analysis enables detection of coordinated binding of multiple factors that are driven at least partly by the underlying sequences.
Of the 123 regions where Sox2, Klf4, and other sites display constrained spacing, 109 (89%) are annotated instances of the RLTR9 ERVK family of long terminal repeat elements. It is interesting to note that while Bourque, et al. found an association between Oct4/Sox2 co-binding sites and other members of the ERVK repeat class [31], we find a set of repetitive elements that encode the binding of Sox2 and other factors without Oct4 in ES cells. Kunarso, et al. suggested that transposable elements have rewired the core regulatory network of ES cells [32]. Our analysis found that the repetitive sequences constrain the in vivo binding of a number of key transcription factors in ES cells.
We computed statistically significant pair-wise spatially constrained binding events between 46 transcription factors characterized in 184 ENCODE ChIP-Seq data sets in five different cell lines. Each transcription factor ChIP was processed independently by GEM so that we could assess any differences in observed binding between cell lines and biological replicates.
We found that 390 pairs of transcription factors have significant binding distance constraints within 100 bp of each other (Figure 6–7, Figure S7, S8, S9, S10, the full list of TF pairs are in Table S8, S9). The number of pairs found in each cell line differed as did the number of transcription factors assayed: K562 (152 pairs/37 TFs), GM12878 (148 pairs/29 TFs), HepG2 (107 pairs/29 TFs), HeLa-S3 (48 pairs/15 TFs), and H1 (23 pairs/11 TFs). Certain factor-pairs exhibited a highly significant single binding spacing offset within 100 bp, such as the 4 bp distance between Egr1 and CTCF in K562 cells (Figure 6). Other factor pairs exhibited a large number of significant offsets, such as the 167 significant spacings between JunD and Max with the most significant being at 4 bp (Figure 6–7). Our analysis confirmed known interaction pairs MYC-MAX [33], the FOS-JUN heterodimer [34], and CTCF-YY1 [35] (Table S8, S9).
Observed novel genome wide spatial binding constraints include c-Fos:c-Jun/USF1, CTCF/Egr1, HNF4α/FOXA1. We find that USF1 often binds 4 bp from c-Fos:c-Jun (Figure 8A and Figure S11). This binding is consistent with Fra1's facilitation of a complex between USF1 and c-Fos:c-Jun [36]. We find a significant number of cases where CTCF co-binds 4 bp from Egr1 (Figure 8B and Figure S12). Egr1 promotes terminal myeloid differentiation in the presence of deregulated c-Myc expression, and Egr1 has been implicated in down regulating c-Myc in conjunction with CTCF [37]. In addition, the co-binding of CTCF and Egr1 at the EPO regulatory region has been suggested [38]. FOXA1 binds at a large number of significant positions close to HNF4α (total 4215 regions with a spacing within 30 bp, Figure 8C and Figure S13), and there are also significant binding constraints between HNF4α and HNF4γ and FOXA1, FOXA2 in HepG2 cells (Table S8, S9). While co-binding of HNF4α/FOXA2 has been reported [39], co-binding of HNF4α/FOXA1, HNF4γ/FOXA1 and HNF4γ/FOXA2 are not known. We note that HNF4α and any one of FOXA1, FOXA2, or FOXA3 is sufficient to reprogram cells towards a hepatocytic fate [40].
Collectively, our results demonstrate that it is now possible to reveal aspects of functional genome syntax by surveying in vivo binding relationships between transcription factors at high spatial resolution. Our analysis has been made possible by sequenced ChIP data and a new computational method, GEM, which provides exceptional spatial resolution.
GEM makes binding predictions and observes spatial constraints by discovering significant events utilizing both motifs and observed read coverage information. Prior work has documented specific genomic regions extensively targeted by multiple transcription factors (TFs) [16]. However, we have shown that the functional syntax of DNA motifs in regulatory elements cannot be fully elaborated with the imprecise ChIP-Seq event calls provided by previous methods. Motif analysis approaches such as SpaMo discover enriched motif spacing by scanning a list of known motifs in sequences anchored by ChIP-Seq data of a single factor [15]. Since the existence of motif instances does not guarantee condition specific in vivo binding, SpaMo cannot confidently determine the spacing between binding events and the factors involved, especially for motifs that are shared by a family of TFs. Furthermore, SpaMo excludes repetitive sequences [15]. In contrast, GEM predicts binding based on uniquely-mapped reads and is able to detect spatial binding constraints in transposable elements. Such elements have been implicated in rewiring the core regulatory network of human and mouse ES cells [32].
We expect that the genome grammatical rules that are suggested here will be examined in further studies to elucidate mechanisms of transcriptional control, and potential protein-protein interactions that have regulatory consequences. Exploration of other genome grammatical constructs can be accomplished with the use of further ChIP experiments and GEM.
The GEM algorithm consists of six phases:
Initial protein-DNA binding event locations are predicted by GPS [8], which employs a negative Dirichlet sparse prior.
GEM discovers a set of enriched k-mers by comparing k-mer frequencies between positive sequences and negative control sequences. The positive set consists of 61 bp sequences centered on the predicted binding locations from Phase 1, and a negative set consists of 61 bp sequences that are 300 bp away from binding locations and that don't overlap positive sequences. We count the number of positive and negative sequences that contain instances of each possible k-mer (hit count), treating each k-mer and its reverse complement as the same sequence. A k-mer is considered enriched if the hypergeometric p-value [41] of its enrichment is less than 0.001 and it has at least 3-fold enrichment in terms of positive/negative hit count. In this study, values of k from 5 to 13 are used on each dataset, and the final k value is chosen as the one that gives the most significantly enriched primary PWM as described below. Each k-mer carries with it its expected offset from a binding event as averaged over the positive set.
GEM next clusters the enriched k-mers into equivalence classes that describe similar DNA binding preferences (Figure S14). Each equivalence class is a collection of k-mers. A genomic sequence is said to match a k-mer equivalence class if the genomic sequence contains any of its component k-mers. GEM clusters enriched k-mers into k-mer equivalence classes by (Figure S14):
After finding the primary k-mer equivalence class, GEM searches for other classes. To accomplish this, the previous seed k-mer is removed from the enriched k-mer pool and PWM motif occurrences are masked in the sequences. The process of building new k-mer equivalence classes is repeated until no more significantly enriched PWMs can be constructed. Rarely, a secondary motif PWM can become more significantly enriched than the primary motif. If this happens, the motif finding process is restarted using the seed k-mer of this secondary motif.
Phase 4 of GEM uses the primary k-mer equivalence class to compute a Dirichlet prior that will be used for binding event discovery in Phase 5. The genome is segmented into independent separable regions (typically a few kb long) by dividing at read gaps that are larger than 500 bp and further excluding regions that contain fewer than 6 reads [8]. At each evaluated genome region, we simultaneously search the occurrences of all the k-mers of the primary k-mer equivalence class using the Aho-Corasick algorithm [43], and matches are marked at the expected binding event location for every matching k-mer. The position-specific prior for a sequence base is defined as the number of positive set sequences that contain one of the enriched k-mers whose binding offsets match that base. The concept of using informative positional priors for motif discovery has been explored previously [44], [45].
GEM employs a generative mixture model that describes the likelihood of a set of ChIP-Seq reads being generated from a set of protein-DNA interaction events originating at specific DNA sequences. The model generates protein-DNA interaction events that are biased to occur at explanatory DNA sequences by a k-mer based positional prior. Each event then independently generates reads following an empirical read spatial distribution that describes the probability of reads given the distance from the event [8] (see Figure 3B for an example).
Formally, in an evaluated region of length M, we consider N ChIP-Seq reads that have been mapped to genome locations R = {r1, …, rN} and M all possible protein-DNA interaction events at single base locations B = {b1, …, bM}. We represent the latent assignments of reads to events that caused them as Z = {z1, …, zN}, where indicator function 1(zn = m) = 1 when read n is caused by the event m.
The probability of a read n is based on a mixture of possible binding events:
where M is the number of possible events; π denotes the parameter vector of mixing probabilities, and πm is the probability of event m; p(rn | m) is the probability of read n being generated from event m and can be determined from the empirical spatial distribution of reads given the event [8].
The overall likelihood of the observed set of reads is:
We make two prior assumptions about the binding events: 1) binding events prefer to occur at the sequence specific DNA motif positions; 2) binding events are relatively sparse throughout the genome. To incorporate these assumptions, we place a negative Dirichlet prior [8], [46] p(π) on binding event probabilities π:
where αs is the uniform sparse prior parameter governing the degree of sparseness, αs>0; αm denotes the binding event specific prior parameter and its value is proportional to Cm, the positional prior count underlying event m (as defined in Phase 4):
where μ is a parameter to tune the effect of motif based prior, 0≤μ<1. In this study, we choose μ = 0.8.
The rationale is that if the k-mers mapped to position m have more occurrences at binding events genome wide, it is more likely to cause a binding event at that genome position. The parameter αm is scaled such that all the values of possible αm will be less than αs. Therefore the k-mer based prior will not force the model to predict a binding event at a motif position when the observed reads do not provide sufficient evidence of a protein-DNA interaction event.
Since the k-mers underlying the possible binding event positions and their counts are known, the value of term −αs+αm remains constant when we estimate the parameters in the mixture model. Therefore, we can solve the mixture model using Expectation-Maximization (EM) algorithm [47].
The complete-data log penalized likelihood is:
where 1(zn = m) is the indicator function.
In the E Step we have:
where γ(zn = m) can be interpreted as the fraction of read n that is assigned to event m.
In the M step, on iteration i we find parameter to maximize the expected complete-data log penalized likelihood:
under the constraint . By simplifying we find the close-form solution of the maximization as:
where Nm is the effective number of reads assigned to event m, or the binding strength of event m. Intuitively, the effective read count of an event is decreased by a pseudo-count αs for the sparseness penalty, and is increased by a pseudo-count αm for the k-mer motif at position m. If for event m, the value of πm becomes zero, the model is restructured to eliminate it [46].
The EM algorithm is deemed to have converged when the change in likelihood falls below a small value, for example 1e−5.
Since the value of term −αs+αm is negative, a binding event supported by enriched k-mers may still be eliminated if it is not sufficiently supported by read data. In addition, a binding event not supported by enriched k-mers may still survive if it is sufficiently supported by the read data.
The predicted binding events are tested for significance as described previously [8]. Briefly, if a control dataset is available, we compare the number of reads in the ChIP event to the number of reads in the corresponding region in the control sample using a Binomial test. If control data is not available, we apply a statistical test that uses a dynamic Poisson distribution to account for local biases. To correct for multiple hypothesis testing, a Benjamini-Hochberg correction [48] is applied. It is worth mentioning that we only use read counts of events to test for significance.
The read spatial distribution of binding events is updated after each round of binding event prediction.
Phase 6 repeats Phase 2 and 3 motif discovery using the binding events predicted from Phase 5. As described in the results section (Figure 1), the spatial accuracy of binding events discovered from Phase 5 (GEM) is significantly improved from Phase 1 (GPS). Thus, these events will be more accurately centered on motifs and the performance of motif discovery is correspondingly improved.
GEM is a stand-alone Java software that takes alignment files of ChIP-Seq reads and a genome sequence as input and reports a list of predicted binding events and the explanatory binding motifs. It can be downloaded from our web site (http://cgs.csail.mit.edu/gem). For analysis with mammalian genomes, GEM requires about 5–15 G memory.
214 ENCODE ChIP-Seq datasets that have an embargo date before Oct 28, 2011 and have known motifs in public databases were downloaded from the ENCODE project website [18]. 16 mouse ES cell factor ChIP-Seq datasets published in references [16] and [28] were downloaded from GEO. ChIP-exo data were provided by Ho Sung Rhee and B. Franklin Pugh. FastQ files of the ChIP-Seq/ChIP-exo data were then aligned with genome (human hg19, mouse mm9) using Bowtie [49] version 0.12.7 with options “-q --best --strata -m 1 -p 4 --chunkmbs 1024”. The GABP ChIP-Seq data was downloaded from QuEST website (http://mendel.stanford.edu/SidowLab/downloads/quest/) and was pre-aligned to hg18 genome.
GEM was applied to 214 ENCODE ChIP-Seq data. The motif PWMs output by GEM were collected. An alternate pipeline used the GPS peak-finder [8] to call binding events and used 7 different motif finding methods (AlignACE v4.0 [23], MDscan v2004 [22], MEME v4.7.0 [20], Weeder v1.4.2 [21], POSMO v2 [24], HMS v0.1 [13] and ChIPMunk v3 [14]) to discover motifs independently. For AlignACE, MDscan, MEME and Weeder, 100 bp sequences were extracted from the top 500 peaks from each dataset, as suggested by the MEME Suite's documentation based on the typical resolution of ChIP-Seq peaks. For POSMO, we extracted a set of 100 bp sequences from the top 500 GPS peaks. This set of sequences provided superior results when compared with sequences taken from the top 5000 1000 bp sequences (as suggested by the author of POSMO). For ChIP-Seq oriented methods, HMS and ChIPMunk, a set of 100 bp sequences and corresponding read coverage profiles were extracted from the top 500 GPS peaks. We found these conditions provided superior results than using sequences taken from the top 5000 200 bp sequences (as suggested by the authors of these methods). MEME was run with “-nmotifs 6” and Weeder was run with option “large”. POSMO was run with options “5000 11111111 sequence_file 1.6 2.5 20 200”. ChIPMunk was run with options “6 15 yes 1.0 p:read_coverage_profile 100 10 1 4 random 0.41”. HMS was run with options “-w motif_width -dna 4 -iteration 100 -chain 50 -seqprop 0.1 -strand 2 -base read_coverage_profile -dep 2”; motif_width was determined by width of motif discovered by MEME for the same data. All other parameters were the defaults specified by the authors.
We collected known binding preference motifs from the TRANSFAC [50], JASPAR [51], and Uniprobe [52] databases. We only include motifs of the factors of interest or motifs for the TF family but not motifs of factors in the same family because factors in the same family may have very different binding motifs. The list of database matrices is provided in Dataset S1. Discovered motifs were compared to known motifs using STAMP [19]. A motif with E-value less than 1e-5 was considered a match. For each program, we counted the number of datasets that had a motif matching at least one known motif of that transcription factor. In some cases, the correct motifs are not matched by the first motif that a method outputs, but by the second or later motifs. Therefore we compare the motif-finding performance using the top 1, top 2… or top 6 motifs. Little improvement is observed after the 6th motifs.
The genome-wide performance of spatial resolution in ChIP-Seq event calls is evaluated as following. We define effective spatial resolution as the average absolute value of the distance between genome coordinates of predicted binding events and the middle of the corresponding high-scoring binding motif hit. Because the center of the motif hit may not represent the true center of a binding event, the offsets to the motif were centered by subtracting the mean offsets. We compare spatial resolution on the “matched” set of predictions that are called by all the methods and correspond to the same high-scoring binding motif. Only those events within 100 bp of a motif match are included in the calculation. An alternative evaluation with all the events that have a motif at a distance less than 100 bp is also performed.
The genome-wide performance of proximal event discovery in ChIP-Seq data is evaluated as follows. For GABP dataset, we compared GEM against other 6 methods (GPS, SISSRs, MACS, cisGenome, Quest and PeakRanger) genome wide. We define a set of candidate sites that all have at least one event detected by all seven methods, and that contain two or more GABP motifs separated by less than 500 bp. We discovered 477 such sites. For each of these sites, we count the number of events discovered by different methods. GABP motif was retrieved from TRANSFAC database (M00341) [50]. A motif score threshold of 9.9, which is 60% of maximum PWM score, is used in this analysis.
In this study, to test GEM's ability to automatically adapt to ChIP-exo data, we initialized GEM with a ChIP-Seq empirical read distribution, and ran GEM with one extra run (phase 5 and 6) so that GEM could use more accurately positioned events to refine the read distribution and use it for final prediction. In practice, the user can directly initialize GEM with a ChIP-exo empirical read distribution (provided with GEM software) and apply GEM the same way as analyzing ChIP-Seq data.
To study the in vivo binding spatial relationship between a pair of transcription factors A and B in the certain cell type and condition, we apply GEM independently to ChIP-Seq data from A and B to predict the respective binding sites. To compute the distribution of spacing between A relative to B, we compute the offsets of A binding sites from B binding sites within a 201 bp window. The sequence strand of the binding predictions is oriented using the B motif when a match to the motif is present, and B is placed in the middle of the window. The occurrences of A at each offset position are summed over all the B sites to produce the empirical spatial distribution. In this study, we evaluate three different methods to call binding sites: GEM binding calls, GPS binding calls, and GPS binding calls that are snapped to a motif within 50 bp if one is present. Another motif distance for snapping binding calls, 100 bp, was also tested and the result was very similar to the 50 bp distance.
To determine if a specific spacing is significant, we compute the p-value of the number of occurrences of factor A at that offset position using a Poisson test. The parameter of Poisson distribution is set as the mean number of occurrences across all the positions in the [−400 bp −200 bp] and [200 bp 400 bp] windows, assuming there are no significant spatial binding constraints in these windows. The p-value is corrected for multiple hypotheses testing using Bonferroni correction by multiplying the p-value by the number of positions in the window and the total number of pair wise tests across all cell types. The significance threshold for corrected p-value is 1e−8. Because the strand orientation of bound sequences cannot be oriented consistently when comparing multiple factor pairs, we report the absolute distance between the most significant interacting factor pairs in Figure 6.
|
10.1371/journal.pntd.0001494 | WormAssay: A Novel Computer Application for Whole-Plate Motion-based Screening of Macroscopic Parasites | Lymphatic filariasis is caused by filarial nematode parasites, including Brugia malayi. Adult worms live in the lymphatic system and cause a strong immune reaction that leads to the obstruction of lymph vessels and swelling of the extremities. Chronic disease leads to the painful and disfiguring condition known as elephantiasis. Current drug therapy is effective against the microfilariae (larval stage) of the parasite, but no drugs are effective against the adult worms. One of the major stumbling blocks toward developing effective macrofilaricides to kill the adult worms is the lack of a high throughput screening method for candidate drugs. Current methods utilize systems that measure one well at a time and are time consuming and often expensive. We have developed a low-cost and simple visual imaging system to automate and quantify screening entire plates based on parasite movement. This system can be applied to the study of many macroparasites as well as other macroscopic organisms.
| The World Health Organization estimates that there are approximately 37 million people who are afflicted by Onchocerca volvulus (the parasitic worm that causes river blindness) and over 120 million people afflicted by the filarial worms Wuchereria and Brugia spp. (causative agents of lymphatic filariasis or elephantiasis). Current mass drug administration includes albendazole and either diethylcarbamazine or ivermectin. These drugs, however, are effective at killing the early larval stage (microfilariae) released from adult female worms but they do not kill the adult worms. Adult worms can live up to 10 or more years, releasing thousands of microfilariae per day. It is essential therefore to treat infected individuals with macrofilaricides in order to prevent the adult parasites from producing microfiliariae for the duration of the infection and to treat the disease. In order to screen candidate drugs for use as macrofilaricides, we have developed an inexpensive system and simple method for quantifying the effectiveness of drugs on parasite movement. The apparatus uses a commodity video camera, a computer and a newly developed free and open source software application to provide automated and quantitative measurements of parasite motility on each plate of worms. This system is not only useful for high throughput screening of macroparasites but can also be applied to the study of other macroscopic organisms as well.
| Lymphatic filariasis is a devastating parasitic disease that affects more than 120 million people in 81 countries [1]. Also known as elephantiasis, the disease is caused by parasitic nematodes whose adult forms inhabit the lymphatic system. WHO estimates that over 1.3 billion people are at risk for lymphatic filariasis and approximately 95% of infected individuals live in Africa and South-East Asia.
Lymphatic filariasis is spread mainly by three species of nematodes in the family Filariodidea: Wucheria bancrofi, Brugia malayi and Brugia timori. The adult forms are threadlike roundworms from 2–5 cm in length and adult females produce millions of microfilariae that circulate in the blood where they are vectored by mosquitoes. The microfilariae develop into the infectious form in the mosquito and are inoculated into individuals when the mosquito takes a blood meal. Larval forms migrate to the lymphatic vessels and mature in 6–12 months and begin releasing microfilariae, completing the cycle of transmission.
The WHO currently advocates interrupting transmission of the disease via an annual mass drug administration of single doses of albendazole and either diethylcarbamazine or ivermectin. These drugs are effective at killing microfilariae but not effective against adult worms (macrofilariae). Since adult worms can live for up to 6–8 years, treatment must be given on a regular basis to break the cycle of transmission. This widespread treatment is logistically challenging and costly, particularly in endemic regions that are politically unstable. Such widespread application also raises the threat of resistance, whose first signs are being seen with ivermectin in Onchocerca volvulus [2]. Mass drug administration would be greatly aided by a macrofilaricidal drug, as the adult parasites would not be able to continue producing microfilariae for the duration of the infection. Hence, there is a great need for new macrofilaricidal drug candidates.
Currently there is no high throughput screening (HTS) method available to screen compounds targeting any of these macroscopic nematodes in vitro. Assays have been developed in recent years that score worm migration, feeding and development [3]–[19] as well as worm viability based on the MTT ((3-(4,5-dimethylthiazola-2-yl)-2, S-diphenyl tetrazolium bromide) assay [7], [8], [11]–[13], [20], [21], but these are not amenable to screening 1000's of compounds with a quick turnaround time on large worms such as filarid nematodes. Parasite movement is an important indication of the effectiveness of a drug and constitutes a crucial phenotype for HTS. Existing assays however, read only single wells at a time, are low-throughput and are unaffordable for many laboratories in developing nations [3], [22]–[24]. In the context of analyzing the phenotypes of model organisms, Buckingham and Sattelle published an algorithm for measuring the thrashing of Caenorhabditis elegans via a statistical analysis of the covariance matrix between sets of frames to determine the period of thrashing [24]. The algorithm is specific to the thrashing phenotype and is not a complete ready-to-screen software application. Both Buckingham and Sattelle's, and Ramot et al.'s [23] applications require that individual videos be (manually) recorded of each well and then be processed offline. Hence no affordance is made for automatically labeling the output data, either by well or using plates barcodes. This makes their tools, as currently implemented, unsuitable for use in a medium- or high-throughput assay. Smout et al. describes an apparatus that does not use an optical assay but instead uses special microtiter plates (xCELLigence and E-plate, Roche Inc.) to measure movement [17]. Recently, efforts have been made for screening Schistosoma spp. based on the quantification of multiple phenotypic responses of the parasites to drugs (as opposed to motility or a biochemical activity) [25]. However, these methods are also not yet available as a ready-to-screen software application or generalized to other macroparasites.
We have developed an inexpensive system (“WormAssay”) for quantifying parasite movement based on worm motility. The apparatus uses a commodity video camera, computer and a newly developed free and open source software application to provide quantitative measurements of parasite motility on entire plates. The application can process multiple wells simultaneously without user interaction, and automatically identifies each well in the plate and labels the output data accordingly. This system can be used to assay large parasites such as the filarid nematodes as well as other macroparasites. WormAssay's automation of the video capture step and lack of need for any interaction with the computer software during scoring differentiates it from all existing motion-based schemes, and permits current screening of 400 worms per assay each week.
Individual adult Brugia malayi female worms (TRS Labs Inc., Athens, GA) were assayed in RPMI-1640 (25 mM HEPES, 2 g/L , Antibiotic/Antimycotic, 5% HI FBS) in 24-well tissue culture plates (1 worm/well). 30 mM stock solutions of albendazole (methyl 5-(propylthio)-2-benzimidazolecarbamate, Sigma), ivermectin (22,23-dihydroavermectin B1, Sigma) and fenbendazole (methyl 5-(phenylthio)-2-benzimidazolecarbamate, Sigma) were prepared with DMSO (Sigma) and serially diluted in media into concentrations of , , , , . DMSO was used as the control and each concentration was run in triplicate. Plates were maintained in a 37C 5% incubator for 48 hours. data were calculated using Microsoft Excel (Microsoft Corp.) and Prism 5 (GraphPad Software, Inc.).
The assays were performed using the open source computer software program described here. This program is named WormAssay. Plates were visualized using a Canon HV-40 Vixia HDV camcorder (Canon Inc.) providing 1080p H.262/MPEG-2 Part 2 compressed HDV video connected via IEEE1384 to an Apple iMac with a 2.93 GHz Intel Core i7 4-core CPU (Apple Inc.) and WormAssay (version 0.15) for 1 minute. The application and source code are available for free use, modification and redistribution under the terms of the GNU Public License (version 2 or later; see http://www.gnu.org/licenses/gpl-2.0.html). The application and its source code can be downloaded from http://code.google.com/p/wormassay/.
The WormAssay was developed for use in high throughput screening of Brugia malayi adult female worms in 24-well plates. Plates were screened using the visual imaging system (Figure 1 and 2) and software application to assess drug effects on adult female Brugia malayi. One minute video recordings using the Lucas-Kande Optical Flow algorithm (see Analysis Algorithms) were taken of each plate and mean optical flow movement units for each worm were converted to percent inhibition: 15 movement units = 0%, no inhibition (worms are very active); 10–15 movement units = 25%, slight inhibition (worms are active); 5–10 movement units = 50%, moderate inhibition (worms slightly moving); 2–5 movement units = 75%, good inhibition (worms barely moving); 0–2 movement units = 100%, very effective killing (worms are dead), using Microsoft Excel and the CSV files generated by WormAssay.
data for albendazole, ivermectin and fenbendazole were calculated for each compound after 48 hours of incubation (see Figure 3 and Videos S1, S2 and S3). Visual analysis of video recordings and data indicate that ivermectin () was the most effective in killing worms compared to albendazole () and fenbendazole (). After 48 hours, only control worms (1% DMSO only) were highly active. In the ivermectin plate, only one of the 3 replicate worms assayed in was found to be barely moving while the other 2 were not moving at all. Worms assayed with albendazole, however, were found to be active even at . After 48 hours of incubation at , one of the 3 worms exhibited activity after 48 hours, while the second replicate was barely moving and the third replicate was dead. Indeed, the of albendazole was 100-fold higher compared to that of ivermectin. Worms assayed with fenbendazole at and were all dead but worms in and concentrations were found to be active after 48 hours.
WormAssay's data acquisition does not require user interaction or configuration and is suitable for robotics integration with any multi-degree-of-freedom plate manipulator. Data acquisition automatically begins when a plate appears in the field of view of the camera and data is written immediately upon removal of the plate (Videos S4, S5 and S6). Video recordings of each read are archived. Motility and other assay data are written to CSV (spreadsheet style) files for use with standard statistical analysis software tools. Barcode reading is performed on the video stream (or from another video camera attached) to automatically label results. The application can automatically email results at the end of a run, for example, when used in a unattended automated assay.
An extended finite state machine [26] is modeled programmatically for each connected camera (see Figure 4). The finite state machine describes the continuous analysis logic of WormAssay that is used to automatically start recording data when a plate is presented to the camera's field of view. If more than one camera is attached to the computer, only the first plate identified on a camera is used for analysis. Other cameras are ignored until that read is complete and the plate is removed from the field of view, except for the purpose of plate identification, where all cameras are inspected simultaneously for common barcode formats. This allows one camera to be used for the motility assay and one or more other cameras to be used for optional barcode recognition. Any barcode text found is used to label the plate output in the CSV files that are output containing the motility data.
The application has no knowledge of the specific geometry of the microtiter plates, except for the number of wells in each row and column of the supported plate sizes. This allows for great tolerance in terms of specific plate geometry and in the position of the plate within the camera's field of view. This is contrary to the scheme used by most microtiter plate assay equipment, where a mechanical sensor is positioned over or in the well of interest, one well at a time.
The WormAssay well finding algorithm iterates through acceptable plate configurations in parallel, corresponding to 6-, 12-, 24-, 48- and 96-well microtiter plates. First, Canny's algorithm [27] is used to find edge features. The Hough (circle) transform [28] is used to find candidate wells efficiently among these edge features. The search for appropriate circles is performed efficiently by limiting the candidate circle size to correspond to the expected range of sizes for a well, under the assumption that the plate fills a simple majority of the imaging field. The resulting circles are then filtered to only accept those that lie on an axis-aligned collinear grid corresponding to the plate row and column configuration. If all expected wells are then found for the plate configuration, the algorithm deems the plate found and moves the state machine down the found edge. If a plate is not found, then the frame image is preprocessed by linear amplifying the high-frequency components of the image. This makes it possible for the Canny edge finding step of the Hough transform to detect the edges that correspond to the well circles in out of focus or poorly illuminated images. If a plate has already been detected, future detection attempts only search for that plate configuration (well count), to reduce CPU resource utilization. Finally, the application uses the location of the wells to label each well canonically (e.g. A1, B1, etc.) in both the real-time screen preview (teal colored text in Figure 2b and 5b, also see Videos S4, S5 and S6) or in the output CSV file data.
We developed two analysis algorithms. The first determines the average velocity of the moving contours inside each well. This algorithm derives the velocity from the optical flow vectors of the luminance component of the video stream from a pair of adjacent frames approximately 100 ms apart. The algorithm uses the sparse iterative version of the Lucas-Kanade optical flow in pyramids provided with the OpenCV framework [29]–[31]. The set of pixels to be considered when calculating the optical flow rate in the current frame is limited in order to make the analysis computationally feasible. This set of points is chosen at random from the set of points that lie on the edge contours found within the well using Canny's algorithm (colored blue in Figure 2b and 5b). Some of these points will correspond to the well itself or other artifacts, but these pixels will not possess a positive optical flow, and will have no impact on the result, as the rate is only determined based on the pixels which are determined to be moving.
This algorithm is useful for scoring rates of motion (or lack thereof) of single parasites with high accuracy as it can reliably differentiate small differences in velocity which may correspond to differing amounts of motility inhibition. A velocity in single dimensional pixel units per second is reported. Only moving components are considered, so this assay is not suitable for assays where a combination of dead (motionless) and moving parasites are present in a single well, since only the moving parasites will be considered in the score. This algorithm is described in WormAssay's Options-Analyzer menu as “Lucas-Kanade Optical Flow.”
The second algorithm is an algorithm that detects changes in the occupation and vacancy of pixels between a group of frames. It uses difference information between a subset of 5 frames chosen at random from the frames that arrived in the past second. First, a difference is performed on each of the 3 color channels of each of the 5 frames and the current frame. Then high frequency components are removed from each of the set of 5 difference values. A voting scheme is employed to determine when a pixel has had its contents changed. Three or more changed pixels is deemed a quorum, otherwise the changes are ignored and deemed noise. The number of filled or vacated pixels is then summed and taken as a fraction of the total number of pixels within the well's circle (times 1000 to improve numerical readability). This number is reported as an arbitrary area unit indicating motility. This algorithm is useful for detecting very low levels of movement or for quantifying the aggregate movement of more than one parasite in a given well. This algorithm is described in the application as “Consensus Voting Luminance Difference.”
All algorithms process in real-time, in parallel on each well. The algorithm programming model is extensible; new algorithms can be added independently of other components of the application. To avoid recording spurious values when the plates are being moved at the beginning or end of a run, the software ignores any frames whose total motion (via the pixelwise mean of the simple interframe absolute difference across all color channels) exceeds a threshold. This value may need to be modified for assays with very large or motile organisms. This is the only non-general threshold used in the application. Improving this aberrant (whole plate) motion detection is a possible area of further research.
Both algorithms (and the well detection) are computationally intensive, and are not able to process every frame of the 1080p (1080×1920 pixels) video input, which is typically 24 or 30 frames per second. On a modern (2011) typical multicore desktop computer, we are able to process 5–10 frames per second, which yields satisfactory results. Since recording of all wells is done in parallel, this is significantly faster than the 5–10 minute recoding times necessary to generate even short 10 frames per second movies on a well-by-well basis on commercial plate microscopes (e.g. on the GE IN Cell Analyzer 2000.)
We also developed a dark-field parallel macroscopic imaging apparatus connected to an HDV camera with an IEEE1394 interface using inexpensive materials [32] (Figure 1). The apparent observable quality of the video recordings improves with greater levels of contrast of the worm with the background. A dark-field imaging scheme provides the most striking contrast, which gives the computer application a greater level of signal-to-noise to analyze. We found that a dark-field imaging scheme, where the plates are illuminated from the side with a uniform light source (in our case white LEDs) and a dark backdrop at least 5 cm from the focus plane of the plate, provides ideal images for recording and analysis. Due to the well finding mechanism, the plates must fill the majority of one of the imaging axes, although there is considerable room for error.
The apparatus used consists of a light-tight box with a hinged lid on the top, with the video camera mounted outside (to ensure easy access and proper cooling) at the bottom of the box and recording upwards. The whole box is made of plywood with some metal parts, all painted black to minimize reflections. The plate is positioned above the camera at such a distance that allows the plate image to fully fill the field of view (approx. 35 cm). The plate is illuminated by a dimmable white LED strip (Home Accent Lighting Kit, White, PPA International) mounted parallel with the plate walls at a distance of 25 mm.
The assay is very sensitive to inadvertent plate motion and illumination that moves or is poor. Hence, it is important to shield the recording field from ambient light so that the operator's movement does not cast a moving shadow on the field of view.
One of the major stumbling blocks in identifying candidate drugs for the treatment of lymphatic filariasis and river blindness is the lack of a high throughput screening system for these large worms. The filarid nematodes are long and threadlike and cannot be easily assayed in a 96-well format. We therefore developed an automated imaging system in which Brugia malayi could be assayed in 24-well plates using a simple and inexpensive method called the WormAssay.
The WormAssay is a visual imaging system that utilizes a novel software program to capture video recordings to assay the effect of compounds on macroparasites. To test the robustness of the software program, we assayed Brugia malayi female worms with 3 antihelminthic drugs: albendazole and fenbendazole (benzimidazoles) and ivermectin (macrocyclic lactone). Albendazole is widely used to treat intestinal nematode infections including ascariasisis and hookworm; ivermectin is used in mass drug administration to treat filariasis; and fenbendazole is used in the veterinary field to treat animals infected with intestinal parasites. data using these compounds with adult filarids are not available. However, the study by Tompkins, Stitt and Ardelli (2010) showed that when adult male and female Brugia malayi were exposed to ivermectin at for 3 days, the motility (as measured in movements/minute) decreased from 250 units to 125 units which is approximately equivalent to the () in our study using ivermectin (molecular weight = 875.1 u) for 2 days [9]. Townson et al. (1990) showed that adult male Onchocerca gutturosa exposed to of ivermectin in the course of 7 days had greatly reduced motility levels (based on mean motility scores from 0–10) compared to controls [13]. Motility scores for male worms exposed to albendazole for 2 days were similar to those for their control worms. Although Townson et al. used Onchocerca adults in their study, their results for both albendazole and ivermectin are consistent with our data.
We are currently using the visual imaging system to screen approximately 400 adult Brugia females per assay. Along with our visual imaging system, we use a Biomek FX (Beckman Coulter) instrument to remove media from each well and dispense compounds. It takes approximately 15 minutes per plate of 24 worms (1 worm/well) to screen compounds at a single concentration and approximately 20 minutes per plate for an . The system is capable of screening more worms but assay throughput is currently limited by the number of worms produced and delivered. Once we receive the 24-well plates containing individual adult female Brugia, we estimate that it takes one person approximately 6–7 hours to setup an assay to screen 96 compounds (at single concentrations) using the Biomek FX and run the WormAssay (on Day 0). Plates are assessed every day for 3 days using the visual imaging system which takes approximately 15–20 minutes for 16 plates. Control worms under these conditions remain highly active while worms treated with low micromolar concentrations of ivermectin are killed as evidenced by the lack of motility. We have observed that the lack of motility is correlated with worm death; dead worms appear more opaque (in some cases are slightly tanned) and never regain motility.
Rather than using laborious and subjective methods of analyzing plates (manual examination of individual wells and plates with a dissecting scope and scoring worm movements relative to control worms), the WormAssay quantified each worm's movement simultaneously on the entire plate, with each plate taking approximately 30 seconds to 1 minute to read. Given the short read times, researchers can increase the number of replicates per compound, thus increasing the accuracy of the assay. Currently, the system requires an individual to place the plate into the visual imaging box but this system is amenable for use with a robotic arm, removing and replacing plates to and from a plate hotel. The software application also includes bar code reading capabilities and can easily be exported to spreadsheets for data analysis.
WormAssay is a unique high-throughput screening motility assay that performs a parallel analysis on each well of entire plates simultaneously, but is independent of specific plate geometry and parasite morphology. The application supports 6-, 12-, 24-, 48- and 96-well plates. WormAssay does not track specific organismal characteristics so it can assay the motility of a large range of macroscopic organisms that can be cultured in a microtiter plate, but is capable of tracking very small or refined movements. The assay requires commodity computer equipment and is compatible with a variety of HD 1080p (or greater resolution) cameras and video capture interfaces. This low-cost and simple-to-use system can also be applied to other target organisms as well. Movements of other macroparasites, including adult schistosome worms were also assessed (see Figure 5), and studies with other macroorganisms are currently being explored.
In summary, the WormAssay offers several advantages: 1) it is inexpensive with costs of the video camera, LED lights and camera totaling less than $3,000 USD and the software is freely available, 2) it is easy to use, i.e. the plate can be quickly placed into the box housing the video camera and removed, 3) video recordings are saved onto the computer along with the data and can be reanalyzed at a later time, 4) entire plates with 6-, 12-, 24-, 48- and 96-wells can be assayed simultaneously, 5) the phenotype (worm movement) is quantified and stored as CSV files and 6) can be more generally applied to the study of macroparasites or other macroscopic organisms.
|
10.1371/journal.pcbi.1000972 | Numerical Analysis of Ca2+ Signaling in Rat Ventricular Myocytes with Realistic Transverse-Axial Tubular Geometry and Inhibited Sarcoplasmic Reticulum | The t-tubules of mammalian ventricular myocytes are invaginations of the cell membrane that occur at each Z-line. These invaginations branch within the cell to form a complex network that allows rapid propagation of the electrical signal, and hence synchronous rise of intracellular calcium (Ca2+). To investigate how the t-tubule microanatomy and the distribution of membrane Ca2+ flux affect cardiac excitation-contraction coupling we developed a 3-D continuum model of Ca2+ signaling, buffering and diffusion in rat ventricular myocytes. The transverse-axial t-tubule geometry was derived from light microscopy structural data. To solve the nonlinear reaction-diffusion system we extended SMOL software tool (http://mccammon.ucsd.edu/smol/). The analysis suggests that the quantitative understanding of the Ca2+ signaling requires more accurate knowledge of the t-tubule ultra-structure and Ca2+ flux distribution along the sarcolemma. The results reveal the important role for mobile and stationary Ca2+ buffers, including the Ca2+ indicator dye. In agreement with experiment, in the presence of fluorescence dye and inhibited sarcoplasmic reticulum, the lack of detectible differences in the depolarization-evoked Ca2+ transients was found when the Ca2+ flux was heterogeneously distributed along the sarcolemma. In the absence of fluorescence dye, strongly non-uniform Ca2+ signals are predicted. Even at modest elevation of Ca2+, reached during Ca2+ influx, large and steep Ca2+ gradients are found in the narrow sub-sarcolemmal space. The model predicts that the branched t-tubule structure and changes in the normal Ca2+ flux density along the cell membrane support initiation and propagation of Ca2+ waves in rat myocytes.
| In cardiac muscle cells, calcium (Ca2+) is best known for its role in contraction activation. A remarkable amount of quantitative data on cardiac cell structure, ion-transporting protein distributions and intracellular Ca2+ dynamics has been accumulated. Various alterations in the protein distributions or cell ultra-structure are now recognized to be the primary mechanisms of cardiac dysfunction in a diverse range of common pathologies including cardiac arrhythmias and hypertrophy. Using a 3-D computational model, incorporating more realistic transverse-axial t-tubule geometry and considering geometric irregularities and inhomogeneities in the distribution of ion-transporting proteins, we analyze several important spatial and temporal features of Ca2+ signaling in rat ventricular myocytes. This study demonstrates that the computational models could serve as powerful tools for prediction and analyses of how the Ca2+ dynamics and cardiac excitation-contraction coupling are regulated under normal conditions or certain pathologies. The use of computational and mathematical approaches will help also to better understand aspects of cell functions that are not currently amenable to experimental investigation.
| Ventricular cardiac muscle cells have deep invaginations of the extracellular space known as t-tubules [1]–[14]. In rodents, these invaginations branch within the cell to form a complex network that allows rapid propagation of the electrical signal (i.e. the action potential, AP) to the subcellular location (i.e. the sarcoplasmic reticulum, SR) where the intracellular Ca2+ required for the cell contraction is stored [14]. The release of Ca2+ from the SR depends on “trigger Ca2+” entering the cytosol from the extracellular space by activating sarcolemmal Ca2+ channels (L-type Ca2+ channels, LCC) and by Ca2+ entry via Na+/Ca2+ exchanger (NCX), [3], [9], [14], [15]. The trigger Ca2+ activates SR Ca2+ release channels (ryanodine receptors, RyRs) by the process of “Ca2+-induces Ca2+-release” (CICR) which amplifies the modest increase in intracellular Ca2+ concentration ([Ca2+]i) caused by the LCC and NCX influxes to provide sufficient Ca2+ for the proteins regulating muscle force (i.e. troponin C, TN) ), [14]. Thus, by working together, the microanatomy of t-tubules and SR permits spatially homogeneous and synchronized SR Ca2+ release and spatially uniform Ca2+ transients throughout the cell [5], [14], [16]. It has been also observed that the spatially uniform Ca2+ transients might be achieved if the SR Ca2+ release and uptake are abolished [5]. Yet, despite a wealth of information on ventricular cell function and structure, the mechanisms causing the synchrony of activation and the similarity of levels of [Ca2+]i across the myocyte still remain unclear.
In cardiac muscle cells, several computational models have been introduced to investigate the Ca2+ signaling, buffering and diffusion [17]–[19] and Ca2+ wave initialization and propagation [12], [20]–[23]. All these studies, however, are conducted on simplified geometries (such as cylindrical or rectangular shapes) and it has been pointed out that a small geometric change (even in the case the change is uniformly applied) could greatly influence the suggested homogeneous Ca2+ distribution by initiating wave propagation in the computer simulation [20], [22]. Several laboratories, using common pool modeling approaches, have investigated also the effects of LCC and NCX distributions on global [Ca2+]i transients in dyadic, sub-sarcolemmal and cytosol compartments [10], [24], [25]. Recently, to examine how the distribution of Ca2+ flux along the sarcolemma affects Ca2+-entry and Ca2+ diffusion and buffering, we developed a 3-D continuum model in rat ventricular cells [19]. An important limitation of this model is that a cylindrical t-tubule geometry was assumed while several studies have provided evidence that in rodent ventricular myocytes the realistic t-tubule geometry is quite complex (with large local variations in the diameter and transverse-axial anatomies), [9]–[12]. These experimental findings suggest that replacing our idealistic t-tubule model with a realistic geometry is needed. The use of idealistic shapes will change the diffusion distances and realistic Ca2+-transporting protein localizations in plane and depth directions and consequently the predicted [Ca2+]i distributions.
In the present study, we sought to develop a morphologically correct geometric model of the t-tubule and to use this model for computational studies of the intracellular Ca2+ dynamics. We examined the Ca2+ signaling in rat ventricular myocytes that had been treated with ryanodine and thapsigargin to eliminate Ca2+ release and uptake by the SR. By using published electro-physiological data and laser-scanning confocal [Ca2+]i measurements, we were able to analyze several important spatial and temporal features of the Ca2+ signals in these cells. In this context, our goal was at least three-fold. The first aim was to develop a mathematical model that would be in qualitative or quantitative agreement with published experimental measurements on Ca2+ influx, and Ca2+ buffering and diffusion in rat ventricular cells with SR function inhibited [5], [26]. Second, to use the model to investigate the importance of t-tubule ultra-structure and membrane Ca2+ flux distribution for the Ca2+ signals. The third task was to simulate the Ca2+ signals in the absence of fluorescent dye and to study the importance of the mobility of endogenous Ca2+ buffers (ATP and calmodulin) and altered extracellular Na+ ([Na+]e) for the Ca2+ signals. The analysis suggests that the quantitative understanding of the Ca2+ signaling requires more accurate knowledge of the t-tubule microanatomy and Ca2+ flux distribution along the sarcolemma. In agreement with experiment, with 100 µM Fluo-3, the lack of detectible differences in the depolarization-evoked Ca2+ transients was found when the Ca2+ flux was heterogeneously distributed along the sarcolemma. In the absence of Fluo-3, the predicted Ca2+ signals were strongly non-uniform. Even at modest elevation of Ca2+, reached during Ca2+ influx, large and steep Ca2+ gradients may develop in the narrow sub-sarcolemmal space. The model also predicts that branched t-tubule topology and changes in the normal Ca2+ flux density along the cell membrane support Ca2+ waves initiated at the sarcolemma. Preliminary results of this work have been presented to the Biophysical Society in abstract form [27].
Combining light microscopy (LM) and electron microscopy (EM) together with 3-D tomographic reconstruction, Hayashi et al. [6] investigated 3-D topologies of important sub-cellular organelles, including dyadic clefts and t-tubules, in mouse ventricular myocytes. In particular, the use of two-photon microscopy (T-PM) in their studies had provided data showing detailed spatial organization of t-tubules (see Fig. 1A upper panel) that was important for the development of our realistic model for computational studies of intracellular Ca2+ dynamics. The gap between imaging and simulation involves two major steps: (1) extracting features (boundary or skeleton) from imaging data; (2) constructing geometric models (represented by meshes) from the detected features. In addition, image pre-processing is usually necessary for better feature extraction, when the original image is noisy or the contrast between features and background is low. With 3-D T-PM images, Yu and collaborators developed a set of image processing and analysis tools and using the mesh generator called GAMer [28] they were able to generate high-fidelity and quality meshes for 3-D t-tubular systems in mice [29] (see Fig. 1A lower panel).
The extreme intricacy of the t-tubule system in mice (with transverse-axial anatomies and large local variations in t-tubule diameter) has been observed in rat ventricular myocytes as well [4], [11], [30]. Because high-fidelity geometric models representing the realistic t-tubule topology in rats are currently not available, in this study we used the geometric model in mice of Yu's et al. [29]. To investigate the Ca2+ signaling in rat ventricular myocytes, we considered a small compartment containing a single t-tubule and its surrounding half-sarcomeres for two reasons: (a) the entire t-tubular system in a ventricular myocyte forms a roughly periodic pattern corresponding to individual sarcomere; and (b) a small model contains much fewer number of mesh nodes that would render numerical simulation significantly faster and more feasible in ordinary computers. The surrounding half-sarcomeres were modeled as a rectangular-shaped box of 2 µm×2 µm in the plane of external sarcolemma and 5.96 µm in depth (Fig. 1B left panel and Table 1). As Yu's t-tubule model did not include the realistic cell surface, one of the box faces (the top red surface in Fig. 1B) was assumed to be the external cell membrane. The t-tubule inside this compartment was extracted from a sub-volume of the T-PM imaging data corresponding to the region indicated in Fig. 1A lower panel. To make it easier to handle boundary conditions in numerical analysis, we have “closed” the end of each branch, yielding a tree-like t-tubule model (see the yellow mesh in Fig 1B). These added “caps” were treated with the same boundary conditions as for the rest of the t-tubular surface. This simplified treatment clearly could introduce some errors because these “caps” are artificial and no Ca2+-transporting protein should reside there. However, the errors should be negligible as the area of these “caps” is very small, compared to the rest of t-tubular surface. The t-tubule diameter varied from 0.19 µm to 0.469 µm and the t-tubule depth was 5.645 µm. The volume of the model compartment was estimated to be ∼23.31 µm3. The compartment membrane area was ∼9.00 µm2 where the percentage of cell membrane within t-tubule was 64% (∼5.75 µm2) and within the external membrane 36% (∼3.25 µm2), [10], [11], [31], [32]. The accessible volume for Ca2+ was estimated to be ∼35–37% of the total cytosolic volume () ∼12.9–13.6 pL in adult rat ventricular myocytes [33], [34]. The sub-cellular aqueous volume of 35–37% assumes that: (1) myofilaments occupy 47–48% of the cell volume, mitochondria 34–36%, nucleus 0–2%, t-tubule system 0–1.2%, and SR lumen 3.5%; (2) 50% of the myofilament space is accessible for Ca2+ (i.e. contains water); (3) mitochondria and nuclei are not rapidly accessible for Ca2+; (4) the SR lumen is not accessible to Ca2+ in the presence of ryanodine and thapsigargin [18], [33].
Recent immunohistochemical studies have demonstrated that marked variations in the distribution of Ca2+-transporting protein complexes (L-type Ca2+ channel, Na+/Ca2+ exchanger) along the cell membrane probably exist [2], [10], [12], [15], [35]–[43]. The analysis suggests that most of the L-type Ca2+ channels are concentrated in the t-tubules (from 3 to 9 times more in the t-tubule membrane than on the external sarcolemma) [2], [10], [12], [38], [39] and that the concentration of LCC along the t-tubule probably increases toward the center of the cell [36], [40].
Studies on the distribution of the main Ca2+ efflux pathway, the Na+/Ca2+ exchanger, are more controversial. All studies but one [41] have reported NCX to localize both to the external and t-tubule membrane, but most studies suggest that the NCX is 1.7 to 3.5 times more concentrated in the t-tubule membrane [15], [34], [36], [42], [43]. However, Kieval et al. data [43] indicate the NCX is more evenly distributed. In summary, the observed differences in the spatial distribution and molecular architecture of Ca2+ microdomains suggest that significant differences in the excitation-contraction coupling between the cell surface and cell interior may be exist. However how the localization of Ca2+- transporting protein complexes along the sarcolemma regulates the intracellular Ca2+ signaling still remains uncertain.
In the current model, the effects of four exogenous and endogenous Ca2+ buffers (Fluo-3, ATP, calmodulin, troponin C) were considered (Fig. 1C). The endogenous stationary buffer troponin C (TN) was distributed uniformly throughout the cytosol but not on the cell membrane and in the sub-sarcolemmal space (∼40–50 nm in depth). The free Ca2+ and mobile buffers (Fluo-3, ATP, calmodulin) diffuse and react throughout the cytoplasm. The cell membrane and sarcomere box faces are subject to reflective boundary conditions. The nonlinear reaction-diffusion equations describing Ca2+ and buffers dynamics inside the model cell are:(1)(2)(3)(4)(5)where: [Bm] represents the concentration of mobile buffer Fluo-3, ATP or calmodulin; [Bs] is the concentration stationary buffer troponin C.
The diffusion coefficients for Ca2+, CaATP, CaCal and CaFluo as well as the total buffer concentrations and buffer rate constants used in the model are shown in Table 2. In the model we also assume: (1) isotropic diffusion for Ca2+ and all mobile buffers [12]; (2) Ca2+ binds to Fluo-3, calmodulin, ATP, and TN without cooperativity; (3) the initial total concentrations of the mobile buffers are spatially uniform; (4) the diffusion coefficients of Fluo-3, ATP or calmodulin with bound Ca2+ are equal to the diffusion coefficients of free Fluo-3, ATP or calmodulin.
The total Ca2+ flux () throughout the t-tubule and external membrane is:(6)where: - total LCC Ca2+ influx; - total NCX Ca2+ flux; - total membrane Ca2+ leak.
To describe L-type Ca2+ current, Na+/Ca2+ exchanger, Ca2+ leak current densities we used the following expressions:(7)(8)(9)(10)
Flux parameter values were estimated or taken from the literature (see Table 3). In this study, the Ca2+ leak is not actually a particular “leak protein”. The Ca2+ leak was included and adjusted so that at rest Ca2+ influx via Ca2+ leak to match Ca2+ efflux via NCX thus no net movement across the cell membrane to occur.
In the model, each current density (Ii) was converted to Ca2+ flux (Ji) by using the experimentally suggested surface to volume ratio (∼8.8 pF/pL) in adult rat ventricular myocytes [32], [33]:(11)Then, the total compartment Ca2+ flux () was computed by multiplying each total Ji with the model cell volume (), and distributing to the external and t-tubule membrane according to the prescribed Ca2+-handling protein concentration ratio.
The voltage-clamp protocol (holding potential −50mV, electric pulse of 10mV for 70ms) and whole-cell L-type Ca2+ current were derived from Zahradnikova et al. data with the blocked SR activity [26]. In this study, each simulation started with a basal cytosolic Ca2+ of 100 nM, basal cytosolic Na+ of 10 mM and buffers in equilibrium. The extracellular Ca2+ concentration () was 1 mM and remained constant. Unless specified otherwise in the Figure legends or in the text, the extracellular Na+ concentration () was 140 mM and 3.4−6 µM mV−1 ms−1.
In finite element methods, a complex domain needs to be discretized into a number of small elements (such as triangles or tetrahedra). This process is usually referred to as mesh generation [28], [44]. Although different types of meshes may be generated depending on the numerical solvers to be employed, we restrict ourselves to triangular (surface) and tetrahedral (volumetric) mesh generation as commonly used in biomedical simulation. In the present simulation, the number of finite element nodes and tetrahedral elements are 50,262 and 221,076, respectively.
The nonlinear reaction diffusion system was solved by a finite difference method in time and finite element method in space using our SMOL software tool (Smoluchowski Solver, http://mccammon.ucsd.edu/smol/) with the time step of 4 ms. It takes around 20 minutes to run 400 ms snapshots with a single Intel Xeon X5355 processor. The SMOL program utilizes libraries from the finite element tool kit (FETK), which previously has been used in several molecular level studies [45]–[49]. One bottleneck for dynamic 3-D simulation of nonlinear reaction diffusion system is the computing complexity involved in solving the problem. Here we successfully extended SMOL to solve multiple coupled partial differential equations with nonlinear ordinary equations. Multiple tests demonstrate that our SMOL program is quite robust and flexible for various boundary and initial conditions. The simulation results were visualized using GMV mesh viewers [50]. Post-processing and data analyses were implemented by customized Python, MATLAB 2008b (The MathWorks, Natick, MA) scripts and Xmgrace software [51]. A version control system, subversion, was used to monitor the development of software [52].
In agreement with the reported experimental data [2], [10], [12], [36], [38]–[40], the spatial patterns of [Ca2+]i were calculated assuming LCC current density: (1) heterogeneously distributed along the cell surface; (2) six times higher and uniform in the t-tubule membrane; or (3) homogeneously distributed along the sacrolemma. In cases (1–2) the NCX flux density was assumed three times higher in the t-tubule and in case (3) NCX was evenly distributed along the sarcolemma. In this study, Ca2+ leak density was homogeneously distributed along the cell membrane with respect to all distribution choices of LCC and NCX. In case (1), the 3-D distribution of LCC current was computed by combining the cluster density and fluorescent intensity plots, see Fig. 2A. The data were then scaled and fitted by a cubic polynomial:(12)where: x is the distance from the external cell surface.
The parameter values of the polynomial (pj, j = 1–4) are shown in Table 4. This polynomial was further scaled by a single factor C (see Table 4) such that the total Ca2+ flux along the t-tubule membrane remained unchanged by redistributing the Ca2+ fluxes. To fit the whole-cell LCC current density to the reported data in rat myocytes with SR release inhibited [26], we used a shape preserving function, (see Eq. 8 and Fig. 2B).
Consistent with the Cheng et al. experiment [5], where the fluorescence signal was recorded along the single scan-lane starting and ending outside the cell and crossing the center of the cell, the model t-tubule was chosen to cross the cell center and the scanned line was located at 200nm away from the t-tubule membrane (see Fig. 1A and Fig. 3). To gain more detailed insights of how the predicted Ca2+ signals are regulated within this geometrically irregular micro-domain we examined [Ca2+]i at two different line-scan positions: 200 nm, angle 120°; 200 nm, angle 60° (see Fig. 3).
Model results in Figs. 4–5 were computed for conditions approximating those of the experiment by Cheng et al. [5], who examined Ca2+ signals in voltage-clamped rat myocytes in the presence of 100 µM Fluo-3 and pharmacological blockade of the SR (see Fig. 4L). The computed line-scan images and local Ca2+ time-courses are shown in Figs. 4F–4K.
These results demonstrate that with LCC heterogeneous or LCC six times higher in the t-tubule: (1) predicted Ca2+ concentration profiles were non-uniform when t<100 ms but the variations in [Ca2+]i seem to be within the range of experimental noise in Fig. 4L; (2) [Ca2+]i was more evenly distributed when t>100 ms, ( Figs. 4F–G, Figs. 4I–J, Video S1). To delineate further the suggested spatial differences in [Ca2+]i (see Figs. 4F–H), we introduced a quantity called ‘spatial Ca2+ heterogeneity’ (SCH). The SCH is defined to be the difference of the maximal and minimal [Ca2+]i values, normalized by the maximal value at given reference point along the line-scan (0.17 µm, 3.09 µm, 5.45 µm) in given moment tj, (see Figs. 4I–K). High SCH suggests non-uniform [Ca2+]i distribution and SCH of zero indicates spatially uniform [Ca2+]i distribution. The histogram in Fig. 4M shows that assuming LCC heterogeneous versus LCC 6 times higher in the t-tubule decreased SCH(tIca-peak) by 1.6 folds, SCH(t70 ms) by 2.29 folds, SCH([Ca]i-peak) by 2.34 folds, SCH(t100ms) by 2.87 folds, and SCH(t200ms) by 8.45 folds. These findings demonstrate that the predicted [Ca2+]i distribution with LCC heterogeneous more closely resembles the Chang et al. experimental findings [5], (compare Figs. 4I and 4L). Finally, the model predicts strongly non-uniform Ca2+ transients when the LCC, NCX and Ca2+ leak fluxes were uniformly distributed throughout the cell surface (Figs. 4H, 4K, 4M). In addition, Video S1 (see right panel) demonstrates that here the Ca2+ signal spreads from the external membrane to the cell center as a continuum wave but after LCC channel closing (t∼72 ms) this wave faltered.
Predicted global [Ca2+]i transient, Na+/Ca2+exchanger, and Ca2+ leak currents with LCC pathways heterogeneously distributed (as in Fig. 4F) are shown in Figs. 4E and 4C–D. Figure 4C demonstrates that: (1) the depolarization of cell membrane reversed the rest exchanger's direction (i.e. in the interval 0ms–70ms NCX operated in Ca2+ entry mode) while when repolarization occurred the flow of Ca2+ through NCX was reversed again (i.e. in the interval 70ms–400 ms NCX operated in Ca2+ exit mode); (2) upon returning to resting voltage of −50mV the exchanger's rate rapidly increased () while rate remained unchanged (note is not voltage-dependent) thus causing fast extrusion of Ca2+ out of the cell and subsequent sudden drop in the local and global [Ca2+]i. Figures 4F–K illustrate also that the global and all local Ca2+ transients reached the peak after ∼76 ms and that [Ca2+]i levels were higher near the t-tubule mouth because the density of t-tubule branches was higher in this region and close topological proximity of the external membrane additionally increased the relative amount of the entering Ca2+. Due to the higher [Ca2+]i gradient under the outer cell edge (t∼70ms) Ca2+ diffused toward the cell center and when ratio along the cell membrane became approximately equal to ratio a new equilibrium level of [Ca2+]i (∼0.16 µM) was reached. Intracellular Ca2+ equilibrated faster when Ca2+ flux was more concentrated in the t-tubule membrane because [Ca2+]i gradient near the t-tubule mouth was lower there than [Ca2+]i gradient with Ca2+transporting complexes distributed homogeneously. Additional reasons for the observed rapid equilibrium of [Ca2+]i may be that [Na+]i was kept constant (in contrast to existing evidence for the presence of local sub-sarcolemmal [Na+]i gradients on the action potential time-scale [33], [53]) or that the realistic distribution of NCX flux may be differ as assumed in the model. Finally, the results demonstrate that the computed average [Ca2+]i peak of 160–185 nM (see Figs. 4E, 4I–4J), is comparable with the measured one of about 163 nM when the SR release and uptake were inhibited [5].
This model is also able to predict how the Ca2+ transients are regulated at different line-scan positions within this geometrically irregular micro-domain. Note, due to the technical limitations the Cheng et al. experiment is not able to suggest where exactly the scanned line is positioned with regard to the specific t-tubule, but the Cheng et al. measurements [5] strongly suggest the similarity of [Ca2+]i at the peripheral and deeper cytoplasm when the SR activity is abolished. For this reason, we examined the Ca2+ profiles (LCC heterogeneous along the t-tubule, Fig. 4F) positioning the line-scan at 200 nm and angle 60° (see Fig. 3) or positioning the line-scan at 50, 100, 200, 300 or 400 nm at different angles. No visible differences in the visualized spatial Ca2+ profiles were found (data not shown).
The 3-D Ca2+ concentration distributions and spatial Ca2+ profiles at Ca2+ peak (76 ms) are shown in Figs. 5A–D. These model results demonstrate that the Ca2+ concentration near the external membrane decreased while [Ca2+]i in the cell interior increased when Ca2+ transporters were uniformly distributed and after that heterogeneously redistributed. The jumps in Fig. 5D show the predicted local Ca2+ flux () at [Ca2+]i peak in the regions where the scanning line of interest is crossing the t-tubule membrane.
Additional interesting model findings are that: (1) large and steep [Ca2+]i gradients were predicted inside the sub-sarcolemmal 3-D space (see Video S1); (2) the global Ca2+ time-course and time to [Ca2+]i peak did not depend on whether Ca2+ fluxes are distributed homogeneously or heterogeneously along the sarcolemma (data not shown); (3) redistributing NCX flux uniformly via the sarcolemma was not able to alter significantly the predicted Ca2+ signals in Fig. 4F (data not shown).
In this study the value of 390 µm2 s−1 for diffusion coefficient of free Ca2+ and published buffer diffusion coefficients and parameters were used to compare the calculated Ca2+ signals with the Cheng's et al. fluorescence Ca2+ signals recorded in rats [5]. It has been suggest, however, that the effective diffusion of free Ca2+ in the cytosol () will be slowed down because the exogenous and endogenous Ca2+ buffers and free Ca2+ concentrations are able to affect Ca2+ diffusion strikingly [18], [19], [54]–[62]. The measurements of Allbritton et al. [54] report a value of 5–21 µm2 s−1 for at low free when a value of ∼223 µm2 s−1 is assumed for . During Allbritton's et al. in vitro experiments Ca2+ sequestration by the SR, mitochondria and ATP was inhibited and only mobile calmodulin and stationary troponin C were present in the cytosolic extract. It is possible to estimate (0 µM Fluo-3, 0 µM ATP), using a simplified equation (see Eq. 13), because in this study the predicted maximal Ca2+ elevations were sufficiently small () and the diffusion coefficients for Ca2+-bound and free mobile buffer forms were assumed equal [18], [19], [58], [59]. Our calculations predict a value of ∼8 µm2 s−1 for when was 390 µm2 s−1 and ∼6 µm2 s−1 when was 223 µm2 s−1. Therefore, the estimated value for (∼8 µm2 s−1) is in reasonable agreement with the experimental observation.(13)
We could not find experimental data suggesting (260 µM ATP, 70 µM TN, 24 µM Cal) or (100 µM Fluo-3, 260 µM ATP, 70 µM TN, 24 µM Cal) in the solution. Therefore, we used published concentrations of Ca2+ binding proteins and published diffusion and dissociation constants () to estimate the effective diffusion constants of free Ca2+ in the presence of ATP or Fluo-3 and ATP in the cytosol. Our calculations indicate that adding 260 µM ATP in the solution accelerated (∼10.4 µm2 s−1) and that increased additionally when 100 µM Fluo-3 was added ( ∼66 µm2 s−1). Furthermore, our studies suggest that in the presence of 100 µM Fluo-3 and LCC heterogeneous Ca2+ binding and diffusion of ATP and calmodulin could not affected significantly the predicted [Ca2+]i distributions (data not shown).
During simulations of SR Ca2+ release into the diadic cleft, a major effect of the stationary phospholipids Ca2+ binding sites has been suggested [17], [18]. To examine the impact of the phospholipids on the much smaller Ca2+ signals (arising from Ca2+ influx via Ca2+ current only), we included this buffer in our model. The results demonstrated that the phospholipids had only a limited effect on the calculated Ca2+ signals in the sub-sarcolemmal region (0 µM Fluo-3, 260 µM ATP, 24 µM calmodulin) and that this effect was even smaller when 100 µM Fluo-3 was included (data not shown).
The current study attacks a difficult problem on how to incorporate the structural-based biological information, critical for the subcellular and cellular function, into sophisticated computational investigations. Pursuing this goal we developed a 3-D continuum model of Ca2+ signaling in rat ventricular cells that incorporates the realistic transverse-axial t-tubule topology and considers geometric irregularities and inhomogeneities in the distribution of ion-transporting proteins. The t-tubule micro-architecture was extracted from the Hayashi et al. two-photon imaging data in mice [6]. Because currently high-fidelity geometric models representing the realistic t-tubule micro-architecture in rats are not available, in this study we used the Yu's et al. geometric model in mice [28], [29]. On the basis of experimental data in rats the aqueous sub-cellular volume, accessible to Ca2+, was estimated to be ∼35–37% [33]. Since the Ca2+ signaling in cells is largely governed by Ca2+ diffusion and binding to mobile and stationary Ca2+ buffers [17]–[21], [23], [54]–[62], [66], [67], the effect of four Ca2+ buffers (Fluo-3, ATP, calmodulin, TN) was considered.
The model was validated against published experimental data on Ca2+ influx, membrane protein distributions and Ca2+ diffusion in rat cells treated with ryanodine and thapsigargin to inhibit the SR Ca2+ metabolism [2], [5], [10], [12], [15], [26], [35]–[43]. We found that with 100 µM Fluo-3 the results more closely resemble the Cheng's et al. experimental data [5] when the LCC density increases ∼1.7 fold along the t-tubule length and the NCX density is assumed three times higher in the t-tubule. An interesting finding is that with LCC six times and NCX three times higher and uniform in the t-tubule, the predicted fluctuations in the [Ca2+]i profiles were within the range of experimental noise [5]. Strongly non-uniform spatial Ca2+ gradients and propagation of Ca2+ wave are predicted, not observed in Cheng et al. experiment, when the LCC and NCX were uniformly distributed along the sarcolemma.
The model studies with 100 µM Fluo-3 indicate also that the [Ca2+]i gradients depend on the diffusion distances in the axial and cell surface directions. Thus, when the LCC were distributed uniformly the local Ca2+ peak in radial depth (5.96 µm) decreased from ∼1.5 fold while in the other cell directions (1 µm×1 µm) no significant changes were found. Redistributing the amount of Ca2+ pumped via the cell membrane (i.e. increasing LCC current density along the t-tubule) while keeping total Ca2+ flux unchanged, lowered Ca2+ gradients near the surface membrane and increased Ca2+ levels in the cell interior (see Video S1). The results also showed that with 100 µM Fluo-3 and Ca2+ flux heterogeneously distributed along the sarcolemma, the computed average [Ca2+]i peak (160–185 nM) is comparable to the measured of about 163 nM [5] and that the NCX redistribution alone yields to qualitatively similar [Ca2+]i profiles.
It should be noted, that in our previous work we used the simplified t-tubule geometry (assuming cylindrical shape) to simulate the Ca2+ dynamics in rats [19]. This idealistic t-tubule model also predicts the lack of systematic differences in the fluorescence Ca2+ signal when the Ca2+ transporters were distributed heterogeneously along the sarcolemma. Thus, the following question arises: How these new computational studies based on more realistic t-tubule structural model will further advance our current knowledge on the cell excitability and Ca2+ cycling in rats? First, in agreement with experiment [2]–[12] current study confirms that due to the branched t-tubule microstructure high and steep sub-sarcolemmal [Ca2+]i gradients could occur throughout the whole cell volume [2]–[6], [33], (see Video S1). Note, Lu et al. idealistic t-tubule model predicts high and steep sub-sarcolemmal [Ca2+]i gradients only in the transverse cell direction [19]. Second, our realistic t-tubule model predicts non-uniformities in [Ca2+]i distribution along the depth of the t-tubule when t<100 ms (see Fig. 4F and Video S1 left panel) while this was not the case when the t-tubule geometry is assumed cylindrical (Fig. 4g in Lu et al., [19]). Third, interesting finding is that no visible differences in the local Ca2+ profiles are predicted when the line-scan was positioned at different. Note, due to the technical limitations the Cheng et al. experiment is not able to suggest where and how exactly the scanned line is positioned with regard to the specific t-tubule [5].
A surprising and important finding of this study is that the spread and buffer capacity of 100 µM Fluo-3 were able to mask completely the pronounced spatial [Ca2+]i non-uniformities that would occur during the Ca2+ influx in the absence of dye (see Video S2 left panel - SR Ca2+ metabolism inhibited, LCC and NCX transporters heterogeneously distributed). Here the simulations demonstrated that with zero Fluo-3 the local and global Ca2+ peaks increased while the time of Ca2+ rise remained unchanged. The predicted sub-sarcolemmal [Ca2+]i gradients were heterogeneous along the cell membrane and larger and steeper compared to those with 100 µM Fluo-3. The NCX and Ca2+ leak time-courses were also affected due to increased local free [Ca2+]i levels. It is interesting that under these conditions no differences in the local Ca2+ time-courses were found (as with 100 µM Fluo-3) when the line-scan was positioned at different angles and distances. To test further the model we also examined how the mobility of endogenous Ca2+ buffers (ATP and calmodulin) and altered extracellular Na+ ([Na+]e) would affect the Ca2+ signals in the absence of fluorescence dye when the Ca2+-transportes are heterogeneously distributed. The results showed that when ATP and calmodulin were immobilized Ca2+ diffuses slowly toward the center of the cell, resulting in higher Ca2+ concentrations near the outer cell edge. When the NCX forward mode was inhibited (assuming [Na+]e = 0 mM) the local [Ca2+]i peaks increased and this increase was more pronounced near the outer cell edge. New findings are that under these conditions near the outer cell edge Ca2+ wave was initiated while this was not the case when ATP and calmodulin were mobile and [Na+]e 140 mM (see Video S2 and Video S3).
Taken together, these studies provide compelling evidence that (1) the exogenous Fluo-3 acts as a significant buffer and carrier for Ca2+, and that (2) the use of 100 µM Fluo-3 during the experiment can sensitively alter the realistic Ca2+ distribution. A new the question, however, arises: Based on the above model findings what could be the underlying mechanism(s) for the predicted heterogeneous Ca2+ concentrations gradients in the absence of Fluo-3? A reasonable answer is that the Ca2+ movement and distribution inside the cell rely strongly not only on the specific cell micro-architecture and Ca2+ transporters distribution but also on the presence of endogenous mobile and stationary Ca2+ buffers (ATP, calmodulin, troponin C - known to affect strikingly the Ca2+ dynamics) [12], [17]–[21], [23], [27], [67]. In support of this hypothesis, our simulations studies revealed that in the absence of Fluo-3: (1) the stationary Ca2+ buffer troponin C imposed stronger diffusion barrier for Ca2+ that resulted in larger and steeper sub-sarcolemmal Ca2+ gradients; (2) in the cell interior, owing on their sheer buffering capacity, Ca2+ buffers (troponin C, ATP, calmodulin) tended to slow down additionally the propagation of Ca2+ so that ATP and calmodulin spreading alone was not able to contribute the spatially uniform Ca2+ profiles to be achieved; (3) immobilizing the endogenous Ca2+ buffers slowed down the Ca2+ movement from the cell periphery to the center leading to build-up of large sub-sarcolemmal Ca2+ gradients and subsequent initiation of Ca2+ wave. It is important to mention that the Lu et al. idealistic t-tubule model predicts completely different 3-D [Ca2+]i distributions with zero Fluo-3, SR Ca2+ metabolism inhibited and Ca2+ transporters heterogeneously distributed [19].
Important limitations of the current modeling approach are: (1) the relatively small size of the model compartment that contains only a single realistic t-tubule shape and spans by just a half-sarcomere inside the ventricular myocyte; and (2) the assumption that the model compartment is a repeating unit inside the cell. The structural studies, however, provide evidence that in rodent ventricular myocytes the realistic t-tubule network is quite complex, (see Fig. 1A), [6]. The above limitations can be overcome in the future by extending the current geometric model toward more realistic models containing several t-tubules, whole-cell t-tubule network or other sub-cellular organelles (including mitochondria, SR, nuclei). This would allow building an improved geometric models representing more correctly the cell segment of interest and help to gain further insights of how the Ca2+-signaling in rat ventricular myocytes is regulated in the absence or presence of SR Ca2+ release and uptake [20]–[23], [66], [67]. However, it is equally important to mention here, that although the limitations (1–2) this model in a first approximation may yield insights across the whole-cell scale of biological organization. It allows simulating the global Ca2+ signal (computed from the line-scan image in Fig. 4F) that roughly would reproduce the whole-cell Ca2+ transient in the Cheng et al. experiment due to observed spatial similarities in [Ca2+]i (see Fig. 1B–1C in [5]). This assumption enables also investigating how the whole-cell Ca2+ signal is regulated by the realistic t-tubule microanatomy, by 3-D distributions of ion-transporting proteins, by mobile dye or endogenous mobile and stationary Ca2+ buffers. It should be noted, that the common pool modeling approaches could not be used to investigate these effects [10], [24], [25].
Important limitation of this study is also that we assume that the ion flux pathways are continuously distributed throughout the t-tubule membrane. Immunohistochemical studies, however, suggest that L-type Ca2+ channels appear to be concentrated as discrete clusters in the dyadic clefts (narrow spaces between LCC and RyR) distributed regularly along the t-tubule membrane at relatively small distances of ∼0.68 µm, [12], [38], [68]. It is interesting to mention that in contrast to Soeller and collaborators data in rats [12], Hayashi et al. data in mice [6] suggest that the dyadic clefts are distributed irregularly along the t-tubule branches. In addition, NCX appears to be absent from the longitudinal tubules [42]. Thus, the above data clearly imply that localized concentration of LCC or NCX flux pathways could result in larger sub-sarcolemmal Ca2+ gradients and local membrane currents that will affect differently the spatial Ca2+ profiles as predicted with the current model. Further extending of our current t-tubule model toward more realistic geometric models containing dyadic cleft topology and L-type Ca 2+ channel clustering could help to better understand how the Ca2+ signaling is regulated in the heart.
Finally, in the present model the effects of two endogenous Ca2+ mobile buffers (ATP, calmodulin) and one stationary Ca2+ buffer (troponin C) were considered only. The ventricular cells, however, contain other stationary Ca2+ buffers (including phospholipids, myosin, calsequestrin) or small and high mobile Ca2+ binding molecules (ADP, AMP) that were not included in the model (or may be other stationary and mobile buffers that have not been identified yet), [18], [24], [69].
Simulations presented in this study demonstrate that the more accurate knowledge of transverse-axial t-tubule microanatomy and protein distributions along the sarcolemma is important to better understand the mechanisms regulating the excitation-contraction coupling in rat ventricular myocytes. The results demonstrate that Ca2+ movement from the cell membrane to the cell interior relies also strongly on the presence of mobile and stationary Ca2+ buffers, including the Ca2+ indicator dye. The key findings are: (1) the model predicts lack of detectible differences in the fluorescence Ca2+ signals at the peripheral and deep myoplams when the membrane Ca2+ flux is heterogeneously distributed along the sarcolemma; (2) 100 µM mobile Fluo-3 is able to mask the pronounced spatial non-uniformities in the [Ca2+]i distribution that would occur when the SR Ca2+ metabolism is inhibited; (3) during the Ca2+ influx alone, large and steep Ca2+ gradients are predicted in the narrow sub-sarcolemmal space (∼40–50 nm in depth); (4) in rodents the branched t-tubule topology, the punctuate spatial distribution of Ca2+ flux along the sarcolemma and the endogenous Ca2+ buffers actually function to inhibit Ca2+ waves. Improved functional and structural computational models are needed to guide the experiments and to test further our understanding of how the t-tubule microanatomy and protein distributions regulate the normal cell function or cell cycle under certain pathologies. To our best knowledge, this study is the first attempt to use the finite element methods to investigate the intracellular Ca2+ responses in physiologically realistic transverse-axial t-tubule geometry.
|
10.1371/journal.pcbi.1001109 | Simulated Epidemics in an Empirical Spatiotemporal Network of 50,185 Sexual Contacts | Sexual contact patterns, both in their temporal and network structure, can influence the spread of sexually transmitted infections (STI). Most previous literature has focused on effects of network topology; few studies have addressed the role of temporal structure. We simulate disease spread using SI and SIR models on an empirical temporal network of sexual contacts in high-end prostitution. We compare these results with several other approaches, including randomization of the data, classic mean-field approaches, and static network simulations. We observe that epidemic dynamics in this contact structure have well-defined, rather high epidemic thresholds. Temporal effects create a broad distribution of outbreak sizes, even if the per-contact transmission probability is taken to its hypothetical maximum of 100%. In general, we conclude that the temporal correlations of our network accelerate outbreaks, especially in the early phase of the epidemics, while the network topology (apart from the contact-rate distribution) slows them down. We find that the temporal correlations of sexual contacts can significantly change simulated outbreaks in a large empirical sexual network. Thus, temporal structures are needed alongside network topology to fully understand the spread of STIs. On a side note, our simulations further suggest that the specific type of commercial sex we investigate is not a reservoir of major importance for HIV.
| Human sexual contacts form a spatiotemporal network—the underlying structure over which sexually transmitted infections (STI) spread. By understanding the structure of this system we can better understand the dynamics of STIs. So far, there has been much focus on the static network structure of sexual contacts. In this paper, we extend this approach and also address temporal effects in a special type of sexual network—that of Internet-mediated prostitution. We analyze reported sexual contacts, probably the largest record of such, from a Brazilian Internet community where sex buyers rate their encounters with escorts. First, we thoroughly investigated disease spread in this dynamic sexual network. We found that the temporal correlations in this system would accelerate disease spread, especially at shorter time scales, whereas geographical effects would slow down an outbreak. More specifically, we found that this contact structure could sustain more contagious diseases, like human papillomavirus, but not HIV. These results highlight the importance of prostitution in the global dynamics of STIs.
| Spatiotemporal heterogeneities in sexual contact patterns are thought to influence the spread of sexually transmitted infections (STIs). Since epidemics can be a society-wide phenomenon, and sexual contact patterns can have structure at all scales, we need population-level sexual network data to understand STI epidemics. Unfortunately, it is hard to collect sexual contact data on that large a scale. Instead, people have focused on small-scale studies using interviews [1]–[3] or contact tracing [4]–[8], or they have studied larger sample sets using random sampling surveys [9]–[13]. Small surveys and contact tracing risk missing large-scale structures [13] and emergent phenomena. Large-scale surveys, on the other hand, have mainly collected the number of partners, but not the connections between them. An alternative way of gather information about sexual contact patterns, which covers a large number of people and explicitly maps their connections, is to use Internet data. In our study, we used a dataset of claimed sexual contacts between Brazilian escorts (high-end prostitutes) and sex buyers [14]. Contact patterns of commercial sex cannot be generalized to a whole population, but they do contain relevant information that can be used to study possible transmission pathways within a social group. Our dataset has information about the time and location of sexual contacts covering six years and 16,748 individuals.
Sexual contact patterns have temporal correlations both at an individual and at a population level [14]. Much like the network structure, temporal structures may influence epidemics in several ways. For example, consider three individuals, A, B, and C, where B and C are in contact first, and later A and B. Considering the temporal order of the contacts, disease cannot spread from A to C via B, but in a standard static network representation this sequence of events is lost, so C appears reachable from A via B [15]–[19]. A conspicuous temporal structure in human behavior that we also observe in our data is bursts of activity during which people are very active for a limited period at a time [20]. Another example of a temporal structure is the long-term behavioral change in which new individuals enter the system and others leave. These temporal effects result in a heterogeneous distribution of inter-event times [14]. To investigate such temporal effects empirically requires time stamps on the contacts. Internet data sets like ours, as opposed to most above-mentioned data, contain just such information.
Extending epidemiological models to include space is a common step towards inclusion of structure beyond the well-mixed assumption [21], [22]. Geography leaves several imprints on the contact structure and thus on disease spread, making the contact network larger than a random graph in terms of graph distances; it also creates the network clusters corresponding to densely populated areas [14]. These effects stress the importance of network-data sets covering a wide geographic area. Our data set, although just a small fraction of the global sexual networks, probably represents a substantial fraction of the Internet-mediated escort business of Brazil [14].
In this paper, we address the question of how the dynamic contact structure in the contact data of Rocha et al. [14] affects epidemic spread in general. For most of our study, we look at spread processes confined to our contact data. Because of the lack of similar data on other types of sexual interaction, it is hard to draw conclusions about the role of the escort business on the spread of STIs in society as a whole. Rather, we investigate the contribution of topological, temporal, and geographic structure on transmission pathways within this specific type of commercial sex. We do, as an example of how our study can be applied to more specific cases, make a crude estimate of the role of our commercial-sex network in the spread of HIV in a population-wide context.
The web community from which our dataset is obtained is a public online forum openly visible online. The full dataset is available as support information (Dataset S1). It is oriented to heterosexual males (sex buyers), who evaluate and comment on their sexual encounters with female prostitutes (sex sellers), both using anonymous aliases. The posts on the forum are organized by the city location of the encounter and by type of prostitution as defined by price level and mode of acquiring customers (for example, escorts, street sex-workers, brothels). We focus on the escorts section, the most expensive form of prostitution [23] of the forum, mostly because it is better organized than the other sections—each escort is discussed in a unique thread. This forum can straightforwardly be represented as a bipartite network—we connect a sex buyer (one type of node) posting in a forum thread to the escort (another type of node) discussed in the thread. An edge in this network represents one sexual encounter between two individuals. The edges are tagged with the dates of the posts, which we take as an estimate of the time of the sexual encounters, even though the sex buyers often post about several encounters at the same session. Consequently, the order of the posts does not have to be exactly the same as the order of the actual encounters. The dataset covers the beginning of the community, spanning the period September 2002 through October 2008. All in all, 50,185 contacts are recorded between 6,642 escorts and 10,106 sex buyers. Even though the network is spread out over twelve Brazilian cities, these contacts make up a network with a largest connected cluster covering more than 97% of the individuals (see Rocha et al. [14] for a thorough analysis of this sexual network). To minimize finite-size effects, we discard the initial 1000 days available in the original data set that correspond to a transient period with fewer users and sparse encounters. One thousand days is an adequate choice, since after that period, the average temporal profile is approximately stationary (see Ref. [14] for details). For better statistical significance, we sample several windows of 800 days. For example, one network sample (of the original network) is obtained by taking all nodes and links that occurred in the period between 1000 and 1800 days; another sample is from the period between 1001 and 1801 days, and so on up to the interval 1200 and 2000 days. The average number of vertices of all windows is N = 10,526±145, and the number of contacts (links) is C = 27,973±3,612, where ± corresponds to the sample standard deviation.
Apart from the anonymous aliases of sellers and buyers and time stamps, posts also include the buyers' grades of the escorts' performance and information about the types of sexual activity performed during an encounter, divided into three categories: oral sex (with or without condom), mouth kissing, and anal sex. All posts, however, are assumed to report vaginal intercourse (random inspection supports this assumption). In our simulations, for the sake of simplicity, unless otherwise stated, we use all available links and disregard the fact that they possess different levels of risk. Most contacts between a seller and buyer happen only once. By inspection, several users report that next time they buy sex, they prefer a different escort, even if the encounter was graded good. We can expect that not all Brazilian escorts and customers of such are present in the data. Furthermore, posting about an encounter is a low-cost action by the sex-buyer that gives him status in the community, which makes it likely that the reports from most users are quite complete. For most of this paper, we ignore this and study disease spread on a network defined by our data set as it is, which limits our conclusions to effects of temporal structure relative to various other scenarios.
A common method of studying correlations in empirical contact data is to compare a network with ensembles, where some properties (like the number of nodes and their degrees) are kept constant and the rest is randomized. In the randomized network versions used in this paper, we conserve the bi-partite structure of the heterosexual network and the number of contacts of each individual.
Diverse network structures can affect disease spread [24]–[26]—one example being clustering (a high density of triangles). Our network has a large number of 4-cycles (the shortest cycle in a bipartite graph), and a pronounced community structure, probably a result of the system being geographically embedded [13], This effect can be studied by randomizing the contact pairs in such a way that we choose two links randomly and swap the respective sex-buyers (we call this new null model random topological, RT). We do not alter the time stamps of the links; hence, the time order of the escorts' contacts is preserved. To remove temporal correlations, we choose two links randomly and only swap their time stamps such that the new encounter time is unrelated to the original, but the network structure is conserved (this model is named random dynamic, RD). Finally, we make a third randomization, where both the temporal and network structures are removed by swapping the time stamps and contact pairs simultaneously (we call this model random dynamic topological, RDT).
To put our results in the context of other levels of epidemiological modeling, we also consider two other contact models—a static network approach and the dynamic network model by Volz and Meyers [18]. The static network (SN) approach considers the network of pairs, with at least one contact over a time interval of 800 days, and assumes that contacts can happen with equal probability over all these links. This approach is common to most network epidemiological studies (e.g., refs. [2], [3], [10]). To compensate for the removal of the time stamps, we assume that each link has a certain probability of being active. This probability is derived from our original network and depends on the number of contacts C = 27,973 and number of different partners K = 21,813 in the window of T = 800 days. Thus, the chance of having a contact active is, for our data, pactive = C/KT = 1.603×10–3. The idea of Volz and Meyers's model is that vertices change partners with a probability pchange while keeping the number of partners fixed over time. This model assumes that a vertex is always connected to someone else; however, in our network, in the interval of 800 days, several vertices have only one or few days during which a connection is active. This means that most of the time they are not in a position to catch a disease. To compensate for this effect, and to allow direct comparison to the simulations on the empirical network, we modify the Volz–Meyers (VM) model to capture the brevity of partnerships in the data. In our formulation of the VM network, each vertex has a chance pk (proportional to the original number of contacts of the vertex) of being active per day. This assures that over the course of 800 days, each vertex has the same number of contacts as in the original empirical network. For each day, we connect pairs of active vertices randomly (if the number is odd, the remaining active vertex is connected the next day that another active vertex is available). Thus, this network has no temporal correlations. We generate VM graphs with 10,526 vertices—the same as the average number of vertices in the sampled windows discussed above. We obtain the degree distribution from these sampling windows as well and use it to calculate pk.
One can model the spread of sexual infection in various ways to capture the various characteristics of pathogens and contact patterns, and also to serve different aims of explanation and prediction. We explore the effects of temporal correlations on different levels of epidemic modeling. The first disease-transmission model we consider is the Susceptible–Infected–Removed (SIR) model. Where all individuals are initially susceptible; upon contact with an infective, a susceptible becomes infective with probability ρ (probabilities are, unless otherwise stated, per-contact probabilities), and after a fixed time δ, a susceptible changes to the removed state. If δ is larger than the vertex lifetime in the network, we get the limit case known as the Susceptible–Infected (SI) model. In a static network of finite size and non-zero transmission rate, all vertices will eventually become infected in the SI model. This is not necessarily the case in a temporal network, which makes the SI model more realistic in temporal, compared to static, contact networks.
To simulate these models in our empirical network, we first map the sampled network onto a time-ordered list. Each entry in the list is one pair of vertices and the time of the contact. Different contacts between the same pair appear as different entries in the list. Then we divide the list into intervals of 800 days each, as mentioned above. The pairs are ordered according to their times of contact. We select the sex-seller of the first contact of an interval as a source of infection and go through the ordered list infecting a susceptible vertex in contact with an infective vertex with probability ρ. The state of the vertex is updated at each new contact. A way of modeling the fact that the network is connected to a background of sexual contacts would be to include multiple sources of infection. To keep the simulations simple, however, we leave this for future studies. Since our temporal information has a resolution of one day, we do not know the order of contacts within a day. To remove this potential bias, we randomize the order of contacts within a single day 100 times. In line with other studies, and to simplify the model, we assume that both infection and removal (after time δ) are immediate, and the transmission probability is constant.
The SI model is adequate for modeling the early phase of an outbreak over shorter time scales than the duration of the disease. SIR, on the other hand, is appropriate for simulating diseases having a well-defined infectious stage followed by immunity. As an example, we will investigate HIV at a more detailed level than simply SI or SIR. Hollingsworth, Anderson, and Fraser [27] devised a model for HIV-1 infection with a susceptible stage followed by four distinct infective stages of different infectivity—one acute infection of high infectivity (over a time-scale of months) followed by a chronic stage (lasting for years), and another high infectivity stage (some weeks) followed by zero infectivity before death. Since our dataset covers only 1000 days, we can omit the last two stages and arrive at a model characterized by an acute stage of transmission probability ρ1 lasting for a time T1, and a chronic stage of transmission probability ρ2. We refer to this as SI1I2 model. Strictly speaking, the transmissibility of HIV-1 also depends on gender and other factors such as type of sex and the fact that the viral load transmitted per-contact can spike during the chronic phase because of comorbidities, among other things. A yet more detailed model could also include an age-stratified population, as young infectives tend to influence an outbreak more. Because they are in the network for longer times, they have higher chance to establish more contacts and contribute to transmit the infection [28].
We follow a similar procedure as above to simulate disease spread in the SN and VM networks. For the initial conditions, however, since the probability of being infected should increase with contact rate in case of the empirical networks, we now select the source of infection randomly (for each realization) and proportionally to the number of contacts of the vertex. This procedure compensates for the fact that in the empirical network, high degree nodes are necessarily selected more than once as a source of infection. This is because, on average, the chance of an individual's being active at a certain moment is proportional to that individual's number of contacts. The state of the vertex is updated after all vertices have been considered. We run the algorithm 30,000 times to obtain averages for these models.
A key quantity is the fraction of infected vertices Ω (the outbreak size). If the time evolution is not explicitly stated we refer to Ω at the end of the sampling time window (800 days). We also run simulations 50 times over different initial conditions to calculate the average values.
A straightforward way of investigating the effects of the temporal and topological structure of contact patterns is to remove different types of correlations by randomization (see Section The network models). In Figure 1, we investigate effects of the time ordering of contacts by using the SI model with ρ = 1 and compare the simulated epidemics in the original network with the epidemics in the three different randomized versions of it. In Figure 1A, we see that an infection spreads much more slowly in the RD network model, reaching fewer than 50% of the individuals compared to more than 60% in the original network. Thus, correlations in the order in which the contacts occur speed up disease spread. More concretely, one such tendency is that individuals tend to be intensely active over a period of time followed by idle periods. When the time stamps are randomized (RD model), this tendency disappears such that the presence of individuals in the system is now, on average, longer and the contacts less frequent. The average time, between an individual's first and last active period of, increases from 170.9±0.1 days in the original network to 337.5±0.1 days after randomization. In addition to correlations in the temporal order of contacts, the topology of the sexual network can also influence epidemics [3], [6]–[9], [18]. In Figure 1B, we compare the evolution of epidemics in the empirical network with the RT network model. The evolution of the fraction of infected individuals 〈Ω(t)〉 seems to grow slowly, at least during the initial 200 days; afterwards, the topologically randomized network yields more rapid and pervasive outbreaks (Figure 1B). The more rapid initial epidemic spread in the original network results from the high clustering of contacts within cities. Finally, considering both the temporal and topological information randomized (RDT model), the curve (evolution of the epidemics, Figure 1C) is in between those of Figure 1A and Figure 1B. The fraction of infected vertices increases slowly during the initial 300 days, but not more slowly than in the RD scenario in Figure 1A. Later it increases more rapidly and by the end of the sampling period reaches about 70% of the individuals (a little less than in the RT scenario in Figure 1B, but still, larger than in the original network).
The limit of high transmission probability ρ = 1 does not reflect actual STI contagion; more realistic values lie in the range 0.001≤ρ≤0.3 [27]–[28]. In Figure 2, we present 〈Ω〉rel = 〈Ωρ〉/〈Ωρ = 1〉, the average number of infected vertices (for probabilities ρ) relative to the number of infected vertices when the maximum transmission probability is used (ρ = 1). The relative number of infected vertices decreases within the initial 100 days and afterwards reaches a minimum for higher transmission probabilities while continuing to decrease slowly for lower rates. The minimum, which corresponds to the time lag of secondary infections, is more pronounced for lower ρ-values. The fact that the curves are fairly constant for times longer than 200 days, that is, that they converge to limiting values, is an indication that our results for the ρ = 1 case hold for other transmission probabilities as well, that is, the time-ordering effects are stronger than the fluctuations from the stochasticity of the contagion process. For lower ρ-values, the curves decrease monotonically, which indicates the existence of an epidemic threshold somewhere between, ρ = 0.01 and ρ = 0.001, which we investigate more cautiously below.
We note that there is a large diversity of outbreaks even for ρ = 1. In Figure 3, we measure the probability distribution P(Ω) of outbreak sizes Ω. This, we hypothesize, is a general phenomenon—temporal constraints increase the diversity of outbreaks because they restrict the possible infection paths in the network. There is, however, a local maximum where, for ρ = 1, a fraction of about 0.75 of the vertices gets infected, setting a characteristic outbreak size. This local maximum depends on the transmission probability and decreases for lower values. Another observation is that the outbreak-size distribution becomes less heterogeneous for lower ρ-values. The peak on the very left of the graph indicates that the disease is likely to die out. Note that the network also contains some isolated connected components that, once infected, do not spread the infection to the giant component.
To illustrate the effect of different sexual activities, we show the outbreak size distribution for the original network considering only the encounters that involve oral sex without condom, and mouth kissing (Figure 3B). This specific network has roughly the same outbreak-size distribution (similar shape and scale) as the original network, despite being about half as dense.
Returning to our original network, we investigate the effect of varying ρ, and see that the average outbreak size 〈Ω〉 is an approximately linear function of transmission probability (see Figure 4A). From the figure, it is evident that the epidemic outbreak is practically absent for transmission probabilities lower than about 0.19—a de facto threshold effect. Looking in more detail, one can see that this threshold effect is due to the fact that the mean value of large outbreaks vanishes and that the number of small outbreaks increases as ρ→0 (cf. Figure 3). To investigate whether this threshold value ρ* is an artifact of the finite-size sampling, we use different sampling windows from the complete dataset and see whether the threshold values converge to a common value for different starting points in the dataset. The threshold values should converge to a limit value as the network structure trends toward a steady state. Each window represents the same duration of time (800 days), but simply starts at a different time (T0) in the original data set. We take the crossing point between the fit of the fraction of infected vertices as a function of ρ to a line and the line of zero secondary infections as an estimate of the threshold value. We see in Figure 4B the apparent convergence of the threshold estimates to values at about ρ* = 0.19±0.01 for increasing T0, which is our estimated threshold value for this contact pattern. This threshold seems slightly smaller for the RDT, but significantly smaller for the SN and VM models (Figure 5A–C).
We plot the average outbreak size 〈Ω〉 as a function of the duration of the infective stage δ in Figure 6A. Here, we assume the maximum transmission probability ρ = 1. We proceed to identify estimated threshold values by performing fits of second-order polynomials to the fraction of infected individuals and identify the crossing point with the zero secondary infection line. Performing a similar analysis as for the SI model's transmission probability threshold, but now for the duration of the infective state, we find that the δ-threshold converges to δ* = 31±1 days.
Now, fixing the infective stage to δ = 91 days, which is roughly 3 months and well above our estimated threshold of δ*, we perform SIR simulations for different transmission probabilities and compare the outbreak sizes by using the original network, the randomized version (RDT), a static (SN), and a dynamic network (VM) (Figure 5D–F). For all cases, the thresholds are above ρ* = 0.2, and the final outbreak size is always larger for the empirical network, suggesting that the temporal correlations, the essential difference between the raw empirical contact patterns, and the models accelerate transmission.
Now we turn to the results of the SI1I2 simulation of HIV spread. We fix the acute infective period at T1 = 91 days and study some different combinations of estimated transmission probabilities available in the literature for different societies by using lower (ρ1 = 0.005 and ρ2 = 0.0005) and higher (ρ1 = 0.01 and ρ2 = 0.001) bounds [27], [29]. In Figure 4A, we see that the threshold transmission probability of the SI model is higher than all these values, so we already know that the SI1I2 model on the actual data is below the epidemic threshold. In Figure 7, we plot the average time evolution of the outbreak size for both our empirical temporal network (Figure 7A) and the RDT contact model (Figure 7B). The average outbreak sizes are, as expected, very low (a fraction of about 10–5 of the population) for both these contact patterns. For the RDT model and ρ1 = ρ2 = 0.01, the system is just above the epidemic threshold, as can be seen by its convex curve in Figure 7B. A conspicuous temporal feature is that, for the empirical network, the effect of a larger transmission probability of the chronic infection is very small—the ρ1 = ρ2 = 0.01 and ρ1 = 0.01, ρ2 = 0.001 curves are almost congruent. For the RDT contact structure, the more homogeneous temporal pattern allows the chronic infection to play a greater role, so these two curves diverge after about 200 days, which is about the average interval between two consecutive contacts.
We simulate the spread of infection in what is probably the largest network of self-reported sexual contacts yet recorded. Our data come from a web community of sex buyers who discuss their encounters with escorts. Although the network is spread out over twelve cities, it is to a large extent connected so that a disease could spread from most parts of the system to most other parts. As with any result based on a subset of a network, we should be cautious about extrapolating our results to the entire society, especially since it is hard to compensate for missing links with the information we have. The escorts in our dataset make up about one percent of all Brazilian sex sellers (of a total of about one million [30]). On the other hand, the escorts are a small fraction of all sex-sellers, and we can tell by the way the average degree (number of partners) converges that the sampling time is longer than an escort's typical duration in the business [14]. Another complication when it comes to generalizing the results of the paper to a society as a whole is that our sexual network is not an isolated system. It is possible that the infection leaves the community and eventually returns through other individuals. In that case, our model would underestimate the impact of the network in the outbreak. Furthermore, commercial sex is not necessarily driven by the same mechanisms as regular sexual interaction. So, since our data is not comprehensive enough to infer the impact of prostitution on disease spread, we focus on how temporal correlations in the empirical data affect results from random and well-mixed models. From studying the SI model with a 100% transmission probability in our sexual network, we conclude that temporal correlations speed up the epidemics, especially in the early phase of superlinear growth. This effect has important implications both for disease modeling, implying that temporal correlations in contact patterns should not be underestimated, and for intervention methods (like targeted vaccination), where temporal structures could potentially be used to detect important individuals. Furthermore, the temporal effects seem to cause well-defined and relatively high epidemic thresholds, unlike studies of model networks with power-law degree distributions [10] where outbreaks can occur of any non-zero transmission probability. For purposes of comparison, in a finite-sized scale-free network and Susceptible-Infected-Susceptible epidemics with recovery μ = 1, ρcritical = 〈k〉/〈k2〉 (where k is the number of different partners of an individual [31]), gives ρcritical∼0.043, using our network. The network structure (apart from the contact-rate distribution), on the other hand, slows down outbreaks. Our network has a high density of short cycles and community structure, reflecting the fact that most sex buyers buy sex in one region, presumably their hometown. Both factors, many short clusters and distinct communities are known to slow down diffusion in networks [24]–[26].
Most of our analysis is at a general STI level and indicates that our network is not dense enough to support STI outbreaks for chronic diseases with transmission probabilities lower than ρ = 0.19. The fact that endemic diseases with arguably lower transmission probabilities exist points to the importance of the background sexual contacts. Because of the incompleteness of our data, this does not completely exclude the possibility of escort prostitution as a reservoir of STIs, but it points to a more complex picture. In the support information (Text S1), in a crude assessment of our dataset contribution to the general STI spread, we suggest that it would only affect the degree-correction of R0 of STIs by a few percent. We also exemplify how temporal structures can affect the spread of a specific pathogen, HIV-1, by simulations of a refined compartmental model. The simulation results indicate that our empirical network alone cannot sustain an outbreak of HIV-1. In general agreement with empirical research [32], [33], our results suggest that pathways (like unsafe man-to-man sex, or intravenous drug use) other than commercial sex are needed to explain the endemic state of HIV epidemics in Brazil [34]. The other studies are, however, from countries other than Brazil; however, they are inconclusive if not controversial [35], [36].
We believe that the study of temporal aspects of contact patterns is, in general, a promising direction for the future. We intend to investigate how far our conclusions can be generalized to other types of cultures, other forms of commercial sex, and hopefully to non-commercial sexual contact patterns.
|
10.1371/journal.pcbi.1004269 | Regulators Associated with Clinical Outcomes Revealed by DNA Methylation Data in Breast Cancer | The regulatory architecture of breast cancer is extraordinarily complex and gene misregulation can occur at many levels, with transcriptional malfunction being a major cause. This dysfunctional process typically involves additional regulatory modulators including DNA methylation. Thus, the interplay between transcription factor (TF) binding and DNA methylation are two components of a cancer regulatory interactome presumed to display correlated signals. As proof of concept, we performed a systematic motif-based in silico analysis to infer all potential TFs that are involved in breast cancer prognosis through an association with DNA methylation changes. Using breast cancer DNA methylation and clinical data derived from The Cancer Genome Atlas (TCGA), we carried out a systematic inference of TFs whose misregulation underlie different clinical subtypes of breast cancer. Our analysis identified TFs known to be associated with clinical outcomes of p53 and ER (estrogen receptor) subtypes of breast cancer, while also predicting new TFs that may also be involved. Furthermore, our results suggest that misregulation in breast cancer can be caused by the binding of alternative factors to the binding sites of TFs whose activity has been ablated. Overall, this study provides a comprehensive analysis that links DNA methylation to TF binding to patient prognosis.
| DNA methylation is a ubiquitous and simple covalent modification that occurs directly on genetic material whereby a simple methyl group (CH3) is attached to Cytosine nucleotides in the context of CpG sites. Modifications of these sites have been postulated to function in gene regulation, potentially via interactions with transcription factors. In this study, we hypothesized that DNA methylation signals contain valuable information that can help infer transcription factors that may be associated with a given disease. Here, we utilize the vast repository of breast cancer data that is available in the public domain, and which contains a rich resource for DNA methylation and clinical data on breast cancer patients. In this guilt-by-association analysis, we postulated that conserved transcription factor binding motifs that are statistically enriched in regions near methylated CpG sites that are correlated with breast cancer patient survival would suggest that their cognate transcription factors would play a role in the initiation, growth, metastasis, or even suppression of the tumor. This integrative approach supports the claim that DNA methylation profiling of patient tumors in the clinic may contain valuable information that can guide the development of treatment regimens for individual patients; thus contributing to the progression of precision medicine.
| DNA methylation is a critical regulatory process that involves direct chemical modification of genetic material via the addition of a methyl moiety to the 5th carbon of Cytosine nucleotides. These covalent modifications occur most prevalently on CpG dinucleotides (CpGs) and are reversible, thus allowing the DNA methylome to achieve a balance of stability and plasticity. DNA methylation plays essential roles in X-chromosome inactivation [1], genomic imprinting [2], transposable elements silencing [3], stem cell differentiation [1,4–6], embryonic development [7,8], and inflammation [9,10]. Considering these critical roles, aberrant DNA methylation patterning has been observed in nearly all cancer types and in a plethora of non-cancer diseases including autoimmune disorders [11,12], neurological diseases [11,13] metabolic disorders [14], and cardiovascular disease [15]. Furthermore, DNA methylation signatures and markers have been used to stratify cancer subtypes and predict patient prognosis [16–18].
Recently, the use of DNA methylation profiling to predict prognostic outcomes of diseased patients has gained popularity. In breast cancer, studies have shown that ER+ and ER- breast cancer cell lines could be distinguished by examining their DNA methylation patterns. Sun et al. identified 84 genes that were differentially methylated between ER+ and ER- cell lines [19]. Additionally, the TCGA consortium clustered 802 primary breast cancer samples based on their DNA methylation signals; this yielded 5 distinct clusters that comprised samples that exhibited varying molecular phenotypes [20]. In a recent study, Anjum et al. identified a BRCA1 mutation-associated DNA methylation signature in 144 case-control primary blood samples that was predictive of breast cancer incidence and patient prognosis [21]. Furthermore, Bullinger et al. applied a MALDI-TOF-MS based methylation analysis to identify a DNA methylation signature in 182 acute myeloid leukemia primary samples that was predictive of patient outcomes [22]. Several other studies have identified DNA methylation signatures and markers in primary breast tumor samples that were shown to predict patient outcome [23–26]. These studies have shown that understanding DNA methylation patterning and dissecting its functions provide valuable insight into its regulatory roles, which may ultimately introduce new avenues for developing efficacious breast cancer treatments.
Despite recent focus on epigenetic based markers, the exact mechanism(s) by which DNA methylation regulates gene expression has yet to be elucidated but its interaction with transcription factors (TFs) have been shown to be a critical mechanism [27–31]. It has been suggested that 5-methyl-CpGs (5meCpGs) physically impede the binding of TFs to their cognate sequences causing gene silencing [27]. Additionally, 5-meCpGs can indirectly control gene expression by modulating local chromatin structure via recruitment of histone remodeling factors such as histone deacetylases and histone methyltransferases [32–34]. Physical obstruction of TF binding and compaction of chromatin structure suggest that DNA methylation exerts a silencing effect; however, studies have shown TFs such as SP1 can bind 5-meCpGs and induce gene expression [35]. In a comprehensive and systematic genomics study, Hu et al. applied a protein microarray-based approach to identify which of 1321 TFs and 210 co-factors have the capacity to bind motifs containing methylated CpGs [36]. They reported 41 TFs and 6 TF co-factors that bound 5meCpG in a sequence-specific manner [36]. Indeed, it is possible that DNA methylation patterns are just passive markers of TF binding or gene regulation whereby CpGs in unbound chromatin are methylated and have no functional relevance [37,38]. Regardless of the exact mechanism, we postulate that changes in CpG signals near putative transcription factor binding sites (TFBSs) can reflect the activity of TFs and can be used to infer the underlying transcriptional machineries that drive the progression of several subgroups of breast cancer.
We have previously integrated ENCODE [39] and TCGA [20] data to computationally examine the association between TF binding and DNA methylation levels in TFBSs (i.e. ERα) [31]. We found that there is a strong negative correlation between ERα activity and DNA methylation levels within ERα binding sites in breast cancer [31]. More importantly, differentially methylated CpGs between ER+ and ER- breast cancer are enriched in the DNA regions surrounding ERα binding peaks (determined by ChIP-seq) in a distance dependent manner—the closer to the center of binding peaks, the more differentially methylated the CpGs tend to be [31]. Conversely, given a set of differentially methylated CpG sites between ER+ and ER- samples, we would expect the binding site motif of ERα or other functionally related TFs to be enriched within the vicinity of these CpG sites. These findings suggest that DNA methylation patterns and their signals are informative for exploring transcriptional regulation mediated by TFs [14].
In this study, we aimed to utilize DNA methylation data derived from primary breast cancer samples to identify TFs that are associated with patient survival via their relationship with methylated CpGs. To achieve this we connected the DNA methylation-TF interactome to breast cancer patient survival using datasets derived from ENCODE and TCGA [20]. Specifically, we identified a list of CpG sites that were significantly correlated with patient survival time in their methylation level. We then determined which TF binding sites are enriched in DNA regions surrounding survival-associated CpGs to extrapolate TFs that are associated with patient survival. Interestingly, we ascertained that ERα TF binding motifs were significantly enriched in survival-associated CpG regions in ER- samples only, and p53 TF binding motifs were enriched in survival-associated CpGs regions in p53- samples only. Overall, our analysis framework demonstrates the intimate linkage between DNA methylation, TF binding, and breast cancer patient prognosis.
The ultimate goal of our analysis was to identify TFs that impact breast cancer patient prognosis via an association with CpG methylation. By identifying TF binding motifs enriched in regions containing differentially methylated CpGs or survival-associated CpGs, we were able to demonstrate a relationship between TF-DNA methylation mediated regulation and overall patient survival rates. Fig 1 depicts our integrated approach to dissecting epigenetic involvement in transcriptional regulation underlying breast cancer patient survival. First, we investigated TF binding motifs enriched in differentially methylated CpG regions to demonstrate that methylation data are informative for inferring transcriptional regulation in breast cancer (Fig 1: Top). We successfully detected the enrichment of ERα TF binding motifs in DNA regions surrounding CpG sites that were differentially methylated between ER+ and ER- breast cancer patients.
Most importantly, we applied motif enrichment analysis to survival-associated CpGs (Fig 1: Bottom) using all breast cancer samples and subsets of samples stratified based on histological, intrinsic, and CpG beta-value intensity phenotypes. We first identified a set of survival-associated CpGs by correlating the methylation levels of each CpG across TCGA breast cancer patient samples using a univariate Cox proportional hazards model [40]. Second, we defined a 102 bp genomic region centered at each CpG (henceforth referred to as a CpG region) and computationally searched for the presence of TF binding motifs in this region. Third, we systematically calculated TF binding motif enrichment in these CpG regions in different subtypes of breast cancer.
To preliminarily demonstrate that TF-mediated transcriptional regulation could be inferred from DNA methylation signals, we investigated the relationship between CpGs with altered methylation levels and the presence of putative TF binding motifs vicinal to these CpGs. Specifically, we identified CpGs that were differentially methylated between estrogen receptor positive (ER+) and estrogen receptor negative (ER-) breast cancer samples in TCGA data using a Student’s t-test, and then examined the occurrences of putative TF binding motifs within DNA regions surrounding these CpGs.
We systematically calculated the enrichment levels of 703 TF binding motifs available from the TRANSFAC and JASPAR databases in differentially methylated CpG regions. Our analysis identified 60 TF binding motifs (38 TFs) enriched in and 105 TF binding motifs (67 TFs) depleted in hypomethylated (ER+ < ER-) CpG regions at a P<1E-15 significance threshold (Fig 2A) (adjusted p-value using the Benjamini-Hochberg multiple testing correction method; hereafter, all reported p-values have been adjusted unless otherwise indicated). In addition, we identified 12 TF binding motifs (10 TFs) enriched in and 50 TF binding motifs (35 TFs) depleted in hypermethylated (ER+ > ER-) CpG regions at the same threshold (Fig 2A). Similar results were obtained when a Wilcoxon ranked sum test was used to identify differentially expressed CpGs (S1 Table).
To validate the accuracy of our systematic analysis, we directed our attention to ERα TF binding motifs. Since stratifying breast cancer patient samples into ER+ and ER- groups is analogous to controlling for ERα activity, we hypothesized that ERα TF binding motifs would be significantly enriched in hypomethylated (ER+ < ER-) CpG regions or, alternatively, depleted in hypermethylated (ER+ > ER-) CpG regions. Our data set includes three ERα TF binding motifs: JA-ESR1, TR-ER_Q6, and TR-ER_Q6_02 for which we calculated their enrichment in the two CpG sets (hypo- and hypermethylated sets). First, JA-ESR1 was depleted in hypermethylated CpG regions 0.93-fold (P = 7.7E-4) and depleted in hypomethylated CpG regions 0.94-fold (P = 0.005) (Fig 2). Second, we observed that the TR-ER_Q6 motif was also depleted in hypermethylated CpG regions 0.84-fold (P = 3.1E-7), but unlike JA-ESR1, was enriched 1.11-fold in hypomethylated CpG regions (P = 0.003) (Fig 2B). Lastly, the TR-ER_Q6_02 motif was depleted in hypermethylated CpG regions 0.92-fold (P = 0.006) and enriched in hypomethylated CpG regions 1.27-fold (P = 3.8E-16). These results indicate that TF binding activity can be inferred based on enrichment of their motifs in DNA regions near informative CpGs (e.g. differentially methylated CpGs). Previously, we had established that DNA methylation within ERα binding sites was anti-correlated with ESR1 expression by integrating TCGA (gene expression and DNA methylation) and ENCODE ChIP-seq data [14]. Therefore, the depletion of all 3 ERα TF binding motifs in hypermethylated CpG regions is in accordance with our previous analysis confirming that ERα activity is associated with loss of binding site-specific DNA methylation [41].
In addition to ERα TF binding motifs, we also identified a number of other TF binding motifs that are known to be associated with ERα. Strikingly, hypomethylated CpG regions contained 18 TF binding motifs (corresponding to 5 FOX family transcription factors, GATA1, HFH8, XFD2, and XFD3) that exhibited greater than 2-fold enrichment, whereas hypermethylated CpG regions contained none (Fig 2A). Operating under the passive model [42], this suggests that loss of methylation is generally associated with enhanced binding activity of these transcription factors in ER+ breast cancer. In addition, we identified all GATA3 and FOXO1 TF binding motifs to be enriched in hypomethylated CpG regions and depleted in hypermethylated regions suggesting that these TFs are associated with ERα activity. Indeed, it has been experimentally shown that FOXA1 influences ERα function by modulating ER-chromatin interactions and FOXA1 deficiency results in loss of ERα activity (S1 Table) [43–45]. In addition, GATA3 has been shown to be necessary for estradiol stimulation of breast cancer cells and more recently, modulate ERα access to enhancer regions [46,47]. Overall, our motif enrichment results in differentially methylated CpG regions results are consistent with the known biological roles of our identified TFs.
We were additionally able to verify several enriched TF binding motifs via a de novo motif search in hyper- and hypomethylated CpG regions. Specifically, we applied the Discriminative Regular Expression Motif Elicitation (DREME) algorithm to identify enriched DNA motifs in hyper- and hypomethylated CpG regions and then matched them to known motifs (See Materials and Methods). We were able to detect the presence of the ESR1 motif in hypomethylated CpG regions but not in the hypermethylated CpG regions, which is consistent with the enrichment results for ERα. We were also able to confirm motifs including FOXO1, SP1, KLF4, EGR2, and E2F1 in hypomethylated CpG regions and TCF3, NHLH1, and HEB in hypermethylated CpG regions.
After identifying enriched TF binding motifs in differentially methylated CpG regions, our next objective was to determine if TFs associated with patient survival could be inferred based on DNA methylation signals. Aberrant TF activity and DNA methylation changes have both been known to play a role in carcinogenesis and cancer progression. However, to our knowledge, there has been no other study that has systematically investigated survival-associated TF-DNA methylation relationships at the level of specific TF-CpG interaction. Thus, to proceed with this high-resolution analysis, we pinpointed CpGs whose methylation levels were significantly associated with breast cancer patient survival and calculated the enrichment of TF binding motifs in the regions surrounding these CpGs. CpGs with hazard ratios <1 were categorized as protective, CpGs with hazard ratio >1 were categorized as hazardous, and pooled protective and hazardous CpGs were simply categorized as survival-associated. We hypothesized that survival-associated fluctuations in CpG methylation intensities would be informative to the activity of specific survival-associated TFs.
When survival analysis was implemented using all samples, we were able to identify 92 TF binding motifs (62 TFs) enriched and 143 TF binding motifs (98 TFs) depleted in protective CpG regions at significance level P<0.01 (S2 Table). For hazardous CpG regions, we detected 11 TF binding motifs (9 TFs) enriched and 2 TF binding motifs (2 TFs) depleted at the same threshold, respectively. Fig 3 highlights four examples of TF binding motifs enriched in survival-associated CpG regions: p53, ERα, HEB, and LRF. First, JA-ESR1 exhibited an enrichment score of 1.23 at P = 0.004 in hazardous CpG regions (Fig 3A). This indicates that the effect ERα binding activity—it is known that ER status is a significant clinical factor for predicting survival of breast cancer patients—has on patient survival can be inferred from DNA methylation signals correlated with patient prognosis. Second, LRF is an oncogenic transcription factor involved in cell growth and differentiation, and is known to be overexpressed in breast cancer [48]. Our analysis shows that the TR-LRF_Q2 TF binding motif is enriched 1.18 times in hazardous CpG regions (P = 0.005); additionally, it is depleted 0.93-fold in protective CpG regions (P = 0.03) (Fig 3A). In one study, Maeda et al. reported that LRF is necessary for embryonic fibroblast cells (MEFs) to undergo transformation even when other potent oncogenes such as H-Ras, T-antigen, and MYC are expressed [49]. Third, we identified HEB (TCF12) to be enriched 1.36-fold in protective CpG regions (P = 1.5E-29, Fig 3B). HEB has been previously reported to correlate with colorectal cancer metastasis by inhibiting E-cadherin, thus verifying HEB as a potential oncofactor [50]. Fourth, it is well established that loss of p53 activity is detected in approximately 50% of all cancers [51]. In our analysis we were able to identify all p53 TF binding motifs to be enriched in survival-associated CpGs (P<0.05). To highlight, the TR-P53_02 motif was enriched 1.28-fold in protective CpG regions (P = 6.8E-5) (Fig 3A). This again indicates that DNA methylation levels provide information about the activity of key transcriptional regulators.
To provide a brief summary, we show the top ten significant TF binding motifs enriched or depleted in protective CpG regions in Table 1. Similarly, we illustrate the top ten TF binding motifs enriched or depleted in hazardous CpG regions in Table 2.
To demonstrate that survival-associated CpGs are informative for identifying clinically relevant TFs, we focus on two key breast cancer-related proteins: ERα and p53. ERα and p53 are major proteins whose expression levels are typically measured in breast cancer cases to determine the molecular status of the tumor, and it is well-established practice to use this information for determining prognosis and treatment strategies. Here, we explored two major subtyping schemata by systematically calculating the enrichment/depletion of all TF binding motifs (in particular ERα and p53) in survival-associated CpG regions for ER+, ER-, p53+, and p53- breast cancer subtypes (S3–S6 Tables). First, we directed our focus to ERα TF binding in ER-stratified samples and were able to identify JA-ESR1 to be enriched 1.44-fold in protective CpG regions in the ER- subtype but not in the ER+ samples (P = 2.6E-5, Fig 4A). In fact, in ER+ samples, JA-ESR1 is significantly depleted 0.80-fold in protective CpG regions (P = 2.2E-11) (Fig 4A). When protective and hazardous CpGs are combined, JA-ESR1 is enriched 1.20-fold in survival-associated CpGs in ER- (P = 0.001) samples only and depleted 0.83-fold (P = 9.2E-9) in survival-associated CpGs in ER+ (Fig 4A). Furthermore, TR-ER_Q6 was enriched 1.24-fold (P = 0.04) in hazardous CpG regions and TR-ER_Q6_02 was enriched 1.18-fold (P = 0.05) in survival-associated CpG regions (protective and hazardous CpG regions combined) in ER- samples.
Next, we stratified patient samples into p53+ and p53- groups and calculated the enrichment of TF binding motifs with particular focus on p53 TF binding motifs (Fig 4B). We identified JA-TP53 to be enriched 1.27-fold in protective CpGs regions only in p53- samples (P = 0.02) and enriched 1.24-fold in survival-associated CpG regions (P = 0.005) (Fig 4B). Likewise, the TR-P53_01 motif was enriched 1.48-fold in protective CpG regions (P = 3.8E-5) and 1.38-fold in survival-associated CpG regions (P = 5.13E-6) in p53- samples only (Fig 4B). In contrast to the other p53 motifs, TR-P53_02 was enriched 1.29-fold in p53+ samples (P = 0.006) and 1.24-fold in p53- samples (P = 0.05), both in protective CpG regions (Fig 4B). Taken together, our enrichment results from ER- and p53-stratified breast cancer patients sample show that the majority of ERα and p53 TF binding motifs are enriched in CpG regions in ER- and p53- samples, respectively. Presumably, the binding sites of key transcriptional regulators can become unbound and accessible to other factors once the activity of the key TF is ablated. Alternative factors may then bind to these open sites leading to misregulation of the associated genes, and contribute to cancer progression and clinical outcomes of patients.
Breast cancer, like most other cancer types, exhibits a high degree of heterogeneity making it refractory to treatment. One approach to abrogate the effects of sample-to-sample variation is to classify tumors into subtypes, each with distinct genetic, molecular, and physiological features. Therefore, we aimed to resolve whether breast cancer subtypes determined by immunohistochemistry also exhibit differences in TF binding motif enrichment near survival-associated CpGs.
First, we calculated TF binding motifs enriched/depleted in survival-associated CpGs in each histological subtype of breast cancer (S3, S4, S7–S12 Tables). In summary, there are a total of 252 (178 TFs), 135 (85 TFs), and 247 (168 TFs) TF binding motifs that are enriched in protective, hazardous, and survival-associated CpG regions, respectively in at least one subtype at significance level P<0.01. In the opposite direction, 323 (217 TFs), 49 (41), and 305 (208 TFs) motifs were depleted in protective, hazardous, and survival-associated CpG regions, respectively in at least one subtype at the same significance level (See S13 Table for more details). The large number of identified motifs suggests that a variety of TFs contribute to breast cancer development and each TFs activity may or may not be important drivers depending on the subtype.
Second, we clustered the p-values (P<0.05) of significantly enriched or depleted TF binding motifs in survival-associated CpGs and observed that PR+ and ER+, which clustered together, exhibited enrichment patterns much different from that of the other subtypes (Fig 5A). More specifically, these subtypes are enriched in TF binding sites that are depleted in the other subtypes and vice versa. This suggests that the TFs associated with survival in PR+ and ER+ samples may not be significant protein factors in the other subtypes. In addition, it is clear that significantly enriched/depleted TF binding motifs vary from subtype to subtype implying that each subtype exhibits distinct TF-DNA methylation interactions. This shows that unique enrichment signatures can differentiate between breast cancer subtypes by revealing transcriptional regulators most likely to exhibit altered activity.
After showing global TF binding motif enrichment patterns of histological subtypes, we provide an example where the TR-NF1_Q6 motif is enriched 1.76-fold in protective CpG regions (P = 1.1E-9) in the triple-negative subtype (Fig 5B). Mutations in NF-1 have been implicated in the proliferation of triple-negative primary breast cancer tumors since it functions as an inhibitor of RAS and mTOR [52–54]. This suggests that DNA methylation within NF-1 binding sites is associated with longer survival times in patients with triple-negative breast cancer.
To determine if different transcriptional regulators could also be identified in breast cancer subtypes based on molecular features, we classified our samples into 5 distinct intrinsic subtypes: luminal A, luminal B, HER2-enriched, and basal [55]. In some cases, intrinsic subtyping is more representative of the underlying molecular architecture in breast cancer and can be used to predict risk of cancer relapse after treatment with chemotherapy [56].
In our analysis, we first identified CpGs that were correlated with survival for each intrinsic subtype and determined which of 704 TF binding motifs were enriched in hazardous or protective CpG regions (S14–S16 Tables). In summary, a total of 9 (6 TFs), 209 (80 TFs), 113 (62 TFs) motifs were significantly enriched in protective, hazardous, and survival-associated CpGs, respectively (in at least one subtype P<0.01). Furthermore, 21 (16), 31 (27), and 40 (34) motifs were significantly depleted in protective, hazardous, and survival-associated CpGs, respectively (in at least one subtype P<0.01) (See S13 Table for more details).
Second, we clustered the enrichment p-values of significant TF binding motifs (P<0.05) in each intrinsic subtype and noticed that the luminal A subtype contained the largest number of significantly enriched/depleted TF binding motifs that yielded P<0.01 (Fig 5C). Conversely, HER2-enriched samples contained no significant TF binding site enriched or depleted in survival-associated CpG regions. This disparity is most likely due to differences in statistical power resulting from unequal subtype sample sizes and/or longer average patient survival times associated with different subtypes (Fig 5C). Despite this, it is clear that some enriched/depleted TF binding motifs are shared amongst luminal A, luminal B, and basal samples and some are not. Overall, this demonstrates global variation in TF binding site enrichment across intrinsic breast cancer subtypes.
To explore individual TF binding motifs that are enriched in an intrinsic subtype, we illustrate JA-ELK1 as an example. JA-ELK1 is enriched 1.78-fold in hazardous CpG regions in the basal subtype (P = 0.002) (Fig 5D). ELK1 has been shown to be involved in up-regulation of Mcl-1, a p53 inhibitor, and may contribute to survival of breast cancer cell lines [57]. Additionally, genome-wide studies in breast cancer cell lines have revealed that ELK1 is involved in the activation of c-Fos, a proto-oncogene that is implicated in tumorigenesis [58]. These studies verify that many TF binding motifs we find to be enriched in breast cancer subtypes are biologically meaningful in the context of breast cancer.
When analyzing TF-DNA methylation relationships in breast cancer subtypes, we build upon conventional methods of cancer stratification. However, in order to analyze TF motif enrichment within a classification scheme focused on DNA methylation, we adopted a bottom-up approach by first classifying all CpGs into subtypes based on their intensity levels. Since many cancers show genome-wide changes in DNA methylation, this approach may be able to identify TFs that are directly related to distinct intensity levels of DNA methylation. Therefore, we created a class of subtypes based on the clustering of CpG β-values and calculated TF binding motif enrichment in these subtypes. Fig 6A shows CpGs organized into 5 clusters based on β-values, with high intensity clusters on top and low intensity clusters on the bottom. From C1 to C5, the clusters are enriched in 68 (50 TFs), 45 (31 TFs), 6 (6 TFs), 6 (5 TFs), and 87 (59 TFs) TF binding motifs, respectively (P<0.05) (Fig 6A). Furthermore, we identified 119 (80 TFs), 38 (24 TFs), 3 (3 TFs), 1 (1 TFs), and 10 (8 TFs) TF binding motifs that were significantly depleted from C1 to C5, respectively (P<0.05) (S18 Table). Like histological and intrinsic subtypes of breast cancer, certain TF binding motifs exhibit different levels of enrichment across CpG subtypes. To globally illustrate the variation in TF motif enrichment between CpG subtypes, we sorted significant motifs in cluster 1 (C1) (P<0.01) from most enriched to most depleted (Fig 6B). We then ordered the TFs in the other 4 clusters relative to those belonging to cluster 1 (Fig 6B). From this, it is clear that related clusters share common patterns of enrichment (i.e. patterns in cluster 1 are more similar to that of cluster 2 than cluster 5) (Fig 6B). Interestingly, cluster C1, which contains highly methylated CpGs, is both enriched and depleted in TF binding motifs (Fig 6A and 6B). In contrast, cluster C5, which contains lowly methylated CpGs, is characterized mainly by TF binding motif enrichment events and few TF binding motif depletion events. This suggests that TF binding is generally associated with reduced methylation levels. Additionally, clusters C3 and C4 contain very few high-significance enriched/depleted TF binding motifs, suggesting that mid-intensity methylation are stochastic events and are not as informative for identifying important breast cancer-associated regulators.
To provide an example, we illustrate TR-NFY_01, which shows highest enrichment in C5 and lowest enrichment in C1 (Fig 6C). It can also be observed that its enrichment level increases from C1 to C5 (Fig 6C). This suggests that these CpG clusters have functional relevance in the context of NF-Y binding. NFY is known to be essential for proper cell cycle regulation and mutation of this protein can lead to inhibition of Cyclin A, RNR R2, DNA polymerase, CDC2, Cyclin B, and CDC25C [59]. Moreover, Agostino et al. showed that NF-Y facilitates gain-of-function p53 mutant binding to NF-Y promoters, resulting in cell cycle misregulation in breast cancer cell lines [60]. We also highlight TR_E47_01, which exhibited highest enrichment levels in C1 and lower enrichment levels in clusters least similar to C1, suggesting that E47 binding sites tend to be highly methylated in breast cancer (Fig 6D). E47 (also known was TCF3) is a repressor of E-cadherin and its activity has been implicated in epithelial-mesenchymal transition events in breast cancer [61].
In order to demonstrate differences in the regulatory interactomes of breast cancer subtypes, we constructed two TF-TF interaction networks for ER+ and ER- samples (see Materials and Methods). Each network illustrates the first order partners of TFs whose motifs are significantly enriched (depletion is excluded) (P<0.01) in ER+ and ER- samples (Fig 7). Interestingly, in ER- breast cancer, ESR1 (ERα), RELA, SP1, and AR exhibit the highest degree in the network (Fig 7). Consistent with our prior results, it can be observed that ESR1 is significantly enriched in protective and hazardous CpGs in the ER- network only (Fig 7). In addition to ESR1, SP1 also exhibits high-degree in both ER+ and ER- networks; however, it is enriched in hazardous CpG regions in ER+ whereas, in ER-, it is enriched in protective CpG regions (Fig 7). This demonstrates that TF-DNA methylation relationships vary depending on disease context.
The effects of DNA methylation are widespread and vary according to genomic context and interactions with TFs. In this study, we proposed a novel method of inferring TF-DNA methylation relationships in breast cancer by utilizing both differential methylation and survival analysis to pinpoint informative CpGs. From these CpGs we were able to delineate TF involvement with methylation patterns and extend that to patient prognosis. Many mechanisms by which DNA methylation interacts with TFs have been proposed. It has been suggested that methylated CpGs can act as a direct physical hindrance to TF binding and thus interfere with its regulatory functions. Additionally, DNA methylation can recruit chromatin remodelers (or proteins that then recruit chromatin remodelers) to compact chromatin, assist in transcription elongation, or merely act as a passive marker of protein binding. This variety of potential mechanisms is key to understanding why the TF binding motifs of some experimentally verified oncofactors (oncogenic TFs) in our analysis were enriched in protective regions and vice versa for tumor suppressor motifs. For example, if methylated CpGs in the binding site of an oncofactor obstructs binding, then hypermethylation will ablate the oncofactor’s regulatory effect and promote survival. Alternatively, it is also possible for tumor suppressor TF binding motifs to be enriched in protective CpG regions if the role of the tumor suppressor is to recruit DNA methyltransferases to silence oncogenes. Additionally, the genomic location of CpGs may have an effect on its regulatory activity; methylated CpGs in promoters may inhibit gene expression but methylated CpGs in gene bodies may aid in transcriptional elongation [27]. Therefore, genomic context and prior knowledge of the TFs relationship with methylated CpGs must be established before reasonable conclusions can be made. As more experimental data is generated regarding these relationships, a TF-DNA methylation interactome network will be of greater use. Here, we have provided evidence that the application of motif enrichment, survival analysis, and differential methylation analysis can be integrated and used to define TF-DNA methylation interactomes in various subtypes of breast cancer. All in all, this study links TF-binding to DNA methylation to overall patient prognosis. By embracing the complexity of misregulation in various breast cancer subtypes, it may be possible to identify key players responsible for cancer subtypes and use that information to guide the development of treatment regimens in the clinic.
Furthermore, our preliminary analysis shows that correlating gene expression with survival yields very few significant genes after multiple hypotheses testing correction (S19 Table). This can be due to post-transcriptional modifications that affect the stability of mRNA transcripts, the fact that mRNA abundance is not always a good proxy for protein activity, and the short lifespans of cancer patients in our datasets. In light of this, the alternative use of DNA methylation signals can reveal significant CpGs even after testing corrections. This may be due to the fact that it is a binary chemical modification that is stable and can, in some cases, better reflect regulatory activity.
The most striking result of our analysis is that the majority of ERα and p53 TF binding motifs are significantly enriched in survival-associated CpG regions in ER- and p53- samples, respectively. From these results, we propose that ERα TF binding motifs that are not bound by their respective TF (as in the case of ERα in ER-samples) may become bound by alternative factors that may cause misregulation of downstream genes and impact patient survival. Indeed, this may also be the case for p53 in p53- samples. The reasoning for this model begins with the observation that the TF binding motifs of a master regulator are enriched in CpG regions whose methylation is correlated with survival only in samples missing that regulator. If a TF is missing, why would a significantly large proportion of its binding sites be enriched in these informative CpG regions? Therefore, we suspect that these motifs are open to binding by alternative factors whose binding events may simultaneously cause gene misregulation and detectable alterations in DNA methylation. Additionally, an alternative explanation may be that DNA methylation in p53 or ERα binding sites passively reflects the lack of p53 or ERα activity, respectively. This is valid in the case of p53- samples, where the loss of a key tumor suppressor would result in longer survival times compared to p53+ patients. However, in the case of ERα, its exact cancer regulatory roles are not as clear and thus difficult to interpret.
To explore breast cancer TF-DNA methylation relationships in depth, we adopted three different classification schemes by which we divide samples based on histological, molecular, and methylation features. Since each subtyping method utilizes information at different levels (i.e protein, gene expression, methylation) it is sensible to adopt all three strategies. By calculating TF binding motif enrichment in different subtypes, we can effectively determine the similarities and differences in their TF-DNA methylation signatures. We also observe differential enrichment patterns between protective and hazardous CpG sets among histological and intrinsic breast cancer subtypes, suggesting that TF-DNA methylation relationships vary across subtypes.
We also complemented our analyses by calculating whether TFs whose motifs were enriched in differentially methylated or survival-associated CpG regions were also differentially expressed in mRNA levels between ER+ and ER-, and between normal and tumor patient samples, respectively. We found that many of TFs with enriched motifs were also significantly differentially expressed, suggesting a greater biological role for these TFs (S1–S12 and S14–S17 Tables). Additionally we validated many of our identified TF binding motifs in an independent DNA methylation dataset published by Dedeurwaerder et al. [62]. This dataset was generated from the Illumina HumanMethylation27K (HM27K) array which profiled ~27,000 CpG sites from 248 primary breast tumors. Particularly we found that the TR-ER_Q6 and TR-ER_Q6_02 motifs were both significantly enriched ~1.3-fold (unadjusted P = 0.02 and P = 0.008, respectively) in hypomethylated (ER+<ER-) CpG regions between ER+ (n = 132) and ER- (n = 101) samples (S1 Table). Moreover, we found that GATA3 and FOXO1 TF binding motifs were enriched in hypomethylated regions (P<0.05), which is consistent with results from the main TCGA dataset. We then extended our validation analysis to include the enrichment calculation of TF binding motifs in survival-associated CpG regions identified across all breast tumor samples. In particular, we sought to confirm our enrichment results for the top 20 protective and top 20 hazardous TF binding motifs in the Dedeurwaerder dataset, and were able to validate 19 out of the 20 protective and 16 out of the 20 hazardous TF binding motifs (P<0.05, S2 Table). To note, we did not perform the CpG filtering procedure in the Dedeurwaerder dataset since the number of CpGs interrogated by the Illumina HM27K array was substantially less than the Illumina HM450K array used by TCGA, resulting in a sizeable decrease in statistical power. Together, these results indicate that our analysis remains robust across independent datasets even when different genomic platforms are used.
We concede that there are limitations to the informativeness and interpretation of our results. First, we used a 102 bp region to define a CpG region, which restricts our analysis to a local binned area. Even though most binding events only encompass ~100 bp, it may be possible that this sequence space may encompass the binding sites of TFs that are not associated with the CpG residue leading to false positives. On the other hand, the 102 bp region may be too small and not encompass the binding sites of TFs that do in fact interact with the CpG resulting in false negatives in our enrichment analysis. Overall, we experimented with varying sequence region sizes and determined that our results remain stable. Second, because we restricted our analysis to a local region, we do not take into account any potential long-range effects CpG sites may have on TF binding as a result of chromatin orientation. Third, limitations in platform technology must also be taken into account since only 450,000 out of a total of ~30 million CpGs are probed by the Illumina HumanMethylation 450K array. Fourth, we acknowledge that there may be differences in statistical power when conducting enrichment analysis in CpG sets (e.g. protective, hazardous) since the number of CpGs in each set may vary. Lastly, many of the analysis steps in our methodology including motif detection and setting significance criteria for survival-associated CpGs suffer from high false positive rates. However, we were able to overcome these obstacles by calculating relative enrichment of TF binding motifs. Therefore, even though we were able to identify significantly enriched TF binding motifs, our enrichment scores were ultimately biased towards the null (RE = 1). Overall, we maintain that our method has produced results that provide new insight into TF-DNA methylation relationships in breast cancer despite these limitations.
In this study, we have developed a novel method for identifying transcriptional regulators involved in breast cancer in the context of patient survival by using DNA methylation data derived from primary breast tumor tissue. By doing so, we have provided insight into the complexity of TF-DNA methylation interactomes that underlie breast cancer across a wide variety of subtypes. Our analysis has revealed several informative results and, in addition, raises a manifold of new questions regarding cancer misregulation. Namely, we have identified transcriptional regulators that affect patient prognosis and proposed a new model whereby breast carcinogenesis may be driven via binding of alternative factors to unbound TFBSs. Additionally, we considered the heterogeneity exhibited by breast cancer tumors by characterizing TF-DNA methylation relationships in histological, molecular, and DNA methylation subtypes. In this analysis, we focused on well-defined TF binding motifs, but it is also possible to combine this analysis with de novo motif identification to identify novel motifs that are enriched in differentially methylated or survival-associated CpG regions. Such integration could also allow for an exhaustive and systematic identification of non-TF regulators that may also interact with methylated CpGs (e.g. non-coding RNAs). Ultimately, our study has provided deep insight into the differential regulatory wiring of breast cancers that occur due to the divergent and combinatorial effects of diverse mutations.
Breast invasive carcinoma (BRCA) Level 3 DNA methylation datasets for the JHU-USC HumanMethylation450K platform, CpG annotation files, and clinical information were downloaded from the TCGA data portal [20]. CpG methylation signal intensities were represented as β-values in the datasets. In addition, Level 3 TCGA UNC AgilentG4502A_07 mRNA expression data was downloaded from the site data portal [20]. Subtype classification for all patient samples was derived from TCGA clinical information [20]. Breast cancer DNA methylation data and clinical information from Dedeurwaerder et al. was downloaded from the Gene Expression Omnibus (GEO) under the accession number GSE20713. PWMs for human TFs were obtained from the TRANSFAC [63] and JASPAR [64] databases. In some cases there were multiple PWMs for a single TF. The TF-TF physical interaction data were compiled from two resources: an experimental dataset from Ravsi et al [65] containing 5238 TF-TF physical protein interactions across 1400 human TFs, and the human protein reference database [66].
Patient samples that were accompanied with histological information regarding ER status were split into ER+ (405 samples) and ER- (122 samples) groups. Differentially methylated CpGs were then identified using two-tailed student t-test. To achieve stringency while maintaining power for enrichment analysis, a Benjamini-Hochberg adjusted P<0.05 was chosen as the cutoff for differentially methylated CpGs. CpGs with p-values below the cutoff with a t-statistic >0 and <0 were categorized into hypermethylated (ER+>ER-) and hypomethylated (ER+<ER-) sets, respectively. A Wilcoxon ranked sum test was also implemented to identify differentially methylated CpG sites.
RNA-seq data for 1154 breast cancers was downloaded from the TCGA data portal. Differentially expressed genes were identified using a Wilcoxon ranked sum test. A fold change >1 indicated gene up-regulation and a fold change <1 indicated down-regulation. Differentially expressed genes corresponding to TFs were included in all supplementary tables to complement TF enrichment information.
The β-values for 376,667 CpGs across 562 samples and clinical data were used as input into a univariate Cox proportional hazards model [40]. Each CpG was considered individually and used as the covariate in the model. Significance of model coefficients was calculated using the Wald test. CpGs that yielded unadjusted P<0.02 and a hazard ratio of <1 or >1 were labeled as protective and hazardous, respectively.
Samples were categorized into histological and intrinsic subtypes based on the clinical information downloaded from TCGA. The β-values for 376,667 CpGs across subtype-only samples were used as input into the Cox proportional hazards model where each CpG was considered individually. This allowed for the identification of survival-associated CpGs significant in the particular subtype.
The 102 bp sequence region centered at each significant CpG was used as input into the FIMO software package [67] from the MEME suite to identify the existence of a motif in the region. Default parameters were used and a threshold cutoff of P<1E-4 was used to determine the presence of a motif. This yielded a matrix containing Boolean values indicating if a particular TFBS motif was present in a CpG region.
The top 10,000 most significant differentially methylated CpG regions (hyper- or hypo-) were chosen as input into the Discriminative Regular Expression Motif Elicitation (DREME) algorithm (MEME suite) using default parameters, with the exception of the maximum motif size, which was set to 20 [68]. Identified motifs were then queried against the JASPAR vertebrate database using Tomtom (MEME suite) to identify their cognate TFs [69].
Because overlapping CpG regions can lead to over-counting of TF binding motifs, we filtered out all overlapping CpG regions in each chromosome for forward and reverse DNA strands. The filtering procedure proceeds as such: (i) CpGs located on different chromosomes were considered non-overlapping. (ii) CpGs that were located on different DNA strands were considered non-overlapping. (iii) Sort all CpGs based on their genomic coordinates and identify clusters of CpGs with overlapping 102bp regions. (iv) For each cluster, identify the CpG with the lowest p-value (differentially methylated or survival-associated depending on analysis) and set this as the reference CpG. (v) Identify all within-cluster CpGs whose regions do not overlap with that of the initial reference CpG and filter out the rest. (vi) Of the non-overlapping CpGs, select the one with lowest P-value and set as the new reference CpG (vii) Iterate until all CpGs have either been selected or filtered out. (All “reference” CpGs are then included in the subsequent motif enrichment analysis.)
To compute enrichment of TFs in functional CpGs (survival-associated, differentially methylated, or clustered CpGs), we applied a two-sided Fisher’s exact test for each TF binding motif to determine if it was overrepresented or underrepresented in a CpG set. The Fisher’s exact test involves calculating the hypergeometric probabilities of all possible matrices of a 2X2 contingency table while keeping the margin counts fixed. The probabilities of all possible fixed-margin contingency tables more extreme than the current table were summed to determine the probability of over-representation/under-representation of a motif to occur by random chance. The R function “fisher.test” was used to implement this computationally. Enrichment was calculated in protective, hazardous, and survival-associated (protective & hazardous) CpG sets for histological, intrinsic, and CpG (CpG clusters) breast cancer subtypes. Additionally, motif enrichment was implemented in differentially methylated CpGs between ER+ and ER- breast cancers. The Benjamini-Hochberg multiple hypothesis testing correction procedure [70] was used to adjust the P-values outputted by multiple Fisher’s exact tests. All P-values presented in the Results section had been adjusted for multiple testing.
When comparing the distributions of TF binding motif enrichment values between hyper- and hypomethylated CpGs (Fig 2A), we first controlled for the potential effects that sample size may have on the power of enrichment analysis. This issue may arise due to the unequal number of CpGs belonging to hyper- and hypomethylated CpG sets. Therefore, we took the top n most significant CpGs from each set, where n is the smallest number of CpGs between the two sets, and carried out enrichment analysis.
TF binding motifs with P<0.01 in ER+ and ER- samples, and their first-order interacting partners were extracted from the TF-TF physical interaction network and used as input into Cytoscape to construct a regulatory network. TF network analysis was implemented using the “NetworkAnalyzer” function included in the software. The size of network nodes were mapped to the enrichment P-values of TFs represented in the network (lower P-values correspond to larger nodes). The font sizes of TF names were mapped to node degree in the network (larger font sizes correspond to higher degree). These mappings were implemented using Cytoscape’s VizMapper tools. If multiple motifs belonging to the same transcription factor fell below the significance threshold, their p-values were averaged.
K-means clustering was applied to cluster CpGs based on their β-values. To determine the number of clusters, k-means was applied using 1–10 clusters and the total within-cluster sum of squares (WCSS) was calculated and graphed. Classification into 5 clusters yielded the last point where there is noticeable decrease in total WCSS.
All statistical and computational analyses were implemented in the R statistical programming environment.
|
10.1371/journal.ppat.1005711 | Staphylococcal Bap Proteins Build Amyloid Scaffold Biofilm Matrices in Response to Environmental Signals | Biofilms are communities of bacteria that grow encased in an extracellular matrix that often contains proteins. The spatial organization and the molecular interactions between matrix scaffold proteins remain in most cases largely unknown. Here, we report that Bap protein of Staphylococcus aureus self-assembles into functional amyloid aggregates to build the biofilm matrix in response to environmental conditions. Specifically, Bap is processed and fragments containing at least the N-terminus of the protein become aggregation-prone and self-assemble into amyloid-like structures under acidic pHs and low concentrations of calcium. The molten globule-like state of Bap fragments is stabilized upon binding of the cation, hindering its self-assembly into amyloid fibers. These findings define a dual function for Bap, first as a sensor and then as a scaffold protein to promote biofilm development under specific environmental conditions. Since the pH-driven multicellular behavior mediated by Bap occurs in coagulase-negative staphylococci and many other bacteria exploit Bap-like proteins to build a biofilm matrix, the mechanism of amyloid-like aggregation described here may be widespread among pathogenic bacteria.
| Major components of the biofilm matrix scaffold are proteins that assemble to create a unified structure that maintain bacteria attached to each other and to surfaces. We provide evidence that a surface protein present in several staphylococcal species forms functional amyloid aggregates to build the biofilm matrix in response to specific environmental conditions. Under low Ca2+ concentrations and acidic pH, Bap is processed and forms insoluble aggregates with amyloidogenic properties. When the Ca2+ concentration increases, metal-coordinated Bap adopts a structurally more stable conformation and as a consequence, the N-terminal region is unable to assemble into amyloid aggregates. The control of Bap cleavage and assembly helps to regulate biofilm matrix development as a function of environmental changes.
| Biofilm formation is universal for all bacteria. The molecular mechanisms governing this process vary among bacteria, but they all culminate in the synthesis of an extracellular matrix. The composition of the extracellular matrix is complex and variable, even within the same bacterial species when environmental conditions are altered [1,2]. However, one common principle is that the matrix scaffold is built from exopolysaccharide or proteins, which eventually can be interwoven with extracellular genomic DNA [3–5]. The reasons underlying the election of a polysaccharide or protein-based biofilm matrix are not well understood, but an increasing number of studies indicate that proteinaceous scaffolds are more common than previously anticipated. Proteins anchored to the bacterial cell surface can assemble the matrix scaffold through homophilic interactions between identical molecules expressed on neighboring cells or through heterophilic interactions with other surface proteins or with non-proteinaceous cell wall structures [6,7]. Members of this group of proteins include autotransporter adhesins [8–11], carbohydrate-binding proteins [12–14], and cell-wall anchored proteins covalently linked to the peptidoglycan (CWA) [2,15–21]. Another strategy by which proteins can contribute to the formation of the matrix scaffold is through polymerization into functional amyloid fibers. Secreted proteins can assemble to form insoluble fibers with a characteristic cross-β-strand structure, where the β-sheets run perpendicular to the fibril axis [22]. Once polymerized, amyloid fibers constitute a strong platform able to mediate interactions between the neighboring cells and surfaces [23–26]. Examples of amyloid fibers mediating biofilm development include curli pili present in Enterobacteriaceae [27,28], FapC in Pseudomonas fluorescens [29], TasA in Bacillus subtilis [30], the aggregative flexible pili named MTP in the pathogen Mycobacterium tuberculosis [31,32] and phenol soluble modulins (PSMs) in Staphylococcus aureus [33].
Biofilm associated proteins (Bap) are high molecular weight multi-domain proteins, characterized by a repetitive structure and localized at the cell surface [34]. The first member of this family of proteins was identified in a mobile pathogenicity island (SaPIbov2) present in some strains of S. aureus. So far, the bap gene has been identified in mastitis-derived staphylococcal species, but has never been found in S. aureus human isolates. However, bap orthologous genes are present in the core genome of several coagulase-negative staphylococcal species that belong to the human commensal microbiota such as S. saprophyticus ATCC15305 (Accession number GCA_000010125.1), S. epidermidis (GCA_000759555.1) and S. warneri SG1 (GCA_000332735.1) [35]. Bap promotes the initial attachment to inert surfaces and cell-to-cell interactions through a mechanism that is independent of exopolysaccharide [21,36]. During infection, Bap facilitates the persistence in the mammary gland by enhancing adhesion to epithelial cells and prevents cellular internalization through the binding to GP96 host receptor, which interferes with the FnBPs mediated invasion pathway [37,38]. Overall these results indicated that Bap plays a dual function: on the one hand, mediating bacterial-bacterial interactions and on the other, bacterial-host interactions. However, the molecular mechanisms by which Bap performs these functions and the region of the protein involved in each process remain unexplored.
In this report, we investigated the mechanistic basis by which Bap proteins promote the formation of the biofilm matrix scaffold. Our results have shown that Bap is constitutively expressed along the growth curve and processed. The resulting fragments, which likely contain mainly the N-terminal region of the protein, form insoluble amyloid–like aggregates when the pH of the media becomes acidic and the concentration of calcium is low. If calcium concentration increases, metal-coordinated Bap adopts a more stable conformation as shown by thermal denaturation monitored by intrinsic fluorescence, nuclear magnetic resonance (NMR), proteinase K digestion and analytical ultracentrifugation. As a consequence, the N-terminal region is unable to self-assemble and to mediate intercellular aggregation and biofilm formation. Furthermore, we show that biofilm assembly by Bap orthologs also depends on the critical N-terminal domain suggesting that the mechanism of biofilm assemblage is conserved in staphylococci. In view of these results, we propose that Bap plays dual role in the bacterial physiology, acting as a sensor and promoting biofilm formation, a configuration that has not hitherto been described for any component involved in biofilm formation.
To investigate the molecular mechanisms underlying Bap-mediated staphylococcal biofilm development, we monitored the expression of Bap in rich liquid media (LB-glu) along the growth curve using native and denaturing gel electrophoresis. Western immunoblotting under denaturing conditions revealed the presence of Bap from early stages of growth until the population entered stationary phase (OD600nm = 5). From that point, the levels of Bap decreased significantly in denaturing gels, whereas a band of high molecular weight appeared in the native gels suggesting that Bap formed aggregates when bacteria entered stationary growth phase (Fig 1A).
When S. aureus is grown in a media containing glucose, the entry in stationary phase is accompanied by a decrease in pH due to the accumulation of acidic byproducts from glucose fermentation [39]. We therefore investigated whether Bap aggregation and Bap mediated biofilm development were related and occurred in response to changes in the media pH. To investigate this hypothesis, bacteria were grown in LB-glu, where the pH levels dropped below 5 when bacteria reached stationary phase (OD600nm = 5) and in LB without glucose, where the pH remained neutral all along the growth curve (Fig 1B). Western immunoblotting revealed that Bap failed to form protein aggregates when bacteria grew in LB (Fig 1C). Moreover, results showed a strong correlation between Bap protein aggregates and Bap mediated biofilm formation since bacteria grown in LB-glu (pH<5) formed bacterial clumps and strong biofilms in microtiter plates whereas in LB media (neutral pH) Bap failed to promote bacterial clumping and biofilm development (Fig 1D and S1A Fig). To further corroborate the effect of pH on aggregation of Bap-positive strains we evaluated cell clumping of S. aureus V329 and ∆bap grown in LB medium acidified with 0.1 M HCl to pH 4.5. After an overnight incubation V329 wild type strain clearly showed a biofilm adhered to the microtiter plate and bacterial clumps at the bottom of the tube, while ∆bap strain did not (S2A and S2B Fig). Moreover, the two strains were also grown in LB-glu and, after an overnight incubation the medium was replaced by LB to evaluate the possible disassembly of bacterial aggregates. Indeed, no bacterial clumps and no biofilm were observed after media were exchanged indicating that the process of interbacterial interaction mediated by Bap is reversible when pH arises (S2C and S2D Fig). Together, these results suggest that acidification of the growth media promotes Bap aggregation and biofilm development.
If Bap is engaged in homophilic interactions during biofilm development, Bap aggregates should be composed primarily of the Bap protein. In contrast, if Bap mediates heterophilic interactions with other surface proteins, Bap aggregates should also contain additional proteins. To distinguish between these possibilities, we determined the protein content of the Bap aggregates by recovering the insoluble protein material retained within the wells of the stacking gel from preparations of S. aureus grown in LB-glu and LB and analyzing their identity by mass spectrometry (MS). MS analysis of the material retained in gel pockets from preparations of S. aureus grown in LB-glu identified peptides that corresponded mostly to the Bap protein strongly suggesting that this polypeptide is the main constituent of the aggregates (S3 Table). Apart from Bap and some ribosomal proteins, the vast majority of the other identified peptides corresponded to proteins that were also detected by MS analysis from preparations of S. aureus grown in LB medium, where no presence of Bap was observed (S3 Table). We also discarded the possibility that additional matrix molecules such as PNAG or eDNA could be involved in the formation of the Bap aggregates since neither Bap insoluble aggregates nor Bap-mediated biofilms were affected by the treatment with dispersin B (DspB) and DNase I (S3A, S3B and S3C Fig). MS analysis also showed that the identified peptides covered the N-terminal sequence of mature Bap almost completely (amino acid 49 to 819), but only a single short peptide from C repeats region (Fig 2A and 2B) was observed. These results suggest that the insoluble aggregates likely contain Bap fragments that at least include N-terminal region.
To further investigate the mechanism of Bap proteolytic cleavage, we performed western immunoblotting of cell wall extracts of S. aureus V329 grown in LB-glu and LB culture conditions. Results revealed that Bap was proteolytically processed in both media (Fig 2D), but the cleavage products obtained in LB were unable to form high molecular weight aggregates (Fig 2C). Western immunoblotting of surface proteins from cells grown in LB-glu extracted at different points of the growth curve showed the presence of degradation bands that increased in number and intensity as bacteria grew (S4A Fig). Mass spectrometry analysis of the largest processed band confirmed that it corresponded to a degradation product of Bap containing at least the N-terminal region of the protein (S4B Fig). Interestingly, when bacteria entered stationary phase (OD600nm≈4, pH<5) bands corresponding to full-length and the resulting processed fragments of Bap disappeared from the gel and insoluble aggregates recognized by anti-Bap antibody were readily detectable (S4B Fig).
With the aim to identify the extracellular proteases responsible for the proteolytic processing of Bap, we constructed mutants in 3 extracellular proteases: a serine protease (V8 protease; SspA), a cysteine protease (SspB) and its specific inhibitor (SspC) and a metalloprotease (aureolysin; Aur). The resulting protease-deficient strains showed similar Bap cleavage patterns and formed cell clumps and biofilm at levels similar to wild-type strain (S4 Fig, left panels). Besides, addition to the culture media of protease inhibitors such as α2-macroglobulin, E-64 (cysteine protease inhibitor), PMSF (serine protease inhibitor) and the inhibitor Staphostatin A (ScpB) that specifically targets the extracellular cysteine protease ScpA, did not interfere with Bap-mediated aggregation and biofilm development (S4 Fig, right panels). Taking together these findings suggest that Bap is processed either by spontaneous cleavage or by the activity of a protease different from the ones tested here, or perhaps by the action of more than one protease. The resulting processed products are more aggregation prone and form the high molecular weight aggregates under acidic conditions.
These latter observations lead us to consider that the N-terminal region of Bap may be sufficient to promote biofilm development. To assess this hypothesis, we generated chimeric proteins comprising different regions of Bap tagged with the 3xFLAG amino acid sequence and linked to the R domain of the clumping factor A (R-ClfA) containing the LPXDG motif (Fig 3A). Variants of Bap comprising domain A (Bap_A, amino acid residues 49 to 361), domain B (Bap_B, amino acid residues 362 to 819), or domain A and B (Bap_AB, amino acid residues 49 to 819) were cloned in pCN51 vector under the control of the Pcad-cadC promoter and expressed in S. aureus ∆bap. The expression of the whole Bap or the chimeric Bap proteins on the bacterial cell wall was verified by western-blot and immunofluorescence using strains deleted in their spa gene (V329 Δspa and ΔbapΔspa) to avoid unspecific antibody labeling of protein A through its union to the Fc fraction of immunoglobulins (S5 Fig). S. aureus producing Bap_AB or Bap_B formed huge cell-to-cell aggregates (Fig 3B) and robust biofilms on polystyrene (Fig 3C and S1B Fig) or on a glass surface under flow culture conditions (Fig 3D). In contrast, no cell clusters and biofilm development were found in S. aureus Bap_A and ClfA strains.
The observation that domain B of Bap is sufficient to induce biofilm phenotype suggests that Bap_B functionality could be affected by the pH. Similarly to the Bap full-length protein, Bap_B formed high molecular weight aggregates when a culture of ∆bap strain expressing Bap_B reached stationary phase (Fig 3F). Accordingly, this strain formed biofilm when it was grown in LB-glu (pH<5), but not in LB (Fig 3E and S1C Fig), and showed bacterial clumping and biofilm formation when grown in LB acidified with 0.1 M HCl (S2A and S2B Fig). Bacterial aggregates formed by ∆bap expressing Bap_B chimeric protein in LB-glu (pH<5) were disassembled when the medium was exchanged for LB (pH>7) (S2C Fig).
Next, we explored whether Bap_B was sufficient to confer cell-to-cell interactions to naturally bap deficient strains: S. aureus MW2, S. aureus Newman and S. carnosus TM300. As shown in Fig 3G and S1D Fig, expression of Bap_B in these strains conferred strong bacterial clumping capacity after an overnight incubation in LB-glu. Taking together, these results indicated that the B domain of Bap (amino acids 362 to 819) is sufficient to bestow multicellular behavior under acidic culture conditions, similarly to the entire Bap protein.
To get insights about the molecular mechanisms by which the N-terminal region of Bap mediates cell-to-cell interactions, we used a purified recombinant protein comprising exclusively the B region of Bap (rBap_B). Purified rBap_B formed a visible ring of protein adhered to the walls of the tube when incubated at acidic pH in a grade of pH from 3.6 until 5 (S6A Fig). Interestingly, the process was reversible and rBap_B aggregates dissociated completely when the pH was raised to 7 (Fig 4A). To validate the functionality of rBap_B, we analyzed the capacity of rBap_B to restore bacterial clumping phenotype of S. aureus ∆bap. Exogenous addition of rBap_B protein (2 μM) induced bacterial clumping only when S. aureus ∆bap was grown under acidic culture conditions (Fig 4B).
Next, we performed a biophysical characterization of the rBap_B domain. First, we determined the relative size of the aggregates by dynamic light scattering (DLS). The graphic in Fig 4C illustrates the size characterization (hydrodynamic radius R) of rBap_B in solution at different pH. In the table, the correlation between the diffusion coefficient (D) of each population and its corresponding radius is shown. It can be observed that pH 4.4 is the condition at which rBap_B protein presented the highest hydrodynamic radius and the lowest D value (peak 2), as expected for aggregated particles that move slower that smaller particles, with a polydispersity percentage below 15% characteristic of monodispersed samples (Fig 4C). At pH 3, rBap_B presented protein populations with intermediate R values. At neutral pH the obtained peak showed a D that once substituted in the Svedberg equation, together with the previously obtained experimental sedimentation coefficient, buffer density and partial specific volume, corresponded to the monomer of the protein (Fig 4C). Far-UV circular dichroism spectra (CD) of rBap_B showed a moderate increase in β-sheet structure (+5%) when the pH was acidified, at the expense of the predominant non-regular secondary structure (Fig 4D and S4 Table).
Next, we analyzed more in depth these β-sheet-rich rBap_B aggregates formed at acidic pH. We examined the amide I region of the Attenuated Total Reflectance–Fourier Transform Infrared spectroscopy (ATR-FTIR) spectrum (1700–1600 cm-1) of rBap_B ring assemblies. This region corresponds to the absorption of the carbonyl peptide bond group of the protein main chain and is a sensitive marker of the protein secondary structure. Deconvolution of the FTIR-absorbance spectra allowed us to assign the individual secondary structure elements and their relative contribution to the main absorbance signal. The FTIR spectra of rBap_B aggregates was dominated by β-sheet/β-turn components contributing >80% to the signal. In particular, the strong bands at 1628 and 1694 cm-1 were consistent with the presence of amyloid-like intermolecular β-sheet structure (Fig 5A). To assess whether the prevalent intermolecular β-sheet in the rings formed by rBap_B aggregates was amyloid-like in nature, we evaluated the binding of the aggregates to the amyloid diagnostic dyes Thioflavin-T (ThT), Congo Red (CR) and ProteoStat. The presence of rBap_B aggregates induced a 25-fold increase in ThT maximum fluorescence emission (Fig 5B). Interestingly, when fresh rBap_B was incubated at pH 4.5 for 5 min it bound readily to ThT in a concentration dependent manner, indicating a fast assembly of rBap_B into ThT positive structures (S7A Fig). In contrast, no change in ThT fluorescence was observed when the protein was incubated at pH 7.0, independent of the protein concentration assayed (S7B Fig). The fast assembly of rBap_B at pH 4.5 was also evident from the strong increase in light scattering relative to the signal obtained at pH 7.0. (S7C Fig). These early assemblies displayed a strong binding to the dye bis-ANS, evidencing the presence of hydrophobic patches exposed to solvent, which potentially might recruit rBap_B monomers into the aggregates and/or contact other cellular molecules through non-polar interactions (S7D Fig). In agreement with an amyloid-like conformation the absorbance of CR and its spectrum maximum red-shifted in the presence of rBap_B ring aggregates (Fig 5C). The absorbance of ProteoStat, a novel fluorescent dye able to stain specifically amyloid aggregates in vivo [40], showed a 20-fold increase in its fluorescence maximum at 550 nm (Fig 5D). Altogether, these data strongly suggest that the intermolecular β-sheet structures formed upon aggregation of rBap_B at pH 4.5 posses an amyloid-like conformation.
In amyloid-like aggregates, short sequence fragments usually promote and guide the formation of amyloid-like structures and become embedded in the inner core of the cross-β structure [41–43]. To identify the likely amyloidogenic regions in the series of Bap_B peptides previously identified by MS in the biofilm we used computational algorithms: AGGRESCAN [44], PASTA [45], WALTZ [46] and ZipperDB [47]. The predictions converged to indicate two Bap_B short sequence stretches as potentially amyloidogenic: TVGNIISNAG named as peptide I (aa 487 to 496), and GIFSYS named as peptide II (aa 579 to 584) (Fig 2B). We synthetized the two peptides and incubated them at 10 μM at pH 4.5. Both peptide solutions formed an evident gel (S8 Fig), a property shared by many amyloidogenic peptides [48] as well as biofilm matrices [49]. Analysis of the structure of the two gels by transmission electron microscopy (TEM) indicated that they comprise fibrils with a typical amyloid morphology (Fig 6A and 6B), and bound to ThT with high affinity (Fig 5E). Taken together these data indicate that Bap_B contains at least two short regions with high amyloidogenic propensity. Analysis of these peptides using the RosettaDesign program [50] implemented in ZipperDB [51] rendered average interaction energies of -25.0 and -25.9 kcal/mol for peptide I and peptide II and shape complementarities between strands of 0.87 and 0.81, respectively. These parameters are compatible with these peptides being able to form steric-zippers that might contribute to N-terminal Bap amyloid assembly. Of course, we cannot discard that the presence of several additional or alternative short amino acid stretches with amyloidogenic tendencies in the sequence of B-domain of Bap would be required for the assembly of the complete domain into aggregated structures at acidic pH.
We next determined by transmission electron microscopy the presence of amyloid fibers. Electron microscopy analysis of the aggregates formed by purified rBap_B at pH 4.5, first revealed the presence of isolated fibers and fibers entangled in larger electron dense aggregates (Fig 6C). Similar fibers were detected in the surface of S. aureus Δbap when bacteria were grown in the presence of exogenously added rBap_B under acidic culture conditions (Fig 6D). We further analyzed the presence of fibers in wild-type S. aureus V329 grown under biofilm forming conditions. Consistent with all the findings obtained for the rBap_B domain, S. aureus V329 contained fibers (Fig 6E and 6F) that specifically reacted with gold-labelled anti-Bap antibody (Fig 6G and 6H). Finally, we determined the presence of amyloid fibers in the Bap-mediated biofilm by staining the extracellular matrix of S. aureus V329 grown in LB-glu medium with ProteoStat, a dye specific of amyloid fibers (S9A Fig) [40]. We also extracted from a gel native pocket the insoluble aggregated material of S. aureus V329 strain and stained it with ProteoStat. As shown in S9B Fig, the dye stained the protein aggregates formed by S. aureus V329 but not the Δbap strain.
Also, we tested the effect of the compound (-)-epigallocatechine gallate (EGCG), known to exert an anti-amyloidogenic effect in the case of proteins involved in neurodegenerative diseases [52], on biofilm development by S. aureus V329 wild type strain. As a control, we used S. aureus 15981 strain that forms a PNAG-dependent biofilm. ECGC significantly disassembled biofilm formed by S. aureus V329, but not by 15981, at all concentrations tested (P<0.001, n = 5) (S9C Fig). These results ratify the amyloidogenic nature of Bap assemblies and their role in biofilm formation.
In order to investigate the biological relevance of Bap-dependent biofilm formation related to the amyloidogenic properties of domain B, we have analyzed the capacity of S. aureus V329 strain grown in LB-glu and LB media to adhere to bovine mammary epithelial (MAC-T) cell line. The results revealed that V329 strain adhered more efficiently (P<0.01) to epithelial cells when bacteria were grown in LB-glu compared to LB (Fig 7A). Accordingly, the corresponding S. aureus Δbap mutant strain showed similar capacity to adhere to epithelial cells when grown in LB and LB-glu media. These results suggest that fibers formation would improve S. aureus adhesion to host cells.
Then, we performed an experiment to evaluate the colonization ability of S. aureus V329 wild type and Δbap mutant using a mouse foreign body infection model. The reasoning is that S. aureus V329 wild type should have higher capacity to colonize and persist on catheters than the Δbap strain in the case that Bap mediates biofilm development in this specific environment. To test this hypothesis, sterile catheters were implanted and inoculated into the mice with 107 CFU of S. aureus V329 and Δbap strains grown overnight in LB-glu at 37°C. Enumeration of S. aureus cells attached to the catheters 4 days after infection showed slight but not significant differences between S. aureus V329 wild type and the Δbap mutant strains (Fig 7B). However, at 10 days post-infection, the number of recovered bacteria was significantly higher for the wild type strain (CFU 5.8 x 105) compared to the bap mutant (P<0.05) (Fig 7B). These results suggest that Bap-mediated biofilm is important for the persistence of S. aureus through an infection process and, since in S. aureus V329 biofilm development depends on Bap amyloid fibers, this would imply a key role of these structures in the colonization of indwelling medical devices in vivo. However, further studies using different mutant strains that express Bap proteins incapable of aggregate (mutated in the major amyloid sequence stretches required for fibrillation, or mutated in its N-terminal domain) are required.
Bap-mediated multicellular behavior is inhibited in the presence of millimolar concentrations of calcium bound to the EF-hand domains present in the region B of Bap [36]. The question arises as how calcium and pH environmental signals reconcile to regulate Bap-mediated biofilm formation. To address this question, we investigated the aggregation kinetics of Bap when S. aureus V329 and ∆bap producing Bap_B were grown in LB-glu supplemented with 20 mM of CaCl2. The presence of calcium inhibited aggregation of the Bap protein (S10A and S10B Fig, left panels), as well as biofilm formation and bacterial clumping (S10C Fig) despite acidification of the growth media. On the other hand, the wild type V329 and the Δbap Bap_B strains mutated in their EF-hand 2–3 calcium binding motifs (ΔEF and Bap_B_ΔEF respectively) showed no disruption of either biofilm phenotype (S10C Fig) or protein aggregation (S10A and S10B Fig, right panels) in the presence of calcium. Moreover, S. aureus V329 strain was grown in LB-glu and, after an overnight incubation the medium was replaced by LB-glu containing 20 mM CaCl2 to evaluate the possible disassembly of formed biofilm. No disaggregation was observed after the medium was exchanged suggesting that Ca2+ inhibitory effect on Bap functionality might be relevant in steps prior to amyloid self-polymerization process (S2D Fig).
Next, we evaluated the effect of Ca2+ on bacterial aggregation of Δbap strain induced by rBap_B under acidic culture conditions. We observed that interbacterial interactions did not occur when Δbap mutant strain exogenously complemented with rBap_B was incubated in the presence of calcium (Fig 8A). We also tested the effect of calcium on Bap amyloid formation by analyzing in vitro aggregation of rBap_B into ThT positive amyloid-like structures in the presence of the cation. Results showed that calcium significantly inhibited the formation of amyloid-like aggregates (Fig 8B).
To determine whether the inhibitory effect of Bap-amyloid aggregation induced by calcium is due to a change in protein structure upon binding to the cation we used several biophysical approaches. For this, it is important to clarify that because fast aggregation of rBap_B at pH 4.5 even at low protein concentration (0.01 mg/ml) makes difficult the characterization of the conformational properties of the soluble monomers at this pH, we analyzed the biophysical properties of the Bap_B domain in the presence of calcium at neutral pH. First, 1 and 10 mM CaCl2 are sufficient to induce a concentration dependent increase in the ellipticity of the far-UV CD spectra of rBap_B (Fig 8C). Deconvolution of the spectra in the absence and in the presence of 1 mM Ca2+ indicated that the protein displays very similar secondary structure content in these conditions. The spectrum is dominated in both cases by disordered conformations, although a small reduction in the overall β-sheet content could be observed in the presence of the Ca2+ (the 10 mM CaCl2 spectrum could not be deconvoluted due to the strong increase in HT voltage below 200 nm in this condition). We next decided to monitor by near-UV CD and intrinsic fluorescence the overall tertiary structure of rBap_B in the presence or absence of Ca2+. Despite no significant impact of Ca2+ on the environment of rBap_B aromatic residues could be observed by near-UV CD (S11B Fig), the cation promotes a detectable increase in intrinsic fluorescence emission (S11C Fig), suggesting rearrangements in the tertiary context of the protein. To confirm the existence of a change in the aromatic residues environment, we performed thermal denaturation in the absence and in the presence of 1 and 10 mM CaCl2. Results indicated that the temperature at which the protein loses half of its intrinsic fluorescence, augmented in the presence of increasing concentrations of calcium (1 and 10 mM), suggesting that the cation exerts a global stabilizing effect on Bap conformation (Fig 8D). Further techniques support this idea. Analytical ultracentrifugation analysis revealed that, in the presence of calcium the rBap_B monomer exhibited a significantly higher sedimentation coefficient (s(20,w)∼3.4 versus ∼3.0). Additionally, rBap_B protein showed a frictional ratio f/f0 = 1.39, compatible with a slightly elongated protein, while in the absence of the cation the protein showed a frictional ration of 1.66 indicating a more elongated and moderately asymmetric protein shape (Figs 8E and S12A). Size exclusion chromatography analysis supported this by revealing that rBap_B is eluted with retard in the presence of calcium (S12C Fig). Additionally, when we performed 1D-NMR of rBap_B in the presence or absence of Ca2+, we observed an increase in the number of peaks corresponding to the methyl (0.5 ppm) and amide (8.5 ppm) regions of the spectrum (S12D Fig). The broader line-widths observed in these regions in the absence of Ca2+ are in concordance with a molten globule that is semi stable and fluctuates between several conformations. The sharpening of the peaks when Ca2+ is present would then be indicative of protein ordering into a more stable state with a smaller hydrodynamic radius (S12D Fig). Finally, analysis of Bap accessibility to proteolytic degradation in the presence of calcium showed that rBap_B was readily hydrolyzed by proteinase K in the absence of calcium, whereas it was protected from proteinase K activity when the cation was present (Fig 8F). Together all these results are consistent with the idea that Bap protein adopts a transient molten globule-like state in the absence of calcium prior to amyloid formation that is stabilized upon calcium binding thus impeding amyloid assembly due to tertiary rearrangements of Bap conformation.
Although orthologs of Bap exist in many coagulase-negative staphylococci, homology in region B is variable [35] (S13 Fig and S5 Table). Thus, we wondered whether Bap orthologs could also mediate multicellular behaviour by the generation of amyloid-like aggregates. We selected Bap_B of S. saprophyticus (Bap_Bsapro) as it shares an intermediate percentage of identity with Bap_B of S. aureus (58% identity over the entire length of the B domain). Expression of a chimeric protein containing Bap_Bsapro linked to R-ClfA in S. aureus ∆bap∆spa (S5 Fig), induced bacterial clumping under acidic culture conditions, but not under basic conditions (Fig 9A and 9B). Consistent with the presence of EF-hand domains, Bap_Bsapro was sensitive to the presence of calcium in the media (Fig 9B). As previously shown for rBap_B, purified rBap_Bsapro formed precipitated protein aggregates in acidic phosphate-citrate buffer (pH 4.5) that reversibly disassembled after raising the pH to neutral (S6B Fig). Together these results indicate that Bap_B domain of S. saprophyticus mediates multicellular behavior under acidic culture conditions, analogous to Bap_B domain of S. aureus.
Biophysical characterization of the rings formed by rBap_Bsapro at pH 4.5 indicated that they possess clear amyloid-like features, displaying strong binding to ThT (Fig 9C), ProteoStat (Fig 9D) and CR (Fig 9E). As for rBap_B, light scattering and bis-ANS binding assays demonstrated that rBap_Bsapro self-assembled rapidly into aggregates displaying exposed hydrophobic clusters at pH 4.5, whereas it remained soluble at pH 7.0 (S14 Fig). Indeed, when fresh rBap_Bsapro was incubated at pH 4.5 for 24 h, the presence of fibrillar structures became apparent (Fig 9F). Finally, we extended the analysis of the amyloid-forming propensity to Bap_B domains of S. simiae, S. xylosus, S. epidermidis and S. simulans (S5 Table). For that, we used the curli-dependent amyloid generator (C-DAG) system that provides a simple cell-based method to test particular target proteins for their amyloid-forming propensity [53]. The presence of extracellular amyloid aggregates was detected by analyzing the capacity of the strains to bind Congo Red dye (CR). Interestingly, all the Bap_B domain orthologs expressed in C-DAG system were able to bind CR, whereas Bap_A domain of S. aureus did not (Fig 9G). Together, these data indicate that Bap orthologs also utilize amyloid assembly as a molecular mechanism to induce multicellular behaviour.
There is a growing recognition that proteins play an important role building biofilm matrix scaffold. To fulfill this function these proteins need to provide stable intercellular connections and at least in some cases, also mediate adhesion to the surface. In this report, we have shown that Bap forms extracellular amyloid-like fibers that assist in building the biofilm matrix in S. aureus. Bap shares structural and functional properties with SasG and Aap proteins of S. aureus and S. epidermidis respectively, implicated in cell-to-cell accumulation and adhesion to epithelial cells [38,54–56]. However, the mechanims of action of SasG and Aap are completely different to the one reported here for Bap. All three proteins undergo a limited proteolytic cleavage of the N-terminal domain that induces biofilm formation [15,57]. The mechanism underlying this processing is different among the three proteins. SasG is known to suffer spontaneous cleavage at labile bonds in its B domain, since protease inhibitors added to the growth medium, as well as strains deficient in each known extracellular and membrane-bound proteases, had no effect on the pattern of SasG processing [15]. In the case of Aap, endogenous and also exogenous host-derived proteases are the responsible for protein cleavage, and addition of α2-macroglobulin to the growth medium specifically led to the loss of cell clumping and biofilm formation of S. epidermidis [57]. We failed to identify a staphylococcal protease responsible for Bap cleavage, because protease mutants (∆aur, ∆sspA and ∆sspBC) and protease inhibitors (α2-macroglobulin, E64, ScpB and PMSF) did not change the proteolytic pattern of Bap (S4 Fig). However, the possibility that a protease different from the ones tested cannot be discarded and requires further study. In the case of SasG and Aap, the N-terminal domain is removed by proteolysis allowing the C-terminal region containing the G5 domains to promote zinc-dependent self-association of opposing molecules [6,15,57,58]. In contrast, it is the N-terminal region of Bap that is released to the extracellular media and self-assembles into amyloid-like fibers, whilst part of the C-terminal repeats region remains anchored to the membrane. Several pieces of experimental evidence support the amyloid-like properties of the Bap_B domain aggregates. First, far-UV CD spectra reveal a modest switch in secondary structure of Bap from disordered to β-sheet, as the pH becomes more acid. Second, FTIR spectrum of Bap aggregates is dominated by β-sheet/β-turn secondary structure. Third, rBap_B binds to the amyloid diagnostic dyes Thioflavin-T, Congo Red and Proteostat and forms aggregates with fibrillar morphology when observed by electron microscopy. Finally, rBap_B contains short-sequence stretches with significant amyloidogenic potency that together with other unknown sequence stretches could contribute to the fibrillogenesis of Bap fragments. The self-assembly of rBap_B at acidic pH is a fast process, where hydrophobic interactions appear to play an important role, at least at the early stages of the reaction. Genuine bacterial functional amyloids utilize sophisticated machineries that direct the polymerization of amyloid fibers outside the cell. For instance, curli (csgACB-csgDEFG), Fap (fapA-F) in Pseudomonas strain UK4, chaplins (chpA-H) in Streptomyces coelicolor and TasA (tapA-sipW-tasA) in Bacillus are expressed together with accessory proteins involved in secretion, nucleation, and assembly of the amyloid subunit [29,30,59–61]. Bap appears to follow a more simplistic model of amyloid auto-aggregation, which does not require the expression of accessory proteins. In this respect, amyloidogenic behavior of Bap could be similar to the mechanism conducted by the surface protein antigen I/II (adhesin P1) of Streptococcus mutans [62,63].
What are the underlying reasons for the conversion of a cell wall anchored protein like Bap into an amyloid fiber? Our results suggest that this strategy allows Bap to play a dual role during biofilm development (Fig 10). Initially, Bap is secreted and covalently anchored to the cell wall. Then, Bap is processed or non-enzymatically cleaved releasing fragments containing the N-terminal region to the media that remain soluble at neutral pH. If the pH of the environment decreases, the N-terminal domain of Bap would transition from its partially ordered native state to an aggregation-prone conformation that would facilitate polymerization into amyloidogenic fibrillar structures. The presence of calcium drastically influences the multicellular behavior promoted by Bap. From a biophysical perspective, the binding of calcium probably to the EF-hand domains of the protein, stabilizes its initial molten globule-like state, likely sequestering the functional N-terminal fragments released from Bap cleavage, and consequently impairing their self-assembly into amyloid structures (Fig 10).
In eukaryotes, there are several examples of proteins involved in aggregation disorders, whose capacity to form multimeric aggregates depend on changes in protein folding caused by binding of metal ions [64]. This is the case of S100A6, an amyloid protein largely expressed in patients with Amyotrophic Lateral Sclerosis (ALS) disease. When S100A6 binds calcium, it suffers a remodeling of the surface electrostatics and hydrophobic patch exposure at the aggregation hotspot inhibiting protein self-assembly into amyloid fibrils [65]. In the case of bacteria, Ca2+ bound to α-haemolysin secreted by pathogenic E. coli, makes the protein more compact, stabilizing its structure and making it less prone to oligomerization [66]. In a similar way, the binding of Ca2+ to Bap causes tertiary rearrangements that increase the stability of the intermediate molten globule-like state of Bap in solution and thus decrease its aggregation behavior. This ultimately prevents cellular interactions and biofilm formation in S. aureus.
Amyloid structures are especially well suited for assembling the biofilm matrix scaffold, as polymerization can occur in the extracellular media in the absence of energy. Furthermore, the amyloid structure provides high stability and inherent resistance against protease digestion and denaturation [28,67]. The pH at which the B domain of Bap shows aggregation activity (pH∼5, early stationary phase) is very close to the isoelectric point (pI∼4.61), where lack of a net charge facilitates interactions between protein molecules, making protein self-assembly more likely. Indeed, a large number of globular and non-globular proteins, including the pathogenic amyloid β peptide and α-synuclein have been shown to display maximum amyloid propensity when they approach their pI, indicating that the solubility of a polypeptide chain is a major factor that determines its conversion to the amyloid state [68]. Because Bap assembly can be reversed when pH is restored to neutrality, it is not difficult to imagine that Bap is able to withstand pH fluctuations, adapting its function by switching from aggregated to soluble forms (and vice versa). The ability to fluctuate between soluble and amyloid-like states has been shown to underlie key physiological processes like processing bodies and stress granules formation [69], the cellular response to DNA breakage [70] or the integrity of the cytoskeleton [71]. In S. aureus this mechanism may have a relevant physiological effect during infectious processes, in which local acidosis usually arises from the accumulation of acidic products as a result of an inflammatory response [72], and bacterial metabolism. Also, this pH-driven Bap-mediated bacterial aggregation mechanism would be physiologically significant for those Staphylococcus species expressing Bap homologous proteins that are capable of colonize human host niches displaying mildly acidic conditions (e.g., skin, anterior nares, vagina, urinary tract and mouth) [73,74], as in the case of S. epidermidis (skin, vagina during prepubertal phase), S. saprophyticus (urinary tract and vagina) and S. warneri (skin, nasal cavity, urinary tract).
In a similar way, Foulston et al [39] demonstrated that many cytoplasmic proteins reversibly associated with the cell surface in response to pH, acting as a biofilm scaffold matrix in S. aureus. Regarding calcium ion, its effect on Bap functionality might serve to explain how changes in Ca2+ levels during the stages of the lactation cycle affect intramammary infections caused by S. aureus. Bap displays low binding affinity to calcium, thus medium-to-high millimolar concentrations of the cation are required to saturate Ca2+-binding sites in the protein. Normally, the concentration of free Ca2+ in mammalian blood is strictly maintained between 1.1–1.3 mM [75,76]}. However, the total Ca2+ concentration in milk is higher, around 1.2 mg/liter (~30 mM), being one third of this total amount free in serum (~11 mM) [77]. Thus, Ca2+ levels present in the milk during the lactation period are sufficient to inhibit Bap-mediated biofilm development. On the contrary, the low Ca2+ concentration conditions that occur in the udder during the dry period allow the formation of Bap-mediated biofilms and the establishment of long-term persistent infections on the mammary gland epithelium [36].
One question that remains open from this study is how Bap amyloid fibers interact with the bacterial surface to induce cell-to-cell aggregation? In S. coelicolor, it has been proposed that covalently linked cell wall chaplin variants ChpA–C contribute to anchoring the fibers to the cell surface [59]. Following the same reasoning, one would expect that Bap amyloid fibers might interact with the C-terminal domain of Bap that remains covalently anchored to the cell wall. However, the finding that extracellular addition of rBap_B to bap deficient S. aureus strains (Fig 4B) and also to L. monocytogenes and E. faecalis (S15 Fig) promoted intercellular adhesion and biofilm formation makes this possibility very unlikely.
Finally, our results indicated that Bap orthologs share similar molecular mechanisms as Bap for mediating biofilm development. S. saprophyticus is a notable human uropathogen [78]. The pH of the urinary tract varies between 4.5 and 7 [74], representing an environment in which Bap, through the formation of amyloid aggregates and together with urease, UafA [78] and other virulence factors [79] could play an important role in the survival and uropathogenesis of S. saprophyticus. Except for S. saprophyticus and S. simiae, the rest of CNS strains used in this study were mostly isolated from mastitis of lactating dairy cows (cultured from milk samples), a physiological situation in which the presence of calcium, as previously explained, can actually play a relevant role in regulating the functionality of Bap. The diversity of yet unknown factors that could affect Bap amyloid behaviour among different bacterial species is worthy of further study.
The bacterial strains and plasmids used in this study are listed in S1 Table. Oligonucleotides were synthesized by StabVida (Caparica—Portugal) (S2 Table). Enzymes for DNA manipulation were supplied by Thermo Scientific and were used according to manufacturer’s recommendations. Staphylococcal strains and E. coli and were grown in Luria-Bertani (LB) broth or in LB agar (Pronadisa). Media were supplemented when appropriate with 10 μl/ml erythromycin, 100 μg/ml ampicillin, 0.25% wt/vol glucose, 1 μM CdCl2, 15 mM CaCl2 and 1.25 mM EDTA. Plasmid DNA was isolated from E. coli strains using a Qiagen plasmid miniprep kit (BioRad), according to the manufacturer’s protocol. Plasmids were transformed into staphylococci by electroporation, using a previously described procedure [21]. Deletion mutants were generated via allelic replacement using the vector pMAD as described previously [80].
The signal peptide (SP) and the different N-terminal domains of the bap gene (AB, B, B∆EF and A) were amplified from S. aureus V329 and V329 ∆EF [36]. To amplify Bap_AB fragment we used primers Bapori-1mB and Bap-63cK (S2 Table). To obtain Bap_B and Bap_B ∆EF regions we first amplified the signal peptide sequence using primers Bapori-1mB and Bap-65c and second, the region B with primers Bap-66m and Bap-63cK. An overlapping PCR was performed with primers Bapori-1mB and Bap-63cK to get a single fragment. To obtain Bap_A region, two fragments were amplified using primers Bapori-1mB and BapB1 (comprising signal peptide sequence), and BapB2 and BapB3K (comprising A domain). A second overlapping PCR was performed with primers Bapori-1mB and BapB3K in order to obtain a single fragment. To obtain Bap_B region of S. saprophyticus, we first amplified the signal peptide of Bap from S. aureus V329 using primers Bapori-1mB and SPbap-sapro-Rv and second, the B-region from S. saprophyticus B20080011225 using primers bapB-sapro-Fw and Sapbap-KpnI-Rv. An overlapping PCR was performed with primers Bapori-1mB and Sapbap-KpnI-Rv to obtain a single fragment. The entire ClfA fragment used as a control for Bap chimeras, was developed by amplifying clfA gene from S. aureus Newman using primers ClfA-9mB and ClfA-7cE. To allow anchoring of amplified bap domains to the bacterial cell wall, the R region of clumping factor A gene containing an LPXTG motif was amplified from S. aureus Newman strain using primer K-3xF-ClfA containing a flag tag and a recognition sequence for KpnI, and primer ClfA-7cE with a recognition sequence for EcoRI. The KpnI/EcoRI-restricted R-clfA was ligated with KpnI/EcoRI-restricted pCN51 vector [81]. The resulting construct was then digested with BamHI and KpnI to insert the previously amplified domains of bap gene. The final pCN51 plasmid constructs thus contained different parts of bap gene fused to a flag tag followed by the C-terminal R domain of clumping factor A gene, expressed under the activity of a cadmium inducible promoter.
To obtain E. coli strains for curli-dependent amyloid generator (C-DAG) system, we PCR amplified from purified genomic DNA (i) region A of bap from S. aureus (primers CDAG BAP_A-Fw and CDAG BAP_A-Rv) and (ii) region B of bap from several staphylococcal species: S. aureus (primers cdag-B-NotI-Fw and cdag-B-XhoI-Rv, S2 Table), S. saprophyticus (primers BAPsapro-cdag-Fw and BAPsapro-cdag-Rv), S. simiae (primers BAPsimiae-cdag-Fw and BAPsimiae-cdag-Rv), S. epidermidis (primers epider-CDAG-Fw and epider-CDAG-Rv), S. simulans (primers simulans-CDAG-Fw and simulans-CDAG-Rv), and S. xylosus (primers xylosus-CDAG-Fw and xyosus-CDAG-Rv). The NotI/XhoI-restricted bapA and bapB fragments were ligated with NotI/XhoI-restricted pEXPORTXhoI plasmid (S1 Table). This vector was obtained by replacing XbaI recognition sequence of the original pEXPORT plasmid [53] for that of XhoI using QuikChange II XL Site-Directed Mutagenesis Kit (Agilent Technologies) and primers pVS72-XhoI-5 and pVS72-XhoI-3. The final pEXPORT constructs were transformed in E. coli VS39 strain. Induction of protein production and presence of amyloid-like material was assessed on solid medium containing 10 μg/ml Congo Red by evaluating colony-color phenotype, as previously described [53].
To generate the deletion in the aur, sspA, sspB genes coding for proteases present in S. aureus, and a deletion in the spa gene coding for surface protein A, we used the pJP437, pJP438, pJP439 [82] and pMADspaAD [20] plasmids which contained two fused fragments of 500 bp each that flanked the left and the right sequence of aur, sspA, sspBC and spa genes, respectively. Plasmids were transformed in V329 or Δbap strains by electroporation. Homologous recombination experiments were performed as described [80]. V329 Δaur, ΔsspA and ΔsspBC (this last strain was also deleted in the cysteine protease inhibitor SspC, which is the last gene of the operon that codifies for SspA and SspB) strains were verified using primers ssp-20cN/ssp-17mS, ssp-24cN/ssp-21mS and aur-Fw/aur-Rv (S2 Table). V329 Δspa and ΔbapΔspa strains were verified using primers spaF/spaE (S2 Table).
Biofilm formation assay in microtiter wells was performed as described [83]. Briefly, strains were grown overnight at 37°C and then diluted 1:40 in the corresponding media supplemented when required with antibiotics, 20 mM CaCl2, or proteases inhibitors (2 U/ml α-macroglobulin, 2 mM cysteine protease inhibitor E64, 10 μM PMSF and 250 nM ScpB). Cell suspension was used to inoculate sterile 96-well polystyrene microtiter plates (Thermo Scientific). After 24 hours of incubation at 37°C wells were gently rinsed two times with water, dried and stained with 0.1% of crystal violet for a few minutes. When desired, crystal violet adhered at the bottom of the wells was resuspended with 200 μl of a solution of ethanol:acetone (80:20 vol/vol) and quantified using a Multiskan EX microplate photometer (Thermo Scientific) with a 595 nm filter.
For biofilm disassembly assays, cells were grown in LB-glu at 37°C on polystyrene microtiter plates. Once formed, adhered biofilm were treated with dispersant agents (0.4 μg/ml Dispersin B, 0.4 μg/ml DNase I, and 20, 100 and 200 μM EGCG) for 2 hours at 37°C. Alternatively, old LB-glu media were extracted and replaced for new LB, LB-glu, and LB-glu + 20 mM CaCl2, and incubated overnight at 37°C. Finally, treated and non-treated biofilms adhered to polystyrene wells were macroscopically determined and quantified as previously described.
Biofilm formation under flow conditions was analyzed using microfermenters (Pasteur Institute’s laboratory of Fermentation) with a continuos flow 40 ml/h of LB-glu and constant aeration with sterile pressed air (0.3 bar) [84]. Medium was supplemented with 10 μg/ml erythromycin and 1 μM CdCl2 when required. Each microfermentator was inoculated with 108 bacteria from an overnight culture of the corresponding strain. Biofilm development was recorded with a Fugifilm FinePix S5800 digital camera.
Aggregation phenotype in cell suspension was determined as described before [10]. Cells were grown overnight in the corresponding medium (TSB-glu, LB-glu or LB acidified with 0,1 M HCl) at 37°C, shaking at 200 rpm and were examined macroscopically for the presence or absence of aggregates (intercellular adhesion). For bacterial clumping reversion assay, bacteria were grown overnight in LB-glu at 37°C, 200 rpm. Cultures were subsequently centrifuged and LB-glu medium was replaced for LB. After after 6 and 18 h incubation at 37°C, 200 rpm, bacterial aggregation at the bottom of the tube was evaluated for each strain and pictures were taken with a FUJIFILM FinePix S5800 digital camera. To quantify bacterial aggregation, the OD600nm at the top of the culture tubes (approximately 1 cm from the surface) was measured as an estimation of non-settled bacteria (planktonic cells) present in the culture after an overnight incubation at 37°C, 200 rpm. The experiment was independently repeated three times, and data were analyzed with the Mann-Whitney test.
For immunofluorescence, cells were grown overnight in the corresponding tested conditions and fixed with 3% paraformaldehyde (SIGMA) for 5 minutes. 200 μl of fixed bacteria were placed on coverslips and incubated for 30 minutes. After several washes with PBS, cells were saturated with PBS-0.5% BSA, and finally stained with anti-Bap or anti-Flag (Sigma) antibodies diluted 1:1000. Alexa Fluor 488-conjugated goat anti-rabbit (Invitrogen) diluted 1:200 was used as a secondary antibody and DAPI diluted 1:200 was used to label nuclei. For ProteoStat staining of amyloid material in-vivo, cells were grown overnight in LB and LB-glu, at 37°C, in polystyrene 24-wells plates. Adhered biofilm was resuspended and fixed with 3% paraformaldehyde (SIGMA) for 5 minutes. Bacteria were washed several times with 1X PBS, and then incubated for 30 minutes, at room temperature and in darkness with ProteoStat Mix buffer (1X Assay Buffer, 1 μl ProteoStat®, 2 μl Hoechst). Bacteria were washed twice with 1X PBS. All preparations were observed with an Axioskop 2 plus epifluorescence microscope (Zeiss) and images were acquired and analyzed with EZ-C1 software (Nikon). For Transmission electron microscopy (TEM), cells were grown overnight in the corresponding tested conditions, washed twice with phosphate-buffered saline (PBS) and then fixed with paraformaldehyde 2% (SIGMA) for 1 hour at room temperature. Formvar/carbon-coated nickel grids were deposited on a drop of fixed sample during five minutes and rinsed three times with phosphate-buffered saline (PBS). Negative staining was performed using 2% uranyl acetate (Agar Scientific, Stansted, UK). Observations were made with a JEOL 1011 transmission electron microscope. For Bap immunogold labelling, grids coated with the sample were washed and incubated for 45 minutes on a drop of PBS containing 1:10 antibody against BapB. After washing with PBS, grids were incubated 45 minutes with gold-conjugated (10nm) goat-anti-rabbit secondary antibody (Aurion, Wageningen, Netherlands). Grids were stained with uranyl acetate as described above.
Region B of Bap (amino acids 361–819) was PCR amplified from purified S. aureus V329 genomic DNA using high fidelity Phusion DNA Polymerase (Thermoscientific) and primers bapB1-LIC-Fw and bapB1-LIC-Rv (S2 Table) designed for use in the LIC cloning system. The resulting 1377 bp fragment was cloned in pET46-Ek/LIC vector (Novagen). B-region of Bap from S. saprophyticus B20080011225 was PCR amplified from its purified genomic DNA using primers bapB-sapro-LIC-Fw and bapB-sapro-LIC-Rv. The resulting 1311 bp fragment was cloned in pET46-Ek/LIC vector (Novagen). Both fusions resulted in Bap_B constructs containing an N-terminal hexahistidine tag (rBap_B and rBap_Bsapro). Overnight cultures of Escherichia coli BL21 DE3 containing Bap_B expression plasmid were diluted 1:100 and grown to an OD600nm of 0.6. Isopropyl B-D-thiogalactopyranoside (IPTG) was added to a final concentration of 0,1 mM and the cultures were shaken at 20°C overnight. After centrifugation, pellets were resuspended, sonicated and centrifuged. Supernatants were filtered (0,45 μm) and rBap_B protein purified by Ni affinity chromatography using HisGraviTrap gravity-flow columns (GE Healthcare). To achieve the highest purity, size exclusion chromatography was applied with a HiLoad 16/600 Superdex 200 pg column (GE Healthcare). The concentration of the purified protein was determined by the Bicinchoninic Acid (BCA) Protein Assay (Pierce, Thermo Scientific) using BSA as a standard.
Rabbit polyclonal antibodies raised against purified rBap_B protein were supplied by Abyntek Biopharma S.L. (Spain). Antibodies were subsequently immunoabsorbed and purified using NAb Spin Kit (Thermoscientific).
To test extracellular complementation, bacteria were grown in LB, LB-glu or LB-glu + 20 mM CaCl2 media mixed with 2 μM-purified rBap_B shaking at 200 rpm at 37°C. Aggregation phenotype in cell suspension was determined and quantified as described above.
To determine the exact pHs at which rBap_B is capable to form aggregates, 2 μM of the protein was incubated in phosphate-citrate buffer at pH ranging from 2.0 until 8. Protein aggregates were macroscopically determined and pictures were taken with a FUJIFILM FinePix S5800 digital camera. For aggregates reversion assay of 2 μM assembled rBap_B and rBap_Bsapro protein, the phosphate-citrate buffer at pH 4.5 was removed and exchanged for phosphate-citrate buffer at pH 7. After an overnight incubation at 37°C and 200 rpm, dissolution of rBap_B and rBap_Bsapro aggregates was macroscopically determined and pictures were taken with a FUJIFILM FinePix S5800 digital camera.
Overnight cultures of S. aureus strains were diluted 1:100 and grown in LB-glu or LB supplemented with the corresponding antibiotic, 20 mM CaCl2 and 1 μM CdCl2 when necessary, at 37° C and 200 rpm. Samples were obtained at different point of the growth curve. For protease inhibition assays, diluted S. aureus V329 cultures were supplemented with proteases inhibitors (2 U/ml α2-macroglobulin, 2 mM cysteine protease inhibitor E64, 10 μM PMSF and 250 nM ScpB) and grown until OD≈0.7. Cells were harvested, washed and finally resuspended in 100 μl of PBS buffer containing 30% raffinose (Sigma), 5 μl of lysostaphin 1 mg/ml (Sigma) and 2 μl of DNase 1mg/ml (Sigma). After 2 hours of incubation at 37° C, cells were centrifuged. The supernatants from surface protein extracts were recovered and analyzed by SDS-PAGE or Native gels. For SDS-PAGE, 1 volume of Laemmli buffer was added to the samples and boiled for 5 minutes. 10 μg of protein was used for SDS-PAGE analysis (7,5% separation gel; 5% stacking gel). For Native gels, surface protein extracts were mix 1:2 with native sample buffer (BioRad). Proteins were separated in Criterion XT Tris-acetate gels and Tris/glycine running buffer (BioRad). For Western blot analysis, protein extracts were blotted onto Hybond-ECL nitrocellulose membranes (Amersham Biosciences). Anti-Bap purified antibody and monoclonal anti-Flag M2-Peroxidase (HRP) antibody (Sigma) were diluted 1:20.000 and 1:1000, respectively, with PBS-Tween 5% skim-milk. Alkaline phosphatase-conjugated goat anti-rabbit Immunoglobulin G (Thermo Scientific) diluted 1:5000 in PBS-Tween 5% skim-milk was used as a secondary antibody for Bap detection and the subsequent chemiluminescence reaction was recorded (Chemiluminescent Substrate Thermo Scientific).
S. aureus V329 strain was grown in LB and LB-glu media, at 37°C, 200 rpm. After an overnight incubation, cells were harvested, washed and finally resuspended in 100 μl of PBS buffer containing 30% raffinose (Sigma), 5 μl of lysostaphin 1 mg/ml (Sigma) and 2 μl of DNase 1mg/ml (Sigma). After 2 hours of incubation at 37° C, cells were centrifuged. Supernatants were mix 1:2 with native sample buffer (BioRad). Proteins were separated in Criterion XT Tris-acetate gels using Tris/glycine running buffer (BioRad). The material retained in the wells of the native gels was excised, washed three times in ddH20, and digested in-gel with 250 ng of trypsin (Sequencing grade modified Trypsin-Promega) in 50 mM ammonium bicarbonate for 16 h at 37°C, after a denaturation step with DTT (10 mM, 30 min, 40°C) and an alkylation step with Iodoacetamide (25mM, 30 min, room temperature). The resulting peptides were extracted with 1% formic acid, 50% acetonitrile and evaporated to dryness prior to LC-MSMS analysis. For each digested sample, a total volume of 5 μl of tryptic peptides was injected with a flow rate of 300 nL/min in a nanoLC Ultra1D plus (Eksigent). A trap column Acclaim PepMap100 (100 μm x 2 cm; C18, 2 μm, 100 Å) and an analytical column Acclaim PepMap RSLC (75 μm x 15 cm, C18, 5 μm, 100 Å) from Thermo Scientific were used following the next gradient: 0–1 min (5% B), 1–50 min (5–40% B), 50–51 min (40–98% B), 51–55 min (98% B), 55–56 min (5% B), 56–75 min (5% B). (Buffer B: 100% acetonitrile, 0.1% formic acid, Buffer A: 0.1% formic acid). MS analysis was performed on a Q-TRAP 5500 system (ABSciex) with a NanoSpray® III ion source (ABSciex) using Rolling Collision Energy in positive mode. MS/MS data acquisition was performed using Analyst 1.5.2 (AB Sciex) and submitted to Protein Pilot software (ABSciex) against UniprotKB/Swiss-Prot database (restricted to “Staphylococcus”) and then against a specifically restricted database for BAP protein from Staphylococcus aureus, using the Paragon™ Algorithm and the pre-established search parameters for 5500 QTRAP.
Adherence experiments were performed as described previously [85]. Briefly, prior to use, wells were seeded with 0.3 x 106 MAC-T cells in 6-well tissue culture plates. Once cells were confluent (1.2 x 106) the culture medium was removed and cells were washed with DMEM plus 10% heat-inactivated fetal bovine serum. Overnight bacterial cultures were mixed vigorously and added to the monolayer cells in a multiplicity of infection of 10 in DMEM. Incubation was carried out 1 hour at 37°C in 5% CO2. To remove non-adherent bacteria, cells were washed three times with sterile PBS. Eukaryotic cells were lysed with 0.1% Triton X-100. Before plating extracts were mixed vigorously by vortexing. The number of adherent bacteria were determined by serial dilution and plating. Experiments were performed in triplicate.
A mouse foreign body infection model was used to determine the role of Bap aggregates in the pathogenesis of S. aureus. Groups of 6 CD1 mice were used. A 3-cm segment of intravenous catheter (24G1”, B. Braun) was aseptically implanted into the subcutaneous interscapular space. Each group of six mice was inoculated with 1 x 107 CFU of either S. aureus V329 or Δbap mutant previosly grown overnight in LB-glu at 37°C. Twelve animals were euthanatized by cervical dislocation on days 4 or 10 post-infection. The catheter was aseptically removed, placed in a sterile microcentrifuge tube containing 1 ml of PBS, and vortexed at high speed for 3 min. Samples were serially diluted and plated onto TSA plates for enumeration of viable staphylococci.
All animal studies were reviewed and approved by the Comité de Ética, Experimentación Animal y Bioseguridad, of the Universidad Pública de Navarra (approved protocol PI-019/12). Work was carried out at the Instituto de Agrobiotecnología building under the principles and guidelines described in European Directive 86/609/EEC for the protection of animals used for experimental purposes.
Proteolysis of rBap_B (1 mg/ml) was performed at 37°C in the presence or absence of 50 mM CaCl2. The protein was incubated with 80 μg/ml Proteinase-K (SIGMA) for 0, 15, 30, 45 minutes and the reaction was stopped by the addition of 5 mM PMSF. Degradation pattern was analyzed by SDS-PAGE (12%) followed by western immunoblotting with anti-Bap purified primary antibody (1:10.000) and alkaline phosphatase-conjugated goat anti-rabbit Immunoglobulin G (1:5000) (Thermo Scientific) as a secondary antibody.
Thioflavin-T (ThT) binding was analyzed for 0.1 mg/ml aggregated rBap_B and rBap_Bsapro in the presence of 25 μM ThT, 25°C, pH 4.5. ThT binding was also measured for rBap_B at different concentrations (0.01, 0.018, 0.027 and 0.036 mg/ml) in the presence of 25 μM ThT, 25°C, at pH 4.5 and pH 7. Fluorescence emission spectra were recorded from 460 to 600 nm with an excitation wavelength of 440 nm, using a slit width of 5 nm for excitation and emission in a Jasco FP-8200 spectrophotometer (Jasco corporation, Japan) at 25°C. Each trace represents the average of 5 accumulated spectra. Aggregation kinetics of 0.01 mg/ml rBap_B protein in phosphate-citrate buffer at pH 4.5, pH 4.5 plus CaCl2 and pH 7.0 were recorded for 1000 s under agitation (800 rpm) at 25°C, in the presence of 25 μM ThT. The kinetic traces were measured exciting at 440 nm and emission was recorded at 475 nm, slit widths of 5 nm were used for excitation and emission in a Jasco FP8200 spectrophotometer (Jasco corporation, Japan). ThT fluorescence spectra were recorded at the end of the experiment.
Congo red (CR) interaction with 0.1 mg/ml aggregated rBap_B and rBap_Bsapro at pH 4.5 was tested using a Cary-400 UV/Vis spectrophotometer at 25°C. After 5 minutes of equilibration, the absorbance spectra were recorded from 400 to 700 nm. Each trace represents the average of 5 accumulated spectra.
Fluorescence emission of 0.1 mg/ml assembled rBap_B and rBap_Bsapro stained with ProteoStat was measured on a Jasco FP-8200 fluorescence spectrophotometer (Jasco corporation, Japan) at 25°C. The emission spectra were recorded between 500 and 650 nm wavelength and the samples excited at 484 nm. Slit widths for excitation and emission spectra were 5 nm. The spectra were obtained from the average of 5 consecutive scans.
ATR FTIR spectroscopy analyses of rBap_B aggregates formed in phosphate-citrate buffer pH 4.5 were performed with a Bruker Tensor 27 FTIR Spectrometer (Bruker Optics Inc.) with a Golden Gate MKII ATR accessory. Spectrum acquisitions consisted of 16 independent scans, measured at a resolution of 2 cm-1 within the 1800–1500 cm-1 range. Spectra were acquired, background subtracted, baseline corrected and normalized using the OPUS MIR Tensor 27 software. Second derivatives of the spectra were used to determine the frequencies at which the different spectral components were located. All FTIR spectra were fitted to overlapping Gaussian curves using PeakFit package software (Systat Software) and the maximum and the area of each Gaussian were calculated.
Samples of 0.1 mg/ml rBap_B and rBap_Bsapro soluble proteins (phosphate-citrate buffer at pH 7) or protein aggregates (phosphate-citrate buffere at pH 4.5) were prepared in solutions containing 10 μM of Bis-ANS and analyzed immediately on a Jasco FP-8200 fluorescence spectrophotometer (Jasco corporation, Japan) at 25°C. The samples were excited at 370 nm and emission measured between 400 and 600 nm with slit widths of 5 nm. The spectra were obtained from the average of 5 consecutive scans.
Far-UV CD spectra were measured in a Jasco-710 (Jasco, Japan) or in a Chirascan (Applied Photophysics) spectropolarimeter thermostated at 25°C. rBap_B at concentrations ranging from 0.2 to 1.5 mg/ml was measured in 10 mM MOPS either with 1mM CaCl2, 10 mM CaCl2 or 100 mM CaCl2, or alternatively 10 mM NaCl or 100 mM NaCl and 10 mM EDTA at pH 7.0/7.5. For measurements at acidic pH, rBap_B (6 mg/ml) in 10 mM NaPO4 pH 7.0, 50 mM (NH4)2SO4 was diluted 30-fold to 0.2 mg/ml into 100 mM NaPO4, 10 mM EDTA at pH 4.4. Spectra were recorded from 260 to 190 nm, at 0.2 nm intervals, 1 nm bandwidth, and a scan speed of 50 nm/min. Twenty accumulations were averaged for each spectrum. Deconvolution of the data were performed using the Dichroweb server implementing the CDSSTR algorithm with reference set 7 [86,87]. Near-UV CD spectra were recorded in a Jasco-710 spectropolarimeter (Jasco, Japan) thermostated at 25°C, from 260 to 320 nm with a 1 nm bandwidth, and a scan speed of 50 nm/min in 10 mM MOPS pH7.0 with 1 mM CaCl2, 10 mM CaCl2, or 10 mM NaCl 10 mM EDTA.
Tryptophan intrinsic fluorescence was measured at 25°C on a Varian Cary Eclipse spectrofluorometer using an excitation wavelength of 280 nm and recording the emission from 300 to 400 nm. Five averaged spectra were acquired and slit widths were typically 5 nm for excitation and emission. Protein concentration was 1.5 mg/ml in 10 mM MOPS either with 100mM CaCl2 or 100 mM NaCl, 10 mM EDTA at pH 7.5.
Thermal denaturation was monitored in a Jasco FP-8200 fluorescence spectrophotometer (Jasco, Japan) The samples were excited at 280 nm and the emission was recorded at 350 nm, using slit widths of 5 nm for excitation and emission. The emission was registered each 0.25 K with a heating rate of 0.5 K/min.
Static light scattering of 0.1 mg/ml rBap_B and rBap_Bsapro in phosphate-citrate buffer at pH 4.5 and pH 7 was recorded on a Jasco FP-8200 spectrofluorometer (Jasco corporation, Japan). Five accumulative spectra were registered with excitation at 330 nm and emission between 320 and 340 nm. Slit widths of 5 nm for excitation and emission were used.
Dynamic light scattering data of 1 mg/ml rBap_B protein in phosphate-citrate buffer at pH 3, 4.4 and 7 were obtained with a DynaPro DLS reader (Wyatt Technology, Germany) using an 825 nm wavelength laser and analyzed with Dynamics V6 software. Hydrodynamic radium (nm), polidispersity percentage and diffusion coefficient (cm2/s) of each population observed at the different pH values were obtained.
The size exclusion chromatography experiment was performed using a HiLoad 16/600 Superdex 200 pg column (GE Healthcare) connected to an AKTAprimeTM Plus chromatography system (GE Healthcare). A 500 μl portion of rBap_B was loaded onto the gel filtration column equilibrated in MOPS buffer (10 mM MOPS, 100 mM NaCl, pH = 7,5) with 100 mM CaCl2 or 10 mM EDTA and eluted with one column volume (124 ml) at a flow rate of 1 ml/min. Recorded data were analyzed using PrimeView software (GE Healtcare).
All AUC experiments were carried out at 20°C in the presence of 100 mM of CaCl2 and 10 mM EDTA, on a Beckman XL-I analytical centrifuge using absorbance optics. Sedimentation velocity was performed for rBap_B at three different concentrations (1, 2 and 3 mg/ml) at 48,000 rpm overnight and the data were analyzed using SedFit 14.7g [88]. Sedimentation equilibrium runs were performed for rBap_B (loading concentrations of 1, 2 and 3 mg/ml) at speeds of 13,000 and 8,500 rpm and analyzed using HeteroAnalysis 1.1.44.
One-dimensional proton NMR experiments were performed at 30°C on 350 μM Bap_B in a buffer containing 10 mM MOPS, 100 mM CaCl2, 10% D2O or 10 mM MOPS, 100 mM NaCl, 10 mM EDTA, 10% D2O. Spectra were processed within TopSpin (Bruker).
The predicted peptides GIFSYS and TVGNIISNAG were obtained from CASLO ApS (Lyngby, Denmark) with high purity (99.88% and 98.29% respectively). Peptide stock solutions at 1 mM were prepared by dissolving into citrate buffer. Samples were immediately sonicated for 10 min to dissemble preformed nuclei and centrifuged (5 min at 16,100g) to deposit insoluble material. Peptide solutions were incubated at room temperature (25°C) for four weeks and amyloid properties were evaluated as described above.
The statistical analysis was performed with the GraphPad Prism 5 program. A nonparametric Mann-Whitney test was used to assess significant differences in biofilm formation and bacterial aggregation capacity, as well as for analysis of experimental infection. The differences in bacterial aggregation after exogenous complementation with rBap_B protein were determined using the unpaired Student’s t test.
|
10.1371/journal.pgen.1002797 | Intact p53-Dependent Responses in miR-34–Deficient Mice | MicroRNAs belonging to the miR-34 family have been proposed as critical modulators of the p53 pathway and potential tumor suppressors in human cancers. To formally test these hypotheses, we have generated mice carrying targeted deletion of all three members of this microRNA family. We show that complete inactivation of miR-34 function is compatible with normal development in mice. Surprisingly, p53 function appears to be intact in miR-34–deficient cells and tissues. Although loss of miR-34 expression leads to a slight increase in cellular proliferation in vitro, it does not impair p53-induced cell cycle arrest or apoptosis. Furthermore, in contrast to p53-deficient mice, miR-34–deficient animals do not display increased susceptibility to spontaneous, irradiation-induced, or c-Myc–initiated tumorigenesis. We also show that expression of members of the miR-34 family is particularly high in the testes, lungs, and brains of mice and that it is largely p53-independent in these tissues. These findings indicate that miR-34 plays a redundant function in the p53 pathway and suggest additional p53-independent functions for this family of miRNAs.
| MicroRNAs (miRNAs) are small, non-coding RNAs that broadly regulate gene expression. MicroRNA deregulation is a common feature of human cancers, and numerous miRNAs have oncogenic or tumor suppressive properties. Members of the miR-34 family (miR-34a, miR-34b, and miR-34c) have been widely speculated to be important tumor suppressors and mediators of p53 function. Despite the growing body of evidence supporting this hypothesis, previous studies on miR-34 have been done in vitro or using non-physiologic expression levels of miR-34. Here, we probe the tumor suppressive functions of the miR-34 family in vivo by generating mice carrying targeted deletion of the entire miR-34 family. Our results show that the miR-34 family is not required for tumor suppression in vivo, and they suggest p53-independent functions for this family of miRNAs. Importantly, the mice generated from this study provide a tool for the scientific community to further investigate the physiologic functions of the miR-34 family.
| The tumor-suppressor protein p53 is a master regulator of the stress response and provides a key barrier to cellular transformation and tumorigenesis [1]. Upon oncogene activation, DNA damage, and other forms of cellular stress, p53 accumulates in the nucleus where it induces or represses the transcription of a myriad of genes. Ultimately, p53 activation results in cell cycle arrest, apoptosis, or senescence, depending on the cellular context and the type of stimulus [2]. Although transcription-independent mechanisms have been reported [3], p53 mainly acts as a transcription factor for a large array of downstream effectors [4], including the proapoptotic proteins Puma, Noxa, and Bax, as well as the cell cycle inhibitor, p21 [5]–[11]. The essential tumor-suppressive function of p53 is further highlighted by the observation that this pathway is inactivated in the vast majority of human cancers [1], [12].
Several groups have recently suggested that miRNAs are also components of the p53 pathway. In particular, three highly related miRNAs—miR-34a, miR-34b, and miR-34c (Figure 1A)—are directly induced upon p53 activation in multiple cell types and have been proposed to modulate p53 function [13]–[20]. The precursors of these miRNAs are transcribed from two distinct loci: the miR-34a locus on chromosome 1p36 and the miR-34b∼c locus on chromosome 11q23. Canonical p53-binding sites are located in the promoter regions of both miR-34a and miR-34b∼c, and these miRNAs are bona fide direct transcriptional targets of p53 [13], [17], [18].
Ectopic expression of members of the miR-34 family is sufficient to induce cell cycle arrest or apoptosis, depending on the cellular context [14], [17]–[21]. Furthermore, loss-of-function studies using miR-34 antagonists have provided some evidence that this miRNA family is required for p53 function [13], [18], [22]–[24]. Many of the predicted miR-34 target genes encode for proteins that are involved in cell cycle regulation, apoptosis, and growth factor signaling. These include Cyclin E2, cMyc, MET, BCL-2, SIRT1, and members of the E2F family of transcription factors [13], [17], [17], [23], [25].
Consistent with a possible tumor-suppressor role, loss of expression of members of the miR-34 family has been reported in human cancers. Hemizygous deletion of the chromosomal region containing the miR-34a locus has been described in neuroblastomas and pancreatic cancer cell lines [14], [21]. Similarly, loss of 11q23, containing the miR-34b∼c locus, has been reported in prostate cancers [26]. Epigenetic silencing of miR-34 members has also been reported in human cancers. Promoter hyper-methylation of miR-34a is observed in non-small-cell lung cancers and melanomas [27], [28], and silencing of miR-34a and miR-34b∼c has been described in human epithelial ovarian cancers [29].
Although these observations point towards an important role for miR-34 members as critical downstream effectors of p53 and potential tumor suppressors, these hypotheses have not been formally tested using miR-34-deficient animals and cells. One notable exception is a recent elegant paper by Choi and colleagues demonstrating that miR-34-deficient MEFs are more susceptible to reprogramming [30]. However, the consequences of miR-34 loss on p53 function were not examined in detail.
Here we report the generation of mice carrying targeted deletion of all three members of the miR-34 family and systematically investigate the impact of miR-34 loss on the p53 pathway. We show that complete genetic inactivation of miR-34 does not detectably impair the p53 response in a variety of in vivo and in vitro assays. These findings highlight likely redundancies among p53's downstream effectors, show that the miR-34 family is largely dispensable for p53 function in vivo, and suggest possible p53-independent functions.
To investigate the biological functions of miR-34, we first examined the expression of this family of miRNAs under basal conditions and in response to p53 activation in vivo. Under basal conditions, miR-34a and miR-34b∼c expression is particularly intense in the testis, brain, and lung of adult mice (Figure 1B–1D). MiR-34b∼c expression seems largely restricted to these three tissues, while miR-34a is detectable, albeit at lower levels, also in a variety of other organs (Figure 1B–1D).
Consistent with previous reports indicating that miR-34a expression is under the direct control of p53 [13], [17], [18], we detected reduced levels of this miRNA in a subset of p53-deficient tissues (heart, small and large intestine, liver and kidney), but the levels of both miR-34a and miR-34b∼c remained high in the brains, testes and lungs (Figure 1B–1D) of p53−/− mice, a finding that suggests that p53-independent mechanisms determine basal miR-34 transcription in these tissues. These results were obtained using two independent techniques: quantitative real time polymerase chain reaction (qPCR) and Northern blotting. The specificity and sensitivity of these assays were validated using miR-34-deficient mice as controls (Figure 1B–1D and Figure 2D).
Exposure to ionizing radiation, which leads to p53 stabilization and transcriptional activation, resulted in substantial miR-34a induction in the spleen, thymus, small and large intestine of wild-type mice, but not in the other tissues examined (Figure S1). We also observed modest but significant miR-34c induction in the thymus, small and large intestine of irradiated mice, but not in the other tissues examined.
To investigate the physiologic functions of the miR-34 family and to determine the extent to which its induction is required for p53 function, we generated mice carrying targeted deletion of both miR-34a and miR-34b∼c loci (Figure 2A–2C). To allow temporally and spatially restricted deletion, we also generated a conditional miR-34a KO allele (miR-34afl, Figure 2A). Northern blot and qPCR analysis confirmed the loss of expression of the respective miRNAs in homozygous mutant animals (Figure 2D). Importantly, homozygous deletion of miR-34a did not lead to compensatory up-regulation of miR-34b∼c, and vice versa (Figure 2D and data not shown). MiR-34a−/− and miR-34b∼c−/− single KO mice were viable and fertile and were obtained at the expected Mendelian frequency (Figure 2E, 2F).
The sequence similarity between the three miR-34 family members (Figure 1A), which share the same “seed”, suggests that they may be functionally redundant. To examine the consequences of complete loss of miR-34 function, we crossed miR-34a−/− and miR-34b∼c−/− mice to generate compound mutant animals carrying homozygous deletion of all three family members (miR-34TKO/TKO). Complete loss of miR-34 expression in miR-34TKO/TKO animals was confirmed by Northern blot and qPCR (Figure 2D). MiR-34TKO/TKO mice of both sexes were obtained at approximately the expected Mendelian frequency (Figure 2G), did not display obvious macroscopic defects (Figure S2), and were fertile (data not shown). A full histological examination (Figure S3), complete blood cell count (Figure S4), and serum chemistry analysis (Figure S5) did not detect any statistically significant defects in adult miR-34TKO/TKO mice of both sexes. An analysis of the major myeloid and lymphoid populations of the bone marrow, spleen and thymus also did not reveal any statistically significant difference between wild-type and miR-34TKO/TKO mice (Figure S6).
Next, we sought to determine whether loss of miR-34 expression affects the p53 response in vitro. We focused on the three best-characterized p53-dependent processes: replicative senescence, response to DNA damage, and response to oncogene activation [31]–[35].
The ability to proliferate indefinitely is one of the hallmarks of cancer cells [36] and also one of the most striking consequences of p53 inactivation at the cellular level [35]. To investigate the role of miR-34 in replicative senescence, mouse primary fibroblasts (MEFs) derived from wild-type, p53−/−, and miR-34TKO/TKO embryos were serially passaged. Although we detected a remarkable induction of miR-34a and miR-34c expression in late-passage wild-type MEFs compared to early-passage MEFs (Figure 3A), miR-34-deficient MEFs became senescent with a kinetic identical to wild-type MEFs (Figure 3B). This is in stark contrast with p53-deficient MEFs, which as expected proliferated indefinitely (Figure 3B). The only significant difference we observed was a slight but reproducible increase in the proliferation rate of early passage miR-34-deficient fibroblasts compared to controls (Figure 3B, 3C).
We next examined the role of miR-34 in the response to the DNA damaging agent doxorubicin. As previously reported [37], doxorubicin treatment leads to stabilization of p53 (Figure 3D) and up-regulation of its downstream targets p21 (Cdkn1a), Mdm2, Puma and Noxa (Figure 3D–3F). Expression of members of the miR-34 family was similarly upregulated in response to p53 stabilization (Figure 3G). Although as predicted, p53-null cells failed to arrest in G1 in response to doxorubicin treatment, the response of miR-34TKO/TKO MEFs was indistinguishable from that of wild-type cells (Figure 3H–3I). Consistent with these results, doxorubicin treatment caused similar activation of p53 and of its downstream targets in wild-type and miR-34TKO/TKO MEFs (Figure 3E and 3F).
The experiments described above were performed on asynchronously growing early-passage MEFs and as such may not be sensitive enough to detect a modest effect of miR-34 loss on the S-phase checkpoint. To measure cell cycle progression more accurately, we first synchronized MEFs by serum starvation and then released the cells in complete medium containing colcemid, a mitotic spindle inhibitor. With this experimental design, upon release in complete medium, cells synchronously proceed from G1 through S phase and then accumulate at the M (4N) phase. This assay provides a more sensitive way to determine the ability of cells to transition through the S-phase and allows detection of subtle defects in the DNA damage-induced S-phase checkpoint.
Although a reproducibly larger fraction of miR-34TKO/TKO cells was able to transition through the S phase after ionizing radiation compared to wild-type MEFs (Figure 3J), we observed a similar difference in non-irradiated MEFs (Figure 3J). The most logical interpretation of these results is that miR-34-deficient MEFs, rather than being more resistant to irradiation-induced cell cycle arrest, possess a slightly faster basal proliferation or more rapid re-entry into the cell cycle following serum starvation. This interpretation is also consistent with the faster proliferation rate displayed by miR-34-deficient MEFs (Figure 3B, 3C) and with the observation by Lal and colleagues that miR-34a is involved in modulating the cellular response to growth factors [38].
We also examined the consequences of miR-34 loss in MEFs on the expression of a subset of its previously reported direct targets [17], [20], [23], [25]. We detected modest upregulation of cMyc, E2f3, Met and Sirt1 in miR-34-deficient MEFs, while Bcl2 was expressed at similar levels in wild-type and mutant cells (Figure 3K). The upregulation of Myc and E2f3 might contribute to the increased proliferation rate we have observed in miR-34 deficient MEFs.
Having established that miR-34 is not required for cell cycle arrest in response to genotoxic stress in MEFs, we next sought to determine whether this miRNA family might contribute to p53-induced apoptosis.
Thymocytes respond to ionizing radiations by rapidly undergoing apoptosis, an effect that is dependent on p53 [39]. We therefore examined the effects of DNA damage on thymocytes from wild-type, p53−/−, and miR-34TKO/TKO mice. As expected, p53−/− thymocytes were almost entirely resistant to irradiation-induced apoptosis; however, wild-type and miR-34-deficient cells were equally sensitive to DNA damage-induced apoptosis, as judged by dose-response and time-course experiments (Figure 4A, 4B).
To exclude the possibility that tissue culture conditions may have masked a physiologic role of miR-34 in modulating the p53 response, we next examined the consequences of p53 activation in miR-34-deficient tissues directly in vivo. Age- and sex-matched wild-type, miR-34TKO/TKO and p53−/− mice were exposed to 10 Gy of ionizing radiation and euthanized 6 hours later. Ionizing radiation induced similar activation of the p53 pathway and of its downstream effectors in wild-type and miR-34TKO/TKO mice (Figure 4C). Analogous to what we observed in thymocytes in vitro, the apoptotic response was equally dramatic in wild-type and in miR-34-deficient mice, while it was virtually absent in p53−/− animals (Figure 4D–4G).
Based on these results we conclude that miR-34 function is not required for p53-induced cell-cycle arrest and apoptosis in response to genotoxic stresses.
The p53 pathway provides a crucial barrier against the neoplastic transformation of primary cells [40]. Supra-physiologic proliferative stimuli, such as those caused by sustained oncogene activation, lead to transcriptional activation of p19Arf, which in turn results in stabilization and activation of p53, and consequently apoptosis or cell cycle arrest [41]. For example, ectopic expression of a constitutively active K-Ras (K-RasV12) in wild-type MEFs leads to oncogene-induced senescence, but the concomitant inactivation of p53 is sufficient to allow full cellular transformation [31]. To test whether miR-34 plays a role in this context, we ectopically expressed oncogenic K-Ras in wild-type, miR-34TKO/TKO, and p53−/− MEFs. As shown in Figure 5A, complete loss of miR-34 function was not sufficient to allow primary MEFs to be transformed by K-RasV12 alone, while p53-deficient MEFs were readily transformed in the same assay. However, when MEFs were co-transduced with oncogenic K-Ras and E1A, which binds to and inhibits the retinoblastoma protein (pRb) [42], we observed a slight increase in the number of foci formed in miR-34TKO/TKO MEFs compared to wild-type cells (Figure 5A, 5B). These results show that while miR-34 alone is not required for p53-mediated tumor suppression in MEFs, its loss might cooperate with inactivation of the Rb pathway in promoting cellular transformation.
To extend our analysis to an in vivo setting, we next examined whether miR-34 inactivation is sufficient to accelerate spontaneous and oncogene-induced transformation in mice. P53-deficient mice exhibit a high incidence of spontaneous tumors, in particular lymphomas and sarcomas [43]–[45], and p53 inactivation greatly accelerates tumor formation in a variety of mouse models of human cancer [46]–[51]. To determine whether loss of miR-34 expression leads to increased spontaneous tumorigenesis, we aged a cohort of 14 miR-34TKO/TKO and 12 wild-type mice. The animals were monitored for at least 12 months (wild-type = 359 days; miR-34TKO/TKO = 359 days) and up to 17.3 months (wild-type = 521 days; miR-34TKO/TKO = 521 days). All wild-type and miR-34TKO/TKO mice appeared healthy and miR-34TKO/TKO mice did not show a reduction in life span compared to wild-type controls (Figure S7). For comparison, the median survival of p53−/− mice has been reported to be 4.5 months and by 10 months of age all p53−/− mice have died or developed tumors [45]. In addition, ∼40% of p53+/− mice develop tumors by 16 months of age [45]. Thus, although a longer follow-up of miR-34TKO/TKO mice may be needed to uncover very subtle defects in tumor suppression, we conclude that loss of miR-34 expression does not lead to a substantial increase in spontaneous tumorigenesis.
We next sought to determine whether loss of miR-34 might accelerate tumor formation in response to genotoxic stress. P53−/− mice irradiated shortly after birth display accelerated tumorigenesis compared to non-irradiated littermates [52]. We therefore exposed a cohort of 14 miR-34TKO/TKO and 11 wild-type mice to 1 Gy of ionizing radiation soon after birth and monitored them for 42–60 weeks. Both wild-type and miR-34-deficient mice appeared healthy throughout the follow-up period (Figure S7), in striking contrast with the ∼15 weeks reported median tumor-free survival of irradiated p53−/− mice [52]. Although it will be important to follow a larger cohort of animals over a more prolonged period, these results suggest that miR-34 does not provide a potent barrier to tumorigenesis in response to genotoxic stress in vivo.
Finally, we sought to determine whether genetic ablation of miR-34 could contribute to tumor formation in cooperation with a defined oncogenic lesion. For these experiments, we chose the Eμ-Myc model of B cell lymphomas [53]. A crucial tumor-suppressive role for p53 is well established in this mouse model and inactivation of the p53 pathway results in greatly accelerated lymphomagenesis [46], [47], [54]. However, even in this context complete loss of miR-34 expression was not sufficient to accelerate tumor formation. The incidence and latency of B cell lymphomas was virtually identical in Eμ-Myc;miR-34TKO/TKO and Eμ-Myc;miR-34+/+ mice (Figure 5C) and the resulting tumors displayed similar histopathological features and extent of spontaneous apoptosis (Figure 5D–5E).
We have reported the generation of mice carrying targeted deletion of miR-34a, miR-34b and miR-34c, and we have investigated the consequences of loss of miR-34 expression on p53-dependent responses in vitro and in vivo. Our results show that complete loss of miR-34 expression is compatible with normal development and that the p53 pathway is apparently intact in miR-34-deficient mice.
Our observation that inactivation of miR-34 does not impair p53-mediated responses in vitro and in vivo is particularly relevant because a key role for miR-34 in the p53 pathway had been previously proposed by a number of independent groups. The results presented in this paper do not necessarily conflict with previous experiments showing that ectopic expression of miR-34 can induce many of the most characteristic consequences of p53 activation; here we have tested whether miR-34 is necessary for p53 function and not whether it is sufficient.
More difficult, however, is to reconcile our findings with previous reports of impaired p53-function in cells treated with miR-34 antagonists. Because previous work has relied on the use of miRNA antagonists to inhibit miR-34 function, it is possible that some of the previous observations reflected miR-34-independent off-target effects.
It is also possible that other miRNAs sharing sequence similarities with miR-34 may compensate for miR-34 loss in the knock-out animals. In particular, members of the miR-449 family (miR-449a, b and c) have the same “seed” sequence as miR-34, and miR-34 antagonists could in principle impair their function as well. A conclusive test for this hypothesis will require the generation of compound miR-34 and miR-449 mutant animals, but several lines of evidence suggest that this explanation is not particularly likely. First, in the tissues and cells used in our experiments, the expression of miR-449 members is much lower compared to miR-34a and miR-34c, as judged by multiple independent methods including qPCR, Northern blotting and high throughput sequencing (Figure S8 and data not shown). A notable exception is represented by the testis, in which expression of miR-449a is particularly elevated (Figure S8). In addition, miR-449 expression is not substantially increased in miR-34-null mice, and activation of the p53 pathway does not lead to significant upregulation of miR-449 (Figure S8).
We would like to emphasize that our results do not necessarily indicate that members of the mIR-34 family are not components of the p53 pathway. Given the essential tumor-suppressive function exerted by p53, it is perhaps not surprising that multiple and partially redundant effector arms are recruited in response to its activation. It is plausible that the simultaneous inactivation of multiple effector arms is required to measurably impair p53 function. Consistent with this model is our observation that while loss of miR-34 expression alone does not allow the transformation of primary cells by oncogenic K-Ras, it slightly increases the efficiency of transformation when combined with inactivation of the Rb pathway by E1A (Figure 5A, 5B). In this context, it will be important to systematically probe the extent of functional cooperation between this family of miRNAs and other, previously characterized p53 effectors.
We also wish to point out that in this manuscript we have investigated the best-characterized functions of p53 (cell cycle arrest, apoptosis and tumor suppression) and it remains possible that miR-34 participates in other p53-dependent processes. For example, p53 has been proposed to modulate autophagy [55] and stem cell quiescence [56], [57] and we cannot exclude that miR-34 plays an important role in these contexts. Future studies using the miR-34-deficient animals we have generated will be needed to test these possibilities.
With respect to the potential tumor suppressive role of miR-34, our experiments indicate that loss of miR-34 expression does not lead to an obvious increase in tumor incidence in mice and does not cooperate with Myc in the context of B cell lymphomagenesis. However, the tumor suppressive function of miR-34 might be restricted to specific tissues and loss of miR-34 might cooperate with specific oncogenic lesions. In humans, for example, loss of miR-34 expression has been reported in a large fraction of primary melanomas, prostatic adenocarcinomas and small cell lung cancers [27], [28], among others. Introducing the miR-34-null alleles we have generated into mouse models of these types of human cancers will be important to fully explore the tumor suppressive potential of this family of miRNAs.
An additional issue raised by the results presented in this manuscript relates to possible p53-independent functions of miR-34. We show that under basal conditions the expression of both miR-34 loci is particularly elevated in the testes and, to a lesser extent, in the brains and lungs of mice. Importantly, in these three tissues, miR-34 expression is almost entirely p53-independent (Figure 1B–1D and [58]), a finding that suggests that additional transcription factors control the expression of this family of miRNAs in the absence of genotoxic or oncogenic stresses.
A role for miR-34c in spermatogenesis and in controlling the first zygotic cleavage has been recently proposed [58], [59]. Although our observation that single KO and miR-34TKO/TKO mice produce viable offspring argues against an essential role for miR-34 in these processes, members of the related miR-449 family, that are particularly highly expressed in the testis (Figure S8), could partially compensate for miR-34 loss in this context.
Recent reports have also implicated miR-34 in neuronal development and behavior [60], [61] and a role for miR-34c in learning and memory [62], as well as in stress-induced anxiety [63], has been reported. In addition, inactivation of miR-34 expression has been recently shown to lead to accelerated neurodegeneration and ageing in Drosophila melanogaster [64]. A detailed behavioral and neuroanatomical analysis, as well as a careful characterization of the long-term consequences of miR-34-loss will be essential to confirm and extend these hypotheses in mice.
In conclusion, we have reported the generation and characterization of miR-34-deficient mice with a particular focus on the consequences of miR-34 loss on the p53 pathway. The genetically engineered mouse models described in this study will be essential to further investigate the physiologic functions and the tumor suppressive potential of this important miRNA family.
The “recombineering” method [65] was used to modify a BAC clone (RP-23-410P10) containing the miR-34a locus to generate the miR-34a conditional knockout allele. A frt-Neo-frt-loxP cassette was first inserted ∼480 bp downstream of the pre-miR-34a sequence. Gap-repair was used to retrieve a 9.6 kbp fragment containing the frt-Neo-frt-loxP cassette, ∼4 kb of 3′ homology arm, and ∼3.7 kb 5′ homology arm, and including the pre-miR-34a sequence. The fragment was cloned into the targeting plasmid pKS-DTA, and a second loxP site was introduced into a unique KpnI site located ∼500 bp upstream of the pre-miR-34a sequence. The final targeting construct was linearized with NotI and electroporated into V6.5 murine embryonic stem cells (ESC). Following selection with G418, ESC colonies were isolated and screened by Southern blotting using DNA probes mapping outside the targeted region. Two targeted clones were expanded and injected into C57BL/6 blastocysts to generate chimeric mice. High contribution chimeras were subsequently crossed to Actin-flpe transgenic mice [66] to excise the frt-Neo-frt cassette and generate the miR-34a conditional knockout allele (miR-34afl) or crossed to CAG-Cre mice [67] to excise the entire region flanked by the loxP sites and obtain the constitutive miR-34a KO allele (miR-34aΔ). Lastly, miR-34a+/fl and miR-34a+/− were intercrossed to obtain miR-34afl/fl and miR-34a−/− animals.
To generate mice carrying deletion of the miR-34b∼c bicistronic cluster, we used recombineering to replace a 1.3 kbp DNA region in BAC RP-23-281F13 containing pre-miR-34b and pre-miR-34c with a frt-Neo-frt cassette. A 8.4 kbp DNA fragment containing the frt-Neo-frt cassette, the 3.7 kbp 5′ homology arm, and 2.8 kbp of 3′ homology arm was retrieved from the engineered BAC and cloned into pKS-DTA. The resulting targeting vector was linearized by NotI and electroporated into V6.5 ESCs. Upon selection, two independent clones were injected into C57BL/6 blastocysts. High contribution chimeras were crossed to Actin-flpe transgenic mice for germline transmission of the targeted allele and to delete the Neo cassette resulting in the miR-34b∼cΔ allele. The miR-34b∼c+/− mice were intercrossed to obtain miR-34b∼c−/− animals. The Eμ-Myc mice were generated and described by Adams and colleagues [53] and the p53−/− mice were generated by Jacks and colleagues [44]. Genotyping protocols are provided in Text S1. All animal studies and procedures were approved by the MSKCC Institutional Animal Care and Use Committee. Mice were maintained in a mixed 129SvJae and C57BL/6 background. The miR-34a<floxed> mice and the miR-34b∼c−/− mice are available to the research community through The Jackson Laboratory (JAX Stock Numbers 018545 and 018546).
Primary MEF lines were generated from E13.5 embryos using standard protocols. miR-34TKO/TKO embryos were obtained by intercrossing miR-34 mutant mice. Wild-type MEFs were generated in parallel. p53−/− embryos were obtained by intercrossing p53+/− mice. Genotyping protocols are provided in Text S1. MiR-34 wild-type and miR-34TKO/TKO MEF lines were also verified by qPCR.
RNA extraction was performed by homogenizing tissues and cells in TRIzol reagent (Invitrogen) according to manufacturer's instructions. For Northern blotting, 15 µg of each RNA sample was loaded into a 15% Urea-PAGE gel and blotted onto a Hybond-N+ nylon membrane (GE Healthcare). The blots were then hybridized with 32P-labeled probes specific for miR-34a, miR-34c, and U6. qPCR was performed using primers and probes by Applied Biosystems according to manufacturer's instructions. Sno-135 was used for normalization.
Passage 2 or 3 primary MEFs were used for all experiments and cultured at 37°C (5% CO2) in DME-HG with 10% FBS (complete medium) or 0.1% FBS (starvation medium) supplemented with L-glutamine, penicillin, streptomycin, and β-mercaptoethanol. For BrdU cell cycle analysis, wild-type, miR-34TKO/TKO, and p53−/− MEFs were plated in complete medium at 70% confluence, treated with varying doses of doxorubicin for 16 hours or treated at different time points, and pulsed with 10 µM BrdU for one hour. The BD Pharmingen APC-BrdU kit was used to process harvested samples and used according to manufacturer's protocol. For the irradiation experiments, 150,000 wild-type, miR-34TKO/TKO and p53−/− MEFs were seeded into each well of a 6-well culture plate and starved for 72 hours. MEF lines were then trypsinized and resuspended in complete medium and either irradiated (20 Gy, Cs-137 irradiator, Shepherd Mark-I) or left untreated. Cells were replated into complete medium containing 500 ng/ml colcemid at 70% confluence and harvested 24 h later. Samples were processed as mentioned above and stained with 7-AAD. Flow cytometry was performed using FACSCalibur (BD Biosciences), and data were analyzed using FlowJo software (TreeStar).
Wild-type and miR-34TKO/TKO MEFs were seeded into a 6-well plate (40,000 cells/well) and counted every day for the growth curves. The standard 3T3 protocol was followed to determine the cumulative population doublings of wild-type, miR-34TKO/TKO, and p53−/− MEFs. Briefly, 3×105 cells were seeded in a 6 cm2 dish and counted and passaged every three days.
Thymocytes were isolated from sex-matched, age-matched wild-type, miR-34TKO/TKO, and p53−/− mice and seeded at a density of 1×106 cells/ml in MEF medium. Thymocytes were then treated with various doses of irradiation (2, 4, 6, 8, and 10 Gy, Cs-137 irradiator, Shepherd Mark-I) or left untreated. For the time course experiments, thymocytes were treated with 5 Gy of irradiation and harvested 4, 8 and 24 h after treatment. Samples were stained with AnnexinV and propidium iodide (Roche) according to manufacturer's protocol. Flow cytometry was performed using FACSCalibur (BD Biosciences) and data were analyzed using FlowJo software (TreeStar).
Phoenix cells (Orbigen) were transfected using FUGENE 6 (Promega) with retroviral constructs of K-RasV12 alone or together with E1A according to manufacturer's instructions. Wild-type, miR-34TKO/TKO, p53−/− MEFs were seeded at 70% confluence and infected with virus. Plates were fixed with methanol and stained with crystal violet two weeks after infection. Foci were quantified using ImageJ.
Cells were lysed in RIPA buffer containing protease inhibitors. Proteins (25 µg) were separated on a NuPAGE Bis-Tris gel (Invitrogen), and transferred onto a PVDF membrane (Millipore). Blocking was performed with 5% milk in TBST. Primary antibodies used were anti-p21 (1: 1000, Santa Cruz, F-5), anti-Mdm2 (1∶1000, Abcam, 2A10), anti-Met (1∶1000, Millipore, 07-283), anti-Bcl2 (1∶500, Cell Signaling, #2876S), anti-E2f3 (1∶500, Millipore, PG37), anti-Sirt1 (1∶1000, Cell Signaling #2028), anti-cMyc (1∶1000, Cell Signaling, D84C12), and anti-α-Tubulin (Sigma, DM1A). The anti-p53 antibody (1∶300) was a kind gift of Kristian Helin (BRIC, Denmark). Secondary antibodies were obtained from Cell Signaling. ECL reagents were obtained from GE Healthcare. Western blot bands were quantified using ImageJ.
Mice were irradiated with 10 Gy and sacrificed 6 hours after. PFA-fixed, paraffin-embedded sections were deparaffinized in xylene, and rehydrated. The samples were stained with Cleaved Caspase-3 antibody(Cell Signaling, #9664) overnight, according to Cell Signaling protocol. The samples were also counterstained with 0.1% alcoholic Eosin Y solution (Sigma-Aldrich) or 30% hematoxylin. The sections were then dehydrated and mounted in Permount (Fisher Scientific). Sample pictures were quantified using ImageJ.
|
10.1371/journal.pgen.1005981 | A Hox Transcription Factor Collective Binds a Highly Conserved Distal-less cis-Regulatory Module to Generate Robust Transcriptional Outcomes | cis-regulatory modules (CRMs) generate precise expression patterns by integrating numerous transcription factors (TFs). Surprisingly, CRMs that control essential gene patterns can differ greatly in conservation, suggesting distinct constraints on TF binding sites. Here, we show that a highly conserved Distal-less regulatory element (DCRE) that controls gene expression in leg precursor cells recruits multiple Hox, Extradenticle (Exd) and Homothorax (Hth) complexes to mediate dual outputs: thoracic activation and abdominal repression. Using reporter assays, we found that abdominal repression is particularly robust, as neither individual binding site mutations nor a DNA binding deficient Hth protein abolished cooperative DNA binding and in vivo repression. Moreover, a re-engineered DCRE containing a distinct configuration of Hox, Exd, and Hth sites also mediated abdominal Hox repression. However, the re-engineered DCRE failed to perform additional segment-specific functions such as thoracic activation. These findings are consistent with two emerging concepts in gene regulation: First, the abdominal Hox/Exd/Hth factors utilize protein-protein and protein-DNA interactions to form repression complexes on flexible combinations of sites, consistent with the TF collective model of CRM organization. Second, the conserved DCRE mediates multiple cell-type specific outputs, consistent with recent findings that pleiotropic CRMs are associated with conserved TF binding and added evolutionary constraints.
| Enhancers are regulatory elements that interact with transcription factor proteins to control cell-specific gene expression during development. Surprisingly, only a subset of enhancers are highly conserved at the sequence level, even though the expression patterns they control are often conserved and essential for proper development. Why some enhancer sequences are highly conserved whereas others are not is not well understood. In this study, we characterize a highly conserved enhancer that regulates gene expression in leg precursor cells. We find that this enhancer has dual regulatory activities that include gene activation in thoracic segments and gene repression in abdominal segments. Surprisingly, we show that the conserved enhancer can tolerate numerous sequence changes yet mediate robust transcription factor binding and abdominal repression. These findings are consistent with abdominal transcription factors binding numerous different configurations of binding sites. So, why is this enhancer highly conserved? We found that overlapping sequences within the enhancer also contribute to thoracic activation, suggesting the enhancer sequences are under added functional constraints. Altogether, our results provide new insights into why some enhancers are highly conserved at the sequence level while others can tolerate sequence changes.
| The generation of cell-specific gene expression patterns during development is critical for proper morphogenesis. Gene expression at the transcriptional level is controlled by cis-regulatory modules (CRMs), which recruit transcription factor (TF) complexes that alter RNA polymerase activity [1–4]. In general, CRMs are relatively short genomic regions containing clustered binding sites for numerous sequence-specific TFs. CRM activity is determined by which TFs are expressed in each cell and the ability of these TFs to form active transcription complexes on CRM sequences [2,5]. Recently, large-scale genomic studies have identified thousands of CRMs [6–11]. Furthermore, human studies have increasingly found disease-associated single-nucleotide polymorphisms (SNPs) within putative CRMs [6–11]. Hence, understanding how CRMs integrate the appropriate combination of TFs to yield cell-specific transcriptional outcomes is fundamental to understanding both normal development and disease.
Two aspects of TF biology make it hard to predict CRM activity based on primary sequence. First, most TFs bind short degenerate DNA sequences present in high copy numbers throughout the genome [12]. Hence the number of potential genomic binding sites for a TF can exceed the number of TF molecules within a nucleus [13]. Second, the number of TFs encoded in the metazoan genome (>1000 in the human genome) makes predicting which specific TFs bind and regulate a CRM difficult [12]. For example, most TFs are members of large protein families that bind similar DNA sequences, yet CRMs are typically regulated by only one or a small subset of factors from each TF family [12]. Thus, the challenge lies in predicting which particular TFs will functionally bind which of the multitude of potential TF binding sites.
To better understand this problem, three models have been proposed for how CRMs integrate transcriptional inputs: the enhanceosome, the billboard, and the TF collective [5,14,15]. All three models require clustered TF binding sites, but they differ in both sequence conservation and modes of TF recruitment. Enhanceosomes are highly conserved, and recruit a highly cooperative TF complex. Known enhanceosomes have rigid constraints on the order, spacing, and orientation of binding sites, and point mutations in single sites disrupt both complex formation and transcriptional output. The best-characterized enhanceosome is the interferon-β enhancer that coordinates the stepwise recruitment of a series of TFs to mediate high levels of transcriptional activation following viral infection [15,16]. In contrast, billboard CRMs are characterized by flexible orientations/spacing of binding sites that recruit TFs independently and are thereby under less evolutionary constraint [14,17]. The rapid evolution and rearrangement of binding sites within the even-skipped (eve) stripe 2 enhancer in dipterans supports the flexible billboard model [18–21]. The TF collective model proposes that groups of TFs form cooperative complexes on CRMs via a combination of protein-DNA and protein-protein interactions [5]. Unlike the enhanceosome, however, the TF collective posits that protein-protein interactions provide flexibility that eases binding site constraints. For example a TF can be recruited to CRMs lacking its binding site as long as there are sufficient sites for the other TFs of the collective. A collective of five TFs form transcription complexes on numerous CRMs containing various combinations of TF binding sites to regulate gene expression in the Drosophila heart [22].
The differing requirements for how TF sites are organized between the enhanceosome, billboard, and TF collective models may help explain the varying degree of sequence conservation between CRMs. Genomic sequencing of related species revealed that only a subset of CRMs involved in regulating developmentally important genes are highly conserved [23]. For example, the Drosophila vestigial boundary enhancer contains blocks of high sequence conservation, while the eve stripe 2 enhancer is not highly conserved at the sequence level [19,24,25]. This raises an interesting question; why are only some developmentally important CRMs highly conserved? While the answer is currently unclear, one reason may be the different ways CRMs integrate TFs. The enhanceosome model requires tight constraints on TF binding sites consistent with high sequence conservation. In contrast the billboard and TF collective models relax constraints on binding sites, consistent with rapid sequence turnover. Unfortunately, few highly conserved CRMs have been thoroughly dissected and thus, we lack an understanding of which models best explain CRM function and conservation.
The DMX is a conserved CRM that activates the Distal-less (Dll) appendage selector gene in thoracic segments to initiate leg development [26–28]. While activators that can stimulate the DMX are also present in the abdomen, DMX activity is restricted to the thorax via a highly conserved sequence (the Distal-less conserved regulatory element, DCRE) [27,28]. Previous studies demonstrated that the DCRE represses transcription by recruiting TF complexes containing abdominal Hox factors (either Ultrabithorax (Ubx) or Abdominal-A (Abd-A)), Extradenticle (Exd), Homothorax (Hth), Engrailed (En), and the FoxG Sloppy-paired TF (Slp1 and Slp2, referred to here as Slp) [28,29] (Fig 1A). Like the Hox factors, Exd (vertebrate Pbx) and Hth (vertebrate Meis) are conserved homeodomain TFs that regulate segment identity and cell fates along the anterior-posterior axis of metazoans [30–32]. Exd and Hth form cooperative TF complexes with Hox factors on DNA via several protein-protein interactions, and the DCRE recruits an abdominal Hox/Exd/Hth/Hox complex via two Hox binding sites (Hox1 and Hox2) that are coupled to either adjacent Exd (Exd1) or Hth sites (Fig 1A). DCRE-mediated repression also requires compartment-specific inputs with an En site needed for posterior-compartment repression, and FoxG (Slp) sites are required for anterior-compartment repression (Fig 1A and [28]).
Based on the presence of high sequence conservation, one may reasonably predict that a highly conserved CRM such as the DCRE indicates constrained interactions between TFs as in the enhanceosome model. Here we provide evidence that despite high sequence conservation, the DCRE is most consistent with the TF collective model of CRM function. First, we used quantitative transgenic reporter and DNA binding assays to show that the DCRE contains an additional Exd/Hox site (Exd0/Hox0, Fig 1A), and that multiple combinations/configurations of linked Hox/cofactor binding sites can mediate robust transcriptional repression. Unlike the independent TF binding of the billboard model, however, we found that abdominal Hox, Exd, and Hth factors mediate cooperative TF complex formation on the DCRE. Moreover, cooperative complex formation and transcriptional repression can tolerate both individual DNA binding site mutations as well as deletion of the Hth DNA binding domain. These findings are consistent with the TF collective model of CRM function. However, we also found that the linked Hox/cofactor sites in the DCRE enhance thoracic Dll expression in a Hox-dependent manner, and that the re-configured Hox/cofactor binding sites failed to perform all DCRE-dependent functions. Taken together, these findings suggest that the pleiotropic functions of the DCRE (thoracic activation and abdominal repression) add constraints that limit sequence variation, thus providing a potential mechanistic understanding for why some CRMs are highly conserved.
Thoracic Distal-less (Dll) expression is essential for the specification of leg precursor cells of the Drosophila embryo [26,33]. Previous studies identified a conserved Dll CRM, the DMX, which mediates early thoracic leg expression [26]. DMX contains two distinct regions: the DMEact (bp 1–661), which activates gene expression in thoracic and abdominal segments, and an abdominal repression element (Fig 1A) [26,28,34]. The repression element has been defined several times based on different criteria including restriction enzyme sites (“NRE-BX” bp 681–877 [26]), functional studies (“DllR” bp 681–713 [27]), and genomic conservation (“DMXR” bp 675–731 [28]). In this study, we use conservation across 21 Drosophila species to define the repression element as the Distal-less Conserved Regulatory Element (DCRE), bp 662–731, (S1 Fig). This conserved sequence contains six previously characterized TF binding sites, including the linked Hox1/Exd1 and Hth/Hox2 sites that recruit a cooperative abdominal Hox complex as well as FoxG (Slp) and En binding sites, all of which are required for complete abdominal segment repression (Fig 1A and [26,28,35]).
Our current understanding of DMX function suggests the DMEact (1–661) mediates equal activation in all body segments (thorax and abdomen) and the DCRE (662–731) mediates abdominal repression to restrict expression to the thorax. To test these ideas, we integrated DMEact-lacZ (DCRE-lacking) and DMX-lacZ (DCRE-containing) into the same genomic locus and measured β-gal expression normalized to thoracic Dll expression in age-matched embryos. If the DCRE only contributes to abdominal repression, then DMEact-lacZ and DMX-lacZ embryos should have equal levels of thoracic expression. However, the DCRE-lacking DMEact-lacZ embryos express β-gal in significantly fewer thoracic cells, and those that do, express β-gal at lower levels when compared to DMX-lacZ embryos (Fig 1B–1E). Next, we determined if the DMEact is capable of equal activation in thoracic and abdominal segments in the absence of the DCRE by comparing thoracic versus abdominal gene expression in DMEact-lacZ embryos. We found significantly fewer abdominal cells express β-gal and those that do have reduced levels compared to thoracic cells (Fig 1C–1E). Taken together, these findings show that the DMEact and DCRE each contribute to thoracic and abdominal gene regulation, and together yield expression differences between the thorax and abdomen.
Because thoracic and abdominal DMEact-lacZ levels differ, we hypothesized that abdominal Hox factors repress the DMX in a DCRE-independent manner. To test this idea, we mis-expressed Abd-A or Ubx using Paired-Gal4 (PrdG4) and measured DMX-lacZ and DMEact-lacZ activity in the thorax. PrdG4 is active in every other segment, which allows for direct comparisons between experimental (T2) and wild type segments (T1/T3). Care was taken to use conditions that express near physiological levels of Ubx and Abd-A (see Materials and Methods). As expected, either Ubx or Abd-A repressed approximately 80% of DMX-lacZ activity in experimental (T2) segments relative to control T3 segments (Fig 1F and 1H and 1J). Importantly, either also repressed DMEact-lacZ, though to a lesser extent than DMX-lacZ (~40%, Fig 1G and 1I and 1J), indicating that Hox factors repress the DMEact either through direct binding or indirectly through the repression of thoracic activators. Thus, abdominal Hox factors repress the DMX through DCRE-dependent and DCRE-independent mechanisms.
To better characterize Hox-mediated regulation of the DCRE, we generated two synthetic transgenic reporter assays to isolate the DCRE from other DMX regulatory sequences. First, we created an abdominal repression assay by placing lacZ under the control of three copies of the Grainyhead-binding element 1 (GBE-lacZ) (Fig 2A). Embryos containing GBE-lacZ exhibit strong uniform epidermal expression during stage 15 [36] (Fig 2B and 2C). Incorporating the DCRE (GD-lacZ) resulted in a pronounced decrease in β-gal expression within a subset of abdominal cells compared to GBE-lacZ embryos (Fig 2B–2E). Previous studies showed that the DCRE mediates repression in a compartment-specific manner within the context of the DMX enhancer [28]. In the posterior compartment, abdominal Hox factors repress with Engrailed (En), whereas in the anterior compartment they repress with the FoxG factors, Sloppy-paired (Slp1 and Slp2). In the GD-lacZ assay, the DCRE is sufficient to repress transcription within abdominal cells that express Slp (Fig 2E). However, the DCRE is not sufficient for posterior compartment repression, suggesting that En and Hox repression through the DCRE requires additional sites within the DMEact. Quantification of β-gal levels in Slp2+ cells of GD-lacZ embryos revealed a 70% decrease in abdominal segments relative to thoracic segments, whereas β-gal levels were equivalent between Slp2+ thoracic cells and Slp2-negative thoracic and abdominal cells (Fig 2F). Importantly, repression in Slp2+ cells is DCRE-dependent as no difference in β-gal was observed between thoracic and abdominal Slp2+ cells in GBE-lacZ embryos (Fig 2G). Thus, GD-lacZ is a quantifiable assay to study the mechanisms of DCRE-mediated abdominal repression in Slp+ cells.
The second synthetic reporter assay consists of lacZ under control of two copies of the upstream activation sequence (UAS) that can be activated by Gal4 (2xUAS-lacZ) (Fig 3A). When 2xUAS-lacZ is crossed to ubiquitous Gal4 drivers such as armadillo-Gal4 (ArmG4), relatively weak, stochastic expression is observed in stage 11 embryos (Fig 3B). Incorporating the DCRE into the 2xUAS reporter (2xUD-lacZ) and crossing to ArmG4 surprisingly did not reveal abdominal repression, suggesting the DCRE cannot repress Gal4-mediated activation (Fig 3C and 3D and 3E). However, consistent with the DCRE enhancing thoracic expression in the context of the DMX, analysis of 2xUD-lacZ activity in the thorax revealed a 2 to 3 fold increase in β-gal levels relative to control 2XUAS-lacZ embryos (Fig 3B and 3C and 3E). Note, we also observed enhanced thoracic expression relative to abdominal segments in early GD-lacZ embryos, but this difference is lost in older embryos due to the uniform increase in strength of the grainy-head activator (compare thoracic reporter activity in Slp2+ and Slp2- cells in Fig 2F). To better quantify the effect the DCRE has on thoracic gene expression in the UAS assay, we incorporated a control 2xUAS-GFP reporter and found that while 2xUAS-GFP and 2xUAS-lacZ are both expressed stochastically, the relative levels of the two reporters are equal between the thorax and abdomen (Fig 3B–3D). In contrast, β-gal expression from 2xUD-lacZ is significantly increased relative to 2xUAS-GFP expression in thoracic but not abdominal cells (Fig 3D and 3E). A similar induction was observed using different drivers (Tubulin-Gal4, Daughterless-Gal4) yet no expression was observed in 2XUD-lacZ embryos lacking a Gal4 driver (S2 Fig). Hence, the DCRE is insufficient to initiate gene expression on its own, but it can selectively enhance transcription in thoracic segments.
Like abdominal repression in the GD-lacZ assay, enhanced thoracic activation of 2xUD-lacZ was observed in only a subset of cells, even though ArmG4 is active throughout these segments as shown by 2xUAS-GFP expression (Fig 3C and 3D). Co-stains revealed that enhanced β-gal largely overlaps with Dll+ cells and a group of Vestigial (Vg)-positive cells that arise from the Dll+ leg primordia (Fig 3F) [33,37]. These results are consistent with the finding that the DCRE-containing DMX-lacZ expresses significantly higher β-gal in Dll+ cells of the thorax than the DCRE-lacking DMEact-lacZ (Fig 1D and 1E).
Altogether, these results support a model whereby the DCRE mediates multiple cell-specific transcriptional outputs: In the abdomen, the DCRE is sufficient to repress transcription in a cell-specific manner (Slp+ cells) in the anterior compartment. In addition, the DCRE contributes to abdominal repression in the posterior compartment in the context of the DMX [28], but the DCRE is not sufficient to perform this function in isolation from the other DMX sequences. In the thorax, the DCRE functions as a conditional activation element that does not initiate expression but can increase transcription of both endogenous (DMEact) and heterologous (2xUAS) enhancers in the leg primordia. Thus, the GD-lacZ and UD-lacZ assays provide tools that can be used to study the role of Hox, Exd, and Hth factors in regulating a subset of DCRE-mediated activities in isolation from the other DMX regulatory sequences.
The published model of DCRE-mediated repression in the anterior compartment requires an abdominal Hox factor (Ubx or Abd-A), the Exd and Hth cofactors, and a FoxG Slp factor [26,28]. However, genetic removal of hth, exd, or Slp results in severe embryonic defects, including the loss of wingless (wg) expression, which is required for DMX activation [33,38,39]. Since GD-lacZ does not require Wg for activation, it provides a useful tool for genetic tests of these factors. While a deletion removing both Slp1 and Slp2 (Slp∆34b) results in gross morphological abnormalities due to segmentation defects [29], GD-lacZ expression is equal in the thorax and abdomen of Slp mutant embryos (Fig 4A and 4B). Thus, Slp factors are required to mediate DCRE repression. To assess the roles of Hth and Exd, we took advantage of the finding that hth and exd are co-dependent for proper function; genetic removal of hth results in exclusion of Exd protein from the nucleus [30,40,41]. Hence, we assayed GD-lacZ activity in a severe hypomorph of hth (hthP2) and found abdominal repression is abolished (Fig 4C). Since abdominal Hox factors are expressed in both Slp and hth mutant embryos [40,42], these findings demonstrate abdominal Hox factors are insufficient to mediate DCRE repression. However, at least one abdominal Hox factor is required for repression. GD-lacZ activity in single Ubx1 and Abd-AM1, and double Ubx1Abd-AMX1 null embryos revealed that either abdominal Hox factor mediates DCRE-repression whereas removal of both abolishes repression (S3 Fig). Together, these data support the model that the DCRE integrates abdominal Hox/Exd/Hth complexes with the Slp FoxG factors to repress abdominal gene expression.
While a role for abdominal Hox factors in repressing Dll was previously established [26], no prior studies revealed a role for a thoracic Hox factor in activating Dll. The best candidate for a potential positive regulator of Dll is the Antennapedia (Antp) Hox factor, as Antp and nuclear Exd/Hth are co-expressed with Dll in thoracic cells that activate 2XUD-lacZ (Fig 5A and 5C and 5E). Moreover, the enhanced thoracic β-gal expression of 2XUD-lacZ is nearly eliminated in Antp25 null embryos as well as in HthP2 embryos that lack both Hth and nuclear Exd (Fig 5B and 5D). These data suggest Antp directly contributes to thoracic Dll expression through the DCRE. To test this idea, we quantified Dll expression in Antp25 null mutants and heterozygous siblings and found a significant reduction of Dll levels (~40%, Fig 5E and 5F and 5K). In addition, we analyzed expression of DMX-lacZ and DMEact-lacZ in Antp25 mutants and found that the DCRE-containing DMX reporter lost over 50% of its thoracic activity in Antp25 null embryos whereas the DCRE-lacking DMEact reporter was not substantially different from heterozygous siblings (Fig 5G–5J and 5L). These data are consistent with Antp increasing DMX-lacZ expression levels in a DCRE-dependent manner.
The behavior of the DCRE in the GD-lacZ, UD-lacZ and DMX-lacZ reporters supports the idea that the DCRE conveys multiple transcriptional outcomes: thoracic activation versus abdominal repression. Moreover, genetic analysis revealed that both activities are Hox-dependent; Antp for activation and abdominal Hox factors for repression. To assess Hox factor binding to the DCRE, we performed comparative electromobility shift assays (EMSAs) using equimolar concentrations of Antp or Abd-A in the absence and presence of Exd/Hth. We found that Abd-A or Antp weakly bound the DCRE in the absence of Exd/Hth, whereas inclusion of Exd/Hth resulted in highly cooperative complex formation with either Hox factor (Fig 6A and 6B and 6D and 6E). However, the Abd-A complex bound DCRE more strongly than Antp, and Abd-A formed a third, slower migrating complex not seen with Antp (arrow in Fig 6E). Since previous studies had identified only two Hox sites, we scanned the DCRE and found a conserved region containing another potential Hox site preceded by a possible Exd site (TTATG, the ‘Hox0’ site and GAAT, the Exd0 site, see Fig 1A). Interestingly, this region coincides with the ‘BX0’ site that was footprinted by an abdominal Hox factor [26]. To assess the nature of the Abd-A and Antp Hox complexes on the DCRE, we assayed complex formation on a series of probes containing one or two linked Hox/cofactor binding sites (S4 Fig) as well as on DCRE probes containing point mutations in one, two, or all three Hox sites (S5 Fig). Neither Abd-A nor Antp formed strong complexes with Exd/Hth on probes containing individual Hox/cofactor sites. However, binding was increased cooperatively on probes containing two or more Hox/cofactor sites, and the number of molecular species observed increased according to the number of Hox/cofactor sites. These findings indicate that nearby Hox/cofactor binding sites contribute to cooperative DNA binding, even if the Hox/cofactor sites are suboptimal (the Exd0 sequence differs from the consensus sequence and the Exd1/Hox1 site contains an unfavorable nucleotide between the sites).
To assess the role of each Hox site in mediating DCRE-dependent repression and activation, we utilized site-selective mutagenesis in the GD-lacZ and 2XUD-lacZ assays and quantified gene expression. Though the DCRE mediates both thoracic activation and abdominal repression in the context of the DMX, our assays effectively separate the two processes, allowing us to compare and quantify embryos as follows: 1) GD-lacZ assay: By stage 15 of embryogenesis no difference in β-gal levels was measured between cells across the thoracic segment (compare Slp2+ versus Slp2-negative thoracic cells in Fig 2F), indicating that localized DCRE-mediated thoracic activation is not observed at this stage of embryogenesis in the GD-lacZ assay. In addition, like the GBE-lacZ, no differences in levels were observed between Slp2-negative thoracic and abdominal cells in GD-lacZ embryos (see Fig 2F). Thus, thoracic DCRE-mediated activation was negligible in the GD-lacZ assay of stage 15 embryos, and we made direct comparisons between the T3 segment and the remaining thoracic and abdominal segments. 2) UD-lacZ assay: Our data indicates that the DCRE does not mediate significant abdominal repression in the UD-lacZ assay. In fact, quantification of β-gal intensity relative to Dll intensity in 2xUAS-lacZ and 2xUD-lacZ embryos reveals the DCRE significantly alters thoracic but not abdominal expression (Fig 3E). Thus, we normalized thoracic 2xUD-lacZ β-gal levels to the A1 segment for each construct.
To assess the dependence of DCRE abdominal repression on Hox/Hox cofactor sites, we first generated mutations in each Hox site or Hox cofactor site in the GD-lacZ assay. In each case, we found a significant decrease in DCRE-mediated repression in Slp+ abdominal cells indicating that all sites are required for optimal repression (Fig 7A and 7B and S6 Fig). However, no single point mutation abolished repression whereas double and triple Hox site mutations resulted in a complete loss of abdominal repression (Fig 7B and S6 Fig). These findings are consistent with previous mutation analysis on the DMX, which revealed double site mutations were required to yield full de-repression [28]. Taken together with the Hox DNA binding assays, these results indicate that the multiple linked Hox/cofactor sites in the DCRE can mediate robust Abd-A/Exd/Hth complex formation capable of abdominal transcriptional repression.
To assess whether thoracic activation by Antp/Exd/Hth complexes on the DCRE was also dependent upon the Hox binding sites, we analyzed the effect of single point mutations within each Hox site or Hox cofactor site using the 2XUD-lacZ assay. We found that thoracic activation was dependent upon both the Hox0 and Hox1 and their associated cofactor sites (Exd0 and Exd1, respectively) but not the Hox2 or its associated Hth site (Fig 7C and S7 Fig). Hence, unlike abdominal repression, thoracic activation in the 2XUD-lacZ assay is abolished by individual mutations in a subset of the Hox/cofactor binding sites.
Of the three major models of CRM function (billboard, enhanceosome, TF collective), our results are most congruent with Hox factors, especially Abd-A, functioning as a TF collective with Exd and Hth on the DCRE. First, unlike the all or none activity predicted by the enhanceosome model, the DCRE mediates significant repression even when individual TF binding sites are mutated in both the GD-lacZ and DMX assays (Fig 7B and S6 Fig and [28]). Second, we found that unlike the independent binding of TFs predicted by the billboard model, Abd-A/Exd/Hth forms multiple cooperative complexes using several distinct binding sites, and can even do so with individual binding sites mutated (S4 Fig and S5 Fig). An additional postulate of the TF collective is that not all TFs of the collective are required to directly bind DNA to contribute to transcriptional activity. Indeed, while individual point mutations within the sole Hth binding site decreased DCRE-mediated abdominal repression in the GD-lacZ assay, significant repression was still observed in this assay as well as in the context of the full DMX (Fig 7 and [28]). As a further test of this idea, we used a hth point mutation (allele hth100.1) that inserts a premature stop codon to generate homeodomain-less Hth proteins [43]. Importantly, this allele mimics a naturally occurring alternative splice isoform of Hth (as well as the vertebrate Meis proteins), and while these Hth∆HD proteins fail to directly bind DNA, they still interact with and translocate Exd into the nucleus [44]. As expected, we found that 2XUD-lacZ activated thoracic expression in ArmG4;hth100.1 embryos to a level similar to wild type embryos, demonstrating that Hth DNA binding is not required for this activity (Fig 8A and 8B). We also analyzed GDZ activity in hth100.1 embryos, and found significant repression in abdominal Slp2+ cells, albeit, the level of repression was reduced in hth100.1 embryos compared to wild type embryos (45% versus 70% repression, Fig 8C–8E). By comparison, repression is abolished in hthP2 null embryos (Figs 8C and 4C). This data is consistent with a previous study that reported normal Dll and DMX expression in hth100.1 embryos [44]. We confirmed this finding by quantifying DMX-lacZ expression in wild type and hth100.1 embryos and found no significant difference in abdominal repression (S8 Fig). We also tested Hox point mutant-carrying GD-lacZ reporters in the context of hth100.1 embryos. As expected, point mutations within the Hox2 site, which is linked to the adjacent Hth site, did not further decrease GD-lacZ dependent repression in hth100.1 embryos (S6 Fig). In contrast, Hox1 point mutations in this genetic background lost all repression activity, a result that is consistent with the fact that multiple Hox/cofactor sites need to be mutated to abolish DCRE-mediate repression (Fig 7B and S6 Fig).
Next, we assessed whether the homeodomain-less Hth protein can contribute to cooperative Abd-A DNA binding on the DCRE. We also tested the role of Exd DNA binding on complex formation using an Exd protein containing a point homeodomain mutation (N51A) that disrupts DNA binding. Importantly, purified Exd/Hth∆HD (Fig 6F) and Exd51A/Hth (Fig 6G) heterodimers did not significantly bind the DCRE in the absence of Abd-A, even when added at a concentration three times higher than the wild type heterodimer (compare second column of each EMSA to wild type Exd/Hth binding in Fig 6E). Inclusion of Abd-A, however, revealed that either DNA binding deficient heterodimer (Exd/Hth∆HD or Exd51A/Hth) stimulated significant cooperative Hox complex formation on the DCRE (Fig 6F and 6G). To determine the independent role of Hth and Exd protein in complex formation, we performed EMSAs using Abd-A with only purified Exd or Hth (Fig 6H and 6I). In contrast to the DNA binding deficient heterodimers, the addition of equimolar concentrations of Exd or Hth alone with Abd-A did not yield significant complex formation on the DCRE (Fig 6H and 6I). These findings are consistent with the TF collective model of CRM function in which protein-protein interactions between Exd and Hth contribute to cooperative TF complex formation with Abd-A on the DCRE.
To determine if different configurations of Hox/Exd/Hth sites could confer similar transcriptional outcomes, we replaced a subset of the Hox/cofactor sites within the DCRE with a distinct set of sites from another Hox-regulated CRM. Previous studies revealed that a rhomboid CRM (RhoBAD) mediates transcriptional activation in sensory organ precursors by integrating an Abd-A/Hth/Exd complex with the Pax2 TF [45,46]. The RhoBAD CRM contains separable binding sites for Pax2 and Abd-A/Hth/Exd (Fig 9A). To determine if the Hox/Hth/Exd sites found in RhoBAD can function in transcriptional repression in the DCRE, we replaced the Hox1/Exd1-Hox2/Hth sites of the DCRE with the Hox/Hth/Exd sites from RhoBAD (DCRE-RhoA, Fig 9A). This fusion transgene lacks the RhoBAD Pax2 site necessary for activation but contains the DCRE FoxG (Slp) sites as well as the Exd0/Hox0 sites that contribute to, but are not sufficient, for mediating repression. We found that the GD-RhoA-lacZ was able to substantially repress gene expression in Slp+ abdominal cells, although not as strongly as the wild type DCRE (Fig 9B–9D). To determine if this modified element was sufficient to repress the DMX enhancer in the abdomen, we compared the activity of DMX-lacZ and DMX-RhoA-lacZ transgenes. Since the DCRE-RhoA element lacks the En site required for posterior compartment repression, significant de-repression in En+ cells was expected and observed in DMX-RhoA-lacZ (Fig 9F). In contrast, repression of the DMX-RhoA-lacZ was comparable to that of DMX-lacZ in Slp+ abdominal cells (Fig 9E–9G). However, similar to DMEact-lacZ, the DMX-RhoA-lacZ configuration of sites expressed decreased levels of β-gal in the thorax compared to the wild type DMX-lacZ (Fig 9E–9G). Altogether, these findings demonstrate that while the DMX-RhoA configuration of Exd/Hth/Hox sites can mediate significant abdominal repression in Slp+ cells, this configuration of sites failed to perform two other DCRE-dependent activities (posterior compartment repression in the abdomen and conditional activation in the thorax).
While it has been established that CRMs regulate a gene’s spatial and temporal transcription expression pattern, we are only now appreciating the complexity of CRMs regarding the number of inputs required to yield cell/tissue specific patterns. In this study, we built upon our knowledge of how the DCRE CRM integrates Hox, Exd, and Hth TFs to ensure precise Dll expression during leg specification. Using quantifiable transgenic reporter and DNA binding assays, we found that the DCRE can recruit either Hox-based repression (Abd-A/Ubx) or activation (Antp) complexes using multiple Hox/Exd and/or Hox/Hth sites. Importantly, the DCRE Hox, Exd, and Hth binding sites and flanking regions are highly conserved across Drosophilid species, yet our studies reveal that an abdominal Hox TF collective can mediate robust cell-specific (Slp+) repression through flexible combinations of Hox/co-factor binding sites. However, the DCRE regulates at least two additional cell-specific transcriptional outcomes, suggesting that the DCRE CRM TF binding sites are under added constraints and maintains high sequence conservation to mediate multiple cell-specific outputs. Thus, our findings provide new insights into Hox specificity, CRM function, and CRM conservation.
In spite thirty years of study, we lack a general understanding of how Hox factors gain sufficient specificity to differentially regulate cell fates along the anterior-posterior axis of metazoans. As monomers, Hox factors bind highly similar DNA sequences in vitro [47,48]. The discovery of two general Hox cofactors that also encode TFs, Exd (vertebrate Pbx) and Hth (vertebrate Meis), suggested that the formation of TF complexes enhances Hox DNA binding affinity and specificity [32,49,50]. Consistent with this idea, the biochemical characterization of Exd/Hox binding sequences using SELEX-seq revealed DNA binding preferences between Hox factors are enhanced by Exd (termed latent specificity) [51]. The Forkhead (Fkh) CRM, for example, contains a unique Hox/Exd site that is specifically bound and regulated by a Sex combs reduced (Scr)/Exd complex [52,53]. More recent studies revealed that Exd also enhances Hox specificity by binding several low affinity sites. Crocker et al. found two CRMs from the shavenbaby (svb) locus that are activated in the abdomen by either Ubx/Exd or Abd-A/Exd complexes via low affinity sites [54]. Altering these sequences to high affinity Hox/Exd sites resulted in a loss of Hox specificity and transcriptional activation by anterior Hox factors. These findings suggest high affinity Hox/Exd sites are more likely to be pan-Hox target sequences regulated by numerous Hox factors whereas low affinity Hox/Exd sites provide specificity.
In this study, we show that the DCRE mediates two opposing transcriptional outcomes using three linked Hox-cofactor binding sites. In the thorax, an Antp/Exd/Hth complex activates largely via two Hox/Exd sites, whereas the linked Hox/Hth sites are less important for DCRE-mediated activation. In the abdomen, all three Hox sites contribute to repression via the recruitment of several Abd-A/Exd/Hth complexes. Hence, the most specific Hox site within the DCRE is the linked Hth/Hox site that mainly contributes to abdominal repression by binding Abd-A and Ubx (Fig 10). In fact, directly linked Hth/Hox sites may be preferentially regulated by posterior Hox factors as the Abd-A specific target gene rhomboid (rho) contains a CRM that is activated via a linked Hth/Hox site [45,46]Additionally, biochemical studies using vertebrate Hox factors revealed that only posterior Hox factors form direct complexes with the Meis factor on DNA [55]. In contrast, both Exd/Hox sites within the DCRE are regulated by both thoracic Hox factors (activation) and abdominal Hox factors (repression) (Fig 10). Sequence analysis reveals that neither DCRE Exd/Hox site is optimal as an extra nucleotide is inserted between the Hox1 and Exd1 site whereas the Exd0 site has several mismatches to its consensus sequence (S1 Fig). Moreover, DNA probes containing isolated Exd/Hox sites from the DCRE are poorly bound by Hox/Exd proteins, whereas combining these suboptimal sites resulted in the formation of Hox complexes that contribute to gene regulation. Thus, the DCRE uses multiple Hox/Hox cofactor sites to recruit distinct complexes that mediate two opposing transcriptional outcomes along the anterior-posterior axis.
While a repression function for the DCRE was expected based on previous studies, the DCRE also contributes to Hox-mediated activation in the thorax. We termed the DCRE a ‘conditional’ activator in the thorax because it fails to initiate transcription, but when coupled to a ubiquitous activation element the DCRE enhances transcription in a subset of thoracic cells. Importantly, the cells that activate the DCRE derive from the endogenous Dll expression domain, and the DCRE contributes to activation of the DMX leg enhancer in an Antp-dependent manner. These data support the model that Antp and Exd/Hth are required for the conditional activation function of the DCRE. However, it is currently unclear why this activity is restricted to the Dll+ leg/wing primordium since Antp and Exd/Hth are broadly expressed throughout the thorax. One possibility is that, much like in the abdomen, an additional factor(s) interacts with the DCRE to provide position-specificity.
How CRMs integrate transcription factor complexes to mediate cell-specific outputs remains an active area of study. The two best-known CRM models are the enhanceosome and the billboard. These models can be seen as extreme opposite ends of the spectrum of rigidity and constraints (enhanceosome) versus flexibility and adaptability (billboard), with most CRMs likely to contain aspects of both models. Since many TFs use protein-protein interactions to promote cooperative complex formation on DNA, these interactions often place constraints on the order, orientation, and spacing of TF binding sites within CRMs. Hence, cooperative DNA binding has often been seen as evidence consistent with an enhanceosome model of CRM function. Dimerization between TFs such as the basic Helix-Loop-Helix (bHLH) proteins and retinoic acid receptors, for example, results in the formation of TF complexes that bind palindromic sequences with restrictions on distances between individual binding sites. In 2012, Junion et al proposed an alternative role for protein-protein interactions between TFs [22]. Using a series of chromatin immunoprecipitation experiments, the Furlong lab found that a group of five TFs regulate a set of cardiac CRMs in the Drosophila embryo. Sequence analysis of co-regulated CRMs revealed combinatorial binding of these TFs does not require specific motif organization, a finding that is also consistent with the billboard model of CRM function. However, unlike the billboard, the TF collective does not require individual DNA binding sites for every TF to mediate appropriate functional outputs. Instead, a TF collective uses a combination of clustered DNA binding sites and protein-protein interactions to recruit large-scale TF complexes containing all the members of the collective. Although the biochemical basis of TF interactions between the five TFs was not explored, previous studies did find that a subset of these TFs form direct protein-protein interactions. Thus, Junion et al proposed the TF collective model of CRM function that predicts a common group of TFs can form many different cooperative complexes via multiple interactions between TFs, which results in greater CRM flexibility rather than rigidity in DNA binding site organization [22].
In this study, we provide evidence consistent with a Hox TF collective regulating early Dll expression in the Drosophila embryo. First, we show that the DCRE uses at least three distinct Hox sites that are each linked to an adjacent Exd or Hth binding site to recruit functional Hox complexes. Focusing on abdominal Hox-mediated repression, we used DNA binding assays and a synthetic reporter system (GD-lacZ) to reveal the following correlations between DNA binding affinity and transcriptional repression: 1) The wild type DCRE containing all three Hox sites yielded the strongest Abd-A/Exd/Hth binding and transcriptional repression in abdominal Slp+ cells. 2) Individual point mutations within any one Hox site partially compromised complex formation and repression. However, significant repression was still observed in the GD-lacZ assay, and in the DMX-lacZ assay single point mutations were still able to mediate abdominal repression in Slp+ cells in the DMX reporter [28]. 3) Mutations that compromise any two Hox sites or two Hox co-factor sites further decreased Abd-A/Exd/Hth complex formation, and abolished GD-lacZ-mediated abdominal repression. Consistent with the TF collective model, we found that Abd-A could still form robust complex formation on the DCRE even in the presence of DNA binding deficient Exd or Hth proteins, and genetic studies revealed that the DNA binding activity of one of the factors (Hth∆HD) is not required to mediate significant abdominal repression or thoracic activation. Moreover, we replaced the Hox1/Exd1-Hth/Hox2 sites with a distinct configuration of Exd/Hth/Hox sites from a different Abd-A regulated CRM and observed significant repression in both the GD-lacZ and DMX-lacZ assays (Fig 10). In total, these data demonstrate that, in the anterior compartment of the abdomen, multiple Hox/Exd/Hth binding site configurations can recruit a Hox TF collective capable of mediating robust transcriptional outputs.
Interestingly, other Hox CRMs also contain characteristics consistent with TF collective enhancers. For example, congruent with variable binding of TFs in a collective, comparison of five mouse hindbrain enhancers controlled by HoxA1 and HoxB1 along with the Exd/Hth homologs, Pbx and Meis demonstrated that the presence, orientation, location, and sequence of the Meis sites are highly variable [56–61]. Additionally, the Hth homeodomainless protein is functional on other Hox-regulated CRMs, including the Fkh250 and Lab550 CRMs in Drosophila embryos [44]. Together, these results suggest that the DCRE is not unique among Hox CRMs in fitting the TF collective model.
An unanswered question emerges from these studies: if interactions between members of the Hox TF collective permit added flexibility in binding site configurations, why is the DCRE so highly conserved across Drosophilid species? One possible reason is that the DCRE mediates multiple opposing Hox-dependent outputs, which places added constraints on sequence conservation. For example, while replacing the Hox1/Exd1-Hth/Hox2 sites with the Exd/Hth/Hox configuration from the RhoBAD CRM can mediate strong repression in Slp+ anterior compartment cells, this configuration fails to repress gene expression in the posterior compartment due to the lack of an En binding site. Similarly, DCRE reporters containing this configuration of Hox/Hox cofactor sites also yielded lower levels of β-gal expression in the thorax, consistent with the idea that Antp fails to regulate linked Hth/Hox sites. Hence, we propose that the dual repression mechanisms of the DCRE in the anterior and posterior compartments of the abdomen as well as its conditional activation function in the thorax requires numerous TF sites, which thereby places evolutionary pressure to maintain sequence conservation.
Several different hypotheses have been proposed for why some CRMs are highly conserved, including pleiotropic functions of CRMs placing added constraints on conservation [62–65]. Moreover, a recent vertebrate study comparing TF binding to syntenic regions of mouse and human genomes revealed that the most highly conserved TF binding activities were found on CRMs with pleiotropic functions in multiple cell types [66]. This study also noted that pleiotropic CRMs enrich for the co-association of many TFs. While this study did not score each CRM for nucleotide identity, their findings are consistent with our functional study on the DCRE and suggest that pleiotropy places added constraints on CRM sequence conservation.
The DMX [28], DMEact (basepairs 1–661 of DMX), 3xGrainyHead binding element1 (3xGBE) [67], and 2xUAS elements were generated by PCR (sequences available upon request). DCRE-containing plasmids were created by ligating annealed complementary oligonucleotides containing restriction enzyme overhangs into the 3xGBE, or 2xUAS plasmids. Sequences of DCRE mutants are located in the figures. All enhancers were subcloned into the placZAttB plasmid. UAS-Abd-A was generated by PCR and subcloned into the pUAST-AttB plasmid. All plasmids were confirmed by DNA sequencing. Transgenic fly lines were generated by ΦC31 integration into the 51C insertion site [68] (Injections by Rainbow Transgenics).
The following fly lines were used: Antp25, Ubx1, hthP2, PrdG4, ArmG4 (Bloomington Stock Center); Abd-AMX1, UbxMx12Abd-AM1, Slp∆34b, UAS-Ubx (Richard Mann, Columbia University, NY, USA); hth100.1 (Kurant et al., 2001); UAS-Abd-A (this work). Embryos were collected, fixed and stained using standard procedures at 25°C except for PrdG4;UAS-Abd-A and PrdG4:UAS-Ubx experiments which were performed at 18°C to lower Gal4 activity. The following primary antibodies were used: En (mouse 1:10) (Developmental Studies Hybridoma Bank, DSHB), Antp (mouse 1:50) (DSHB), Abd-A (guinea pig 1:500) (Li-Kroeger et al., 2008), Ubx (mouse 1:20) (Richard Mann); Vestigial (rabbit 1:25) (Sean Carroll, University of Wisconsin-Madison, WI, USA); and β-gal (chicken 1:1000) (Abcam). Antibodies were generated against Slp2 (amino acids 1–275) and Dll (full-length) using purified His-tagged proteins injected into rats (Cocalico Biologicals). Both the Slp2 and Dll sera were used at 1:500. All immunostains were detected using fluorescent secondary antibodies (Jackson Immunoresearch Inc. or Alexa Fluor, Molecular Probes). For quantitative analysis of gene expression, sets of embryos were harvested, fixed, and imaged under identical conditions at the same time. When possible, age-matched siblings were analyzed. For GD-lacZ and UD-lacZ assays, images used for quantification were taken using a single exposure time and normalized to segment T3, A1, or Dll expression levels within the same embryo as indicated. Pixel intensities and areas were measured using NIH-ImageJ software.
The following proteins were purified from BL21 cells as previously described [27]: His-tagged Abd-A [69]; Antp [27]; his-Hth [70] and untagged Exd heterodimers [27]; his-Hth∆HD/Exd heterodimers [51]; his-Exd51A/Hth heterodimers [56]; his-Hth and his-Exd. Purified proteins were confirmed using SDS-PAGE and Coomassie blue staining and concentrations measured by Bradford assay. EMSAs were performed as previously described using native polyacrylimide gel electrophoresis [56]. Probes were used at 0.36 μM, and protein concentrations are noted in figure legends. The dried acrylamide gels were exposed to a phosphor screen for imaging using a StormScanner (GE Healthcare). Densitometry was performed using ImageQuant 5.1 software. All EMSA experiments were performed in triplicate.
|
10.1371/journal.pntd.0003664 | Control, Elimination, and Eradication of River Blindness: Scenarios, Timelines, and Ivermectin Treatment Needs in Africa | River blindness (onchocerciasis) causes severe itching, skin lesions, and vision impairment including blindness. More than 99% of all current cases are found in sub-Saharan Africa. Fortunately, vector control and community-directed treatment with ivermectin have significantly reduced morbidity. Studies in Mali and Senegal proved the feasibility of elimination with ivermectin administration. The treatment goal is shifting from control to elimination in endemic African regions. Given limited resources, national and global policymakers need a rigorous analysis comparing investment options. For this, we developed scenarios for alternative treatment goals and compared treatment timelines and drug needs between the scenarios. Control, elimination, and eradication scenarios were developed with reference to current standard practices, large-scale studies, and historical data. For each scenario, the timeline when treatment is expected to stop at country level was predicted using a dynamical transmission model, and ivermectin treatment needs were predicted based on population in endemic areas, treatment coverage data, and the frequency of community-directed treatment. The control scenario requires community-directed treatment with ivermectin beyond 2045 with around 2.63 billion treatments over 2013–2045; the elimination scenario, until 2028 in areas where feasible, but beyond 2045 in countries with operational challenges, around 1.15 billion treatments; and the eradication scenario, lasting until 2040, around 1.30 billion treatments. The eradication scenario is the most favorable in terms of the timeline of the intervention phase and treatment needs. For its realization, strong health systems and political will are required to overcome epidemiological and political challenges.
| River blindness (onchocerciasis) is transmitted by blackflies and causes severe itching, skin lesions, and vision impairment including blindness. More than 99% of all current cases are found in sub-Saharan Africa where the disease has historically hindered socioeconomic development in endemic areas. The treatment goal is shifting from control to elimination in Africa as morbidity has significantly decreased through vector control and community-directed treatment with ivermectin. Studies in Mali and Senegal proved that elimination is feasible with ivermectin administration. Given limited resources, national and global policymakers need a rigorous analysis of investment options from epidemiological, economic, and societal aspects. For this, we developed control, elimination, and eradication scenarios and compared treatment timelines and drug needs over the next 30 years. We found that the elimination and eradication scenarios would require a shorter treatment phase and a smaller amount of ivermectin than the control scenario, mainly because community-directed treatment with ivermectin could be ended earlier thanks to regular active surveillance.
| Elimination of neglected tropical diseases (NTDs) has recently emerged on the global health agenda and gained prominence with the release of the global plan to combat NTDs by the World Health Organization (WHO) [1]. In 2012,WHO issued a roadmap towards the elimination of 17 NTDs [2], and stakeholders from the public and private sectors pledged to contribute to the control, elimination, and eradication of ten NTDs through the London Declaration on NTDs [3]. The second WHO report on NTDs further elaborated the roadmap [4], and the London Declaration follow-up report showed the substantial progress that had already been achieved through the stakeholder partnership approach [5].
One of the NTDs targeted for elimination is onchocerciasis (river blindness). This is a parasitic disease caused by filariae that are transmitted by blackflies. Severe itching, skin lesions, and vision impairment including blindness are its most notable symptoms. Onchocerciasis is endemic in parts of Africa, Latin America, and Yemen, but over 99% of all current cases are found in sub-Saharan Africa [6] where onchocerciasis has historically been a serious public health problem and hindered socioeconomic development in endemic areas [7]. However, many infections are asymptomatic, and vector control and community-directed treatment with ivermectin have significantly reduced morbidity. Specifically, the Onchocerciasis Control Program (OCP), which was implemented in West Africa from1975 to 2002, and the African Programme for Onchocerciasis Control (APOC), which has supported onchocerciasis control activities in sub-Saharan countries since 1995 and continued the OCP’s activities where needed, have decreased the burden of disease to such an extent that it is no longer a public health problem in most endemic areas [8]. In Latin America, the Onchocerciasis Elimination Program for the Americas (OEPA) implemented since 1993 has brought the disease close to elimination. Colombia and Ecuador announced the elimination of onchocerciasis after WHO verification in 2013 and 2014, respectively [9,10]. Treatment has also been stopped in seven foci in Guatemala and Mexico where it has been replaced by surveillance to detect possible recrudescence [11]. Regional elimination in Latin America is expected to be achievable by 2020 if the regular treatment of a sufficient proportion of the nomadic Yanomami in the border area between Brazil and Venezuela can be achieved [12]. In Yemen, onchocerciasis is endemic in a limited number of communities. Elimination in the near future is considered technically feasible, and a national action plan aiming at elimination by 2015 was developed in 2010 [13]. Currently, political instability and security concerns that limit access to endemic areas hamper its implementation [4].
Studies in Mali and Senegal have proved the feasibility of onchocerciasis elimination through ivermectin treatment in some hyper-endemic foci in West Africa [14,15]. This has provided additional momentum and arguments for a shift in the strategic goal from control to elimination also in Africa. The decision to invest in elimination and eradication efforts should be informed by broad assessments considering biological and technical feasibility, financial and economic costs, health and economic gains, capacity of and impacts on health systems, and societal and political willingness to cooperate [16]. An approach to such an assessment has been proposed in the form of eradication investment cases in 2010 [17]. Tediosi and colleagues have examined the approach with focus on three NTDs including onchocerciasis [18]. With reference to this approach, we have developed and compared alternative scenarios, namely, staying in a control mode versus moving toward elimination and subsequent eradication.
In the present paper, we describe the scenarios to achieve control, elimination, and eradication of onchocerciasis, predict the timeline of stopping treatment at country level, and estimate the number of required ivermectin treatments over the next 30 years with focus on Africa.
We developed scenarios, describing all required activities and resources that are expected to lead to the goals of control, elimination, and eradication, if effectively implemented and sustained as long as required, based on current standard practice, the results of large-scale studies, and available historical data. To clearly distinguish these alternative scenarios, we referred to the definitions of control, elimination, and eradication endorsed and recommended by the WHO Strategic and Technical Advisory Group for NTDs [19]. The ultimate goals of the scenarios were defined as follows:
1) control scenario: continuing community-directed treatment with ivermectin (CDTi) to keep the prevalence under a locally acceptable level; 2) elimination scenario: scaling up CDTi to all endemic areas where feasible aiming at the reduction of disease incidence to zero; and 3) eradication scenario: including strategies and tailored interventions to overcome operational challenges in endemic areas with feasibility concerns in addition to CDTi with the aim of reducing the global disease incidence to zero (Table 1).
From an operational perspective, the control and elimination scenarios are designed to target endemic areas where interventions appear feasible without major challenges, whereas the eradication scenario is an optimal situation. To make the eradication scenario feasible, intensive efforts to improve operational capacity and to increase political willingness would be required to overcome epidemiological and political challenges. We assume effective treatment would be implemented through tailored approaches in those areas, and regular surveillance would be maintained during and after the intervention phase until eradication has been verified.
Referring to the general principles for developing scenarios outlined by Tediosi and colleagues [18], the key components of scenarios were identified at project level. Scenarios were further revised by verifying the realism of assumptions in consultation with a technical advisory group consisting of policymakers, onchocerciasis epidemiologists, public health experts, health economists, and donors.
Key components for developing the scenarios are defined as follows and the developed scenarios are described in Table 1.
The number of required ivermectin treatments to achieve the goals of the control, elimination and eradication scenarios in endemic African regions was predicted by multiplying the estimated population living in endemic areas with the treatment coverage rate and the CDTi frequency per year for the required duration of treatment at project level. The capacity of drug manufacturers to supply the required number of ivermectin was assumed to be sufficient considering Merck’s commitment to donate ivermectin until elimination is achieved globally [41].
The time horizon for predicting the number of treatments was 2013 to 2045. The start year was set considering the most recent version of the APOC databases available for analysis was for 2012. The end year was chosen based on the prediction that the last project in the eradication scenario would stop CDTi in 2040, and that after stopping CDTi, at least three years would be required to confirm local elimination. In the control and elimination scenarios, the last projects were expected to continue CDTi beyond 2045.
In the S1 Table, the relevant data regarding the key components of the scenarios, which were used for estimating the timelines and the number of required ivermectin treatments, are presented at project level.
Parameters used for the scenario analysis were subject to considerable uncertainty and the impact of the uncertainty was examined for the target population, the timeline when CDTi is expected to be stopped, and the number of required ivermectin treatments. The impact of a single parameter’s uncertainty was assessed with one-way deterministic sensitivity analysis (DSA). Considering the final estimates are driven by the joint effects of multiple parameters, multivariate probabilistic sensitivity analysis (PSA) was conducted with all the variables examined in the one-way DSA.
The included parameters were population growth rate, treatment coverage, treatment duration, CDTi start and end years, and the assumptions for selecting target projects. For DSA, the parameter uncertainty ranges were determined based on available data, expert opinion or both. For PSA, statistical distributions were chosen considering the characteristics of parameters, and fitted to available data. Simulations were run 1,000 times for each scenario.
For each scenario, we predicted target areas in endemic African regions and population in those areas, the timeline when CDTi is expected to be stopped, and the number of required ivermectin treatments.
The control scenario targeted hyper-and meso-endemic areas in all endemic African countries. Under the elimination scenario, CDTi was extended to hypo-endemic areas where CDTi is feasible in addition to hyper-and meso-endemic areas. Countries that include projects with feasibility concerns have been identified to be the Central African Republic, the Democratic Republic of the Congo, and South Sudan due to political instability, and Gabon due to the high prevalence of L. loa in areas with a low prevalence of onchocerciasis. In these four countries, hypo-endemicity areas were therefore excluded from the elimination scenario. The eradication scenario targeted all hyper-, meso-, and hypo-endemic areas. The endemic countries in Africa were categorized into two control programs in which they participate or participated, APOC and OCP, respectively (Table 2).
The control scenario included 27 countries, and potential new projects were predicted to cover around 3% of the total population in the entire target area, or 4.7 million of 144 million (Fig 1). The elimination scenario included the same 27 countries, and new projects were predicted to cover at most 17% of the population in the entire target area (167 million). Depending on the number of new projects in potential hypo-endemic areas, the population in new project areas ranged from 12.1 million to 27.8 million (7% and 17%). The eradication scenario included one more country, Gabon, and the total population in the entire target area was estimated at around 176 million of which 21% at maximum live in new project areas with a range of 12.1 million to 36.5 million people (7% to 21%) depending on the number of new projects in potential hypo-endemic areas.
In the control scenario, most endemic countries outside West Africa were predicted to continue CDTi beyond 2045 (Fig 2). The most influential parameter determining the expected year of ending CDTi was the extension of treatment duration due to insufficient treatment coverage (Fig 3). For the elimination and eradication scenarios, the final year of CDTi represents the year of ending the intervention phase at country level assuming no recrudescence would occur. In the elimination scenario, all endemic countries except the four countries with feasibility concerns were expected to finish the intervention phase by 2028 at the latest and those four countries were expected to continue CDTi beyond 2045 (Fig 2). In the eradication scenario, all endemic countries were expected to reach the end of the intervention phase by 2040 assuming sufficient treatment would be delivered sustainably in the four countries with epidemiological and political concerns. For the elimination and eradication scenarios, one-way DSA (Fig 3) showed that any delay in starting and ending CDTi and low treatment coverage would result in the intervention phase to end later than expected; on the contrary, high treatment coverage would expedite the progress of the intervention phase and lead to an earlier end of the intervention phase.
The need for ivermectin treatments was concentrated in the first half of the time horizon for the elimination and eradication scenarios, as 80% of all potential projects were stopped safely by 2031 and 2025, respectively. In the control scenario, it took until 2038 for the same proportion of the total projects to stop CDTi (Fig 4). The cumulative number of required ivermectin treatments over 2013–2045 was estimated at 2.63 billion (95% central range: 2.41 billion-2.99 billion) for the control scenario. Specifically, 1.48 billion (1.51bn-1.57bn) treatments were predicted to be required until 2025 and 1.15 billion (0.90bn-1.41bn) treatments over 2026–2045 (Table 3). According to the simulation of the elimination scenario, the required number of ivermectin treatments over the whole period was around 1.48 billion (1.42bn-1.79bn). Compared to the control scenario, the total number of required treatments in the elimination scenario was lower by 1.15 billion (44%): 0.45 billion (0.36bn-0.55bn) until 2025 and 0.69 billion (0.38bn-0.92bn) from 2026 to 2045 (Table 3, Fig 5). The eradication scenario required an even smaller number of ivermectin treatments for the whole period, 1.30 billion (1.18bn-1.51bn), which was 0.18 billion (0.03bn-0.49bn), or 12%, lower than that under the elimination scenario and 1.32 billion (0.97bn-1.75bn), or 50%, lower than that under the control scenario (Fig 5). In one-way DSA (Fig 6), the most influential parameter on the cumulative number of required ivermectin treatments was the delay in ending CDTi in all scenarios. For the control scenario, the second most influential parameter was the number of projects with extended CDTi duration due to insufficient treatment coverage. For the elimination and eradication scenarios, it was the number of potential new projects in hypo-endemic areas.
The key changes for shifting from the control mode to elimination and subsequent eradication are the scale-up of CDTi to hypo-endemic areas and the implementation of regular epidemiological and entomological surveys along with ongoing surveillance. For successful implementation of these, overcoming the existing feasibility issues related to the co-endemicity with L. loa, the insecure political situation, and weak health systems will be critical. We found that, if this could be accomplished, regional elimination in Africa could be achieved as early as 2040, and consequently all endemic countries including Latin Americas and Yemen would be in the post-elimination phase until eradication has been verified.
We found that achieving elimination would reduce treatment needs by 43% compared to the control mode for the period 2013–2045. The driver of this remarkable difference is that CDTi could be stopped for the majority of projects based on regular surveillance, while it would have to continue for at least 25 years under the control scenario. The eradication scenario is predicted to require an even smaller number of ivermectin treatments than the elimination scenario, as hypo-endemic areas with feasibility concerns were assumed to have a shorter treatment period through effective treatment via tailored approaches as well as CDTi, whereas those areas would be under the control mode in the elimination scenario. This finding implies that saved ivermectin drugs could be used for other disease programs, for instance, mass drug administration (MDA) for lymphatic filariasis (LF).
The uncertainty about the target population in the elimination and eradication scenarios was mainly driven by uncertainty in the number of potential new projects in hypo-endemic areas, as some of those areas might not be actually endemic. Parasitological surveys are therefore needed to determine the current infection status of those areas. Setting up a new project requires operational planning, human resource mobilization, and startup costs. To move towards elimination without delay and to save human and financial resources, the rapid mapping of potential hypo-endemic areas should be a priority to confirm areas to set up new projects and to develop elimination strategies for those areas.
The main driver of the number of required ivermectin treatments was the delay in stopping CDTi. This finding implies that maintaining high treatment coverage to avoid the extension of treatment duration and continuous monitoring and evaluation to decide a proper time to stop CDTi would lead to faster elimination and prevent unnecessary efforts to deliver drugs.
We assumed no recrudescence in our analysis. However, if recrudescence occurs, the duration of CDTi would need to be extended, local elimination would be delayed, and the number of required treatments would increase. Recrudescence might occur because of human or vector migration, interrupted drug distribution due to political instability, and residual transmission from not-treated endemic areas due to incomplete or inconsistent geographic coverage.
We did not adjust for alternative treatment approaches for areas where L. loa is highly endemic but onchocerciasis is hypo-endemic. Suggested treatment approaches for these areas include anti-Wolbachia therapy with macrofilaricidal drugs, high doses of albendazole, and the test-and-treat strategy [43,44]. These approaches would expedite elimination and increase the demand for other drugs while reducing the need for ivermectin.
Our modeling did not incorporate the impact of changing the CDTi frequency on the treatment duration. It has been suggested to increase the frequency of CDTi to reduce the prevalence and transmission of onchocerciasis faster compared to the annual CDTi [35]. A recent study by Coffeng and colleagues shows that six-monthly ivermectin treatment could reduce the required treatment duration by 40% based on a dynamical transmission model [38]. In practice, increasing the CDTi frequency would require collaboration between policymakers, health workers, and community volunteers and new strategies on how to mobilize human and financial resources, given limited resources and competing health programs. Under the control scenario, annual CDTi could mean overtreatment for projects that had more than 15–20 years of treatment, for example, some areas in West Africa where ivermectin administration has been implemented since the 1990s. For these areas, less frequent CDTi could be an alternative for morbidity control, which would require a smaller number of ivermectin and less human and financial resources. However, less frequent CDTi might lead to a loss of local expertise, human resources, and community compliance over the time interval without CDTi and, consequently, to the decrease of treatment coverage below the required level, which could expose the areas to the risk of recrudescence.
We did not incorporate possible delays in ending CDTi due to co-endemicity with LF. In areas where LF is co-endemic with onchocerciasis, an assessment whether both diseases have reached the thresholds to stop treatment will be needed in order to stop CDTi. In practice, no delay is expected in most cases as MDA for LF, which relies on albendazole and ivermectin, usually requires fewer cycles to reach the point of transition to the post-treatment phase. However, LF mapping or anti-LF MDA have not started in about a third of the 35 endemic countries in Africa [45].
We did not take into account the possibility of drug resistance, as no confirmed cases of ivermectin resistance have been reported from endemic countries so far. However, if ivermectin resistance were to happen as suggested by Bourguinat and colleagues through studies on the effects of ivermectin on the genetics of Onchocerca volvulus [46], the entire efforts for onchocerciasis treatment could be endangered, as current strategies heavily rely on ivermectin.
The long time horizon of 2013–2045 poses challenges in predicting technological, political, and economic changes. New treatment and diagnostic tools could be game changers in achieving elimination. Ivermectin is a microfilaricidal drug which requires many years of treatment and has a risk of eliciting severe adverse reactions in L. loa patients. Macrofilaricidal drugs that are safe and effective for general population use, are easy to administer in communities, and have a shorter treatment period than ivermectin could substantially change treatment strategies and expedite elimination. Several macrofilaricidal drugs for human use have been or currently are under development, e.g., doxycycline [47], emodepside [48], moxidectin [49], and flubendazole [50]. The need for diagnostic techniques that are capable of detecting infections early, are easy to use in the field, and are affordable would greatly facilitate surveillance when early detection of new infections is paramount. The skin snip method, currently the most common diagnostic method, has low sensitivity for detecting very light infections, and can result in a delay in detecting recrudescence. Several diagnostic techniques, e.g., OV-16 (ELISA and Rapid Test) and the DEC patch test [51,52], that may prove more sensitive and practical, have been developed. Unexpected political unrest might hamper the elimination programs, as it interrupts interventions and weakens political support. Industrialization along with economic growth may have a significant impact. For instance, the construction of dams can flood existing breeding sites of blackflies or create new ones, and deforestation can greatly alter the composition or density of blackfly populations.
Political will across the whole spectrum of stakeholders from global and national policymakers to community members will be particularly critical during the “last mile” towards elimination and subsequent eradication [53]. Countries sharing borders spanning endemic areas would need to effectively collaborate to enable prompt responses to or prevent possible recrudescence. Regular meetings have been held between Guinea/Sierra Leone/Liberia, Togo/Benin, and Benin/Nigeria [54], and this proves such mechanism can work. Similar collaborative relationships would need to be fostered for other endemic countries. APOC has announced that it would transform to a new regional entity by 2016 that would support integrated country-driven programs to eliminate onchocerciasis, LF, and other preventive chemotherapy NTDs (soil-transmitted helminthiasis, schistosomiasis, trachoma) in Africa [55,56]. Successful launching of this new regional entity might provide a more collaborative environment for sustainable interventions and post-treatment surveillance for NTDs in the region. Continuous support from community members is essential for onchocerciasis elimination in Africa. National policymakers would need to keep empowering community drug distributors, as their role is critical for successful CDTi and will continue to be so until eradication has been achieved.
|
10.1371/journal.ppat.1002852 | HSV-1 Genome Subnuclear Positioning and Associations with Host-Cell PML-NBs and Centromeres Regulate LAT Locus Transcription during Latency in Neurons | Major human pathologies are caused by nuclear replicative viruses establishing life-long latent infection in their host. During latency the genomes of these viruses are intimately interacting with the cell nucleus environment. A hallmark of herpes simplex virus type 1 (HSV-1) latency establishment is the shutdown of lytic genes expression and the concomitant induction of the latency associated (LAT) transcripts. Although the setting up and the maintenance of the latent genetic program is most likely dependent on a subtle interplay between viral and nuclear factors, this remains uninvestigated. Combining the use of in situ fluorescent-based approaches and high-resolution microscopic analysis, we show that HSV-1 genomes adopt specific nuclear patterns in sensory neurons of latently infected mice (28 days post-inoculation, d.p.i.). Latent HSV-1 genomes display two major patterns, called “Single” and “Multiple”, which associate with centromeres, and with promyelocytic leukemia nuclear bodies (PML-NBs) as viral DNA-containing PML-NBs (DCP-NBs). 3D-image reconstruction of DCP-NBs shows that PML forms a shell around viral genomes and associated Daxx and ATRX, two PML partners within PML-NBs. During latency establishment (6 d.p.i.), infected mouse TGs display, at the level of the whole TG and in individual cells, a substantial increase of PML amount consistent with the interferon-mediated antiviral role of PML. “Single” and “Multiple” patterns are reminiscent of low and high-viral genome copy-containing neurons. We show that LAT expression is significantly favored within the “Multiple” pattern, which underlines a heterogeneity of LAT expression dependent on the viral genome copy number, pattern acquisition, and association with nuclear domains. Infection of PML-knockout mice demonstrates that PML/PML-NBs are involved in virus nuclear pattern acquisition, and negatively regulate the expression of the LAT. This study demonstrates that nuclear domains including PML-NBs and centromeres are functionally involved in the control of HSV-1 latency, and represent a key level of host/virus interaction.
| After an initial lytic infection, many viruses establish a lifelong latent infection that hides them from the host immune system activity until reactivation. To understand the resurgence of the associated diseases, it is indispensable to acquire a better knowledge of the different mechanisms involved in the antiviral defense. During latency, viral genomes of nuclear-replicative viruses, such as herpes simplex virus type 1 (HSV-1), are stored in the nucleus of host cells in a non-integrated form. Latency establishment is associated with a drastic change in HSV-1 gene expression program that is maintained until reactivation occurs. The last two decades of research has revealed that the functional organization of the cell nucleus, so-called nuclear architecture, is a major factor of regulation of cellular genes expression. Nonetheless, the role of nuclear architecture on HSV-1 gene expression has been widely overlooked. Here we describe that the genome of HSV-1 selectively interacts with two major nuclear structures, the promyelocytic nuclear bodies (PMLNBs or ND10) and the centromeres. We provide evidence supporting that these nuclear domains directly influence the behavior of latent viral genomes and their transcriptional activity. Overall, this study demonstrates that nuclear architecture is a major parameter driving the highly complex HSV-1 latency process.
| Herpes simplex virus type 1 (HSV-1), a major human pathogen, is a persistent human neurotropic virus and a model of long-term interaction between a host cell and a parasite. HSV-1 establishes a long-term latent infection in neurons of the trigeminal (or Gasserian) ganglia (TG) of the peripheral nervous system, from which it reactivates periodically to replicate and spread [1]. The establishment of latency is dependent on a sequence of physiological and molecular events involving the host immune system, the cellular antiviral response, and the ability of the virus to initiate a latent gene expression program.
Latent HSV-1 dsDNA genomes localize in the nucleus of the host neuron where they remain as multi-copy chromatinized episomes, which do not integrate into the host-cell genome [2], [3]. During latency, HSV-1 lytic gene expression is strongly repressed; although some lytic transcripts could be detected at low level, by highly sensitive techniques [4]–[6]. The latency-associated transcript (LAT) locus is the only gene to be highly expressed throughout the persistent stage, from establishing latency to reactivation [7]. LAT is a noncoding RNA, synthesized as an 8.3-kb polyadenylated, unstable primary transcript, and is rapidly processed into a stable 2-kb intron lariat and several microRNAs [8]–[11]. LAT expression has been linked to several aspects of the latency process, including neuron survival, viral genome chromatin status, lytic gene expression, number of latently infected neurons, and efficiency of reactivation in animal models [2], [3], [10], [12]–[17]. Although LAT appears to regulate latency and reactivation, several studies have shown that LAT is probably expressed only in a subset of latently infected neurons, implying that latency is intrinsically a heterogeneous event [18]–[23]. The heterogeneity of HSV-1 latency has also been observed at the level of the viral genome copy number in individual neurons, which has been directly correlated with reactivation probability, suggesting that it is a functionally significant parameter [19], [20], [24]. How these variable parameters impact on the biology of the latent virus and the reactivation process remains unclear. Moreover, host-cell factors and the cellular environment can be anticipated to also account for the variability of latency and for determining the ability of HSV-1 to reactivate. Therefore, the study of latency requires experimental approaches in which the heterogeneity can be fully assessed with regard to viral genome features, viral gene expression, and host-cell nuclear components. In situ fluorescence-based strategies offer such a possibility, through a multi-parametric reading of a cell population at the single-cell level.
The mammalian cell nucleus is a highly organized compartment containing the chromosomes and several nuclear domains, which reflect the various molecular activities taking place in the nucleus. Numerous studies reported that the position of a gene within the nucleus is correlated with its transcriptional status [25], [26]. The predetermined nuclear positions of genetic loci within the nuclear architecture are key determinants of gene expression, together with transcription factors and epigenetic chromatin modifications [25], [27], [28]. Nuclear structures known to influence gene expression include the nuclear envelope, telomeres, centromeres and pericentromeres, and nuclear domains such as promyelocytic leukemia (PML) nuclear bodies (NBs, also called ND10), transcription factories, polycomb group complexes, and the nucleolus [26], [29]–[34]. Among these nuclear domains, PML-NBs are proteinaceous structures that reorganize in response to various cellular stressors [33], [35], [36]. PML-NBs provide a nuclear environment that can be associated with transcription of cellular genes [32], [37], [38]. However, PML-NBs contain repressor proteins such as HP1, ATRX, and hDaxx [39], [40], which have an inhibitory effect on transcription and replication of RNA and DNA viruses, supporting the silencing activity of PML-NBs [41], [42]. In cultured infected cells, the association of PML-NB with genomes of several viruses, including HSV-1, has led to the hypothesis that PML-NBs may operate as a nuclear relay for innate host-cell defense mechanisms, blocking replicative infection by creating an environment unfavorable for viral gene expression [32], [42]–[50]. However, how nuclear domains impact in vivo on the biology of persistent viruses such as HSV-1 and whether they may intervene in the latency process, in particular in the acquisition of essential parameters involved in latency maintenance and reactivation, is currently unknown.
In this study, we took advantage of a physiologically well-characterized mouse model of HSV-1 infection, to develop an efficient fluorescent in situ hybridization (FISH) approach for detecting HSV-1 genomes during latency in neurons from infected mouse TG. Using a high-resolution visualization technique, we described the intra-nuclear distribution of the latent HSV-1 genome in neurons, and correlated HSV-1 patterns with LAT expression. We found that HSV-1 genomes were non-randomly associated with two nuclear domains, PML-NBs and centromeres. Using infected PML knockout (KO) mice, we showed that PML/PML-NBs influence viral genome distribution and negatively regulate expression of LAT. Finally, we demonstrated that HSV-1 genomes associated with PML-NBs or centromeres were negative for the expression of LAT.
The lack of an efficient in situ detection method of viral genomes has been a major technical limitation to the study of herpes virus infection and disease both in animal models and human samples. Detection of HSV-1 genomes by FISH in latently infected mouse tissues has remained unsuccessful despite attempts of many groups [51]. To determine the intra-nuclear organization of the multiple copies of HSV-1 and its influence HSV-1 gene expression, we developed a DNA-FISH protocol and applied it to an established lip-inoculation mouse model in which HSV-1 establishes significant latency in the TG (Figure 1A; [52]). Infected and mock-infected mice were sacrificed at 28 d.p.i., a time point at which HSV-1 latency is known to be fully established [51], and the TGs were cryo-sectioned. Our FISH protocol efficiently detected latent HSV-1 genomes in mouse neuronal tissues (Figure 1D; see Materials and Methods for details). The DNA-FISH probes recognized a 90 kb region of the viral genome, excluding the LAT locus (named hereafter “HSV-1 genome probes,” Figure 1B). Importantly, our protocol did not include a signal amplification procedure and thus is well suited for the study of intra-nuclear organization by high-resolution microscopy. Signal specificity was assessed through several control experiments, including FISH analysis of mock-infected mice, FISH with control probes without HSV-1 sequences (Figure 1C), and a comparison between our probe and a commercially available probe (Figure S1).
In TG sections from infected mice sacrificed at 28 d.p.i., the FISH signal for HSV-1 DNA was observed only in the nuclei of neurons, where it appeared as a dotted pattern comprising spots of various numbers, sizes, and intensities (Figure 1D and S1). The presence of the HSV-1 genome in neurons and not in satellite cells is consistent with the results of previous in situ PCR and single-cell PCR studies [20], [24]. Two main intra-nuclear patterns were observed: a single, bright, round spot (termed “single”) and numerous spots of non-uniform size and shape (termed “multiple”) (Figure 1D–E). Both patterns were observed in different inoculation models (lip, eye, and whisker pads), animals (mice of different inbred strains and rabbits), and viral strains (SC16, 17syn+, KOS/M, and McKrae), indicating that they are characteristic of neurons latently infected with HSV-1. The occurrence of each pattern was variable (Figure 1E), suggesting that the strain of virus and or the route of infection may affect how the viral genome accumulates in neurons. In addition to these two primary patterns, a single spot accompanied by one or two smaller spots (termed “single+”) and a multiple pattern that filled the nucleus (termed “super-multiple”) were less frequently observed (Figure S1). The presence of multiple genome spots in a large proportion of infected neurons is consistent with the results of earlier single-cell quantitative PCR (qPCR) analyzes [19], [20]. We further confirmed that the sizes of the spots observed by FISH were consistent with the presence of several copies per spot (Table 1). The spot in the single pattern was 0.80±0.14 µm wide (n = 48), a size similar to that of in vitro-induced quiescent genomes [50], which were estimated by qPCR to contain four to five copies of the genome. The spots of the multiple pattern varied from 0.40 to 3 µm in diameter (more in the case of large aggregates). We measured the sizes of individual FISH-detected HSV-1 genomes in in vitro-infected cells to define a reference. Single-copy parental genomes entering the nucleus appeared as spots that were 0.51±0.08 µm wide (n = 76), which was similar to the width of isolated spots within the multiple pattern (0.44±0.07 µm; n = 51), indicating that these spots may represent single genomes. Based on this analysis, it can be predicted that the single-spot pattern contains more than one copy of the genome and that in the multiple-spot pattern, the genome can be either isolated or aggregated.
We next used serial sectioning to explore whether HSV-1 established latency with a topographical preference within the TG. Neurons shown by FISH to be positive for HSV-1 were distributed all along the TG (one of the three sections analyzed), without any enrichment along the antero-posterior axis (Figure 1F). Similarly, the frequencies of the single and multiple patterns were equivalent throughout the TG (one mouse is shown as an example in Figure 1G). The frequency varied from section to section, and no reproducible pattern could be detected in a group of six mice.
Overall, these results show that the HSV-1 latent genome in mouse neuronal tissues can be detected by FISH, with sufficient efficiency and quality for the analysis of intra-nuclear distribution. During latency in neuron nuclei, the HSV-1 genome is present as multiple copies, as shown previously using other methods [15], and adopts a non-random intra-nuclear organization.
Data from several groups suggest that LAT is expressed in a fraction of infected neurons during latency. We set up a dual RNA/DNA-FISH assay based on tyramide signal amplification (TSA) technology to co-detect LAT transcripts and HSV-1 genomes (Figure 1B and 2A). We challenged the sensitivity of our RNA FISH method by using up to 20 times the amount of probe (1000 ng/assay instead of 50 ng) and increasing the TSA time. We failed to detect neurons with weak LAT signals, indicating that our test efficiently detected LAT-expressing neurons. In mice at 28 d.p.i., 18 to 31% of the HSV-1 DNA-containing neurons were positive for the 2-kb LAT RNA, thus confirming the results previously obtained by different approaches [18], [20]. Notably, fewer than 10% of neurons with a single pattern were positive for 2-kb LAT (Figure 2B). In contrast, 40.3±9.5% of the multiple-pattern neurons expressed LAT, suggesting that the multiple pattern reflects conditions favorable for LAT transcription (Figure 2B). A reciprocal analysis showed that 83.0% of LAT-positive neurons contained the HSV-1 genome in a multiple pattern (Figure 2C). This suggests that the organization of the HSV-1 genome in a multiple pattern is necessary, but not sufficient, to support LAT transcription. These data demonstrate that transcription of the LAT locus is linked to the intra-nuclear pattern of the viral genome.
The distribution of latent HSV-1 in neuron nuclei did not bear any resemblance to the patterns of known nuclear domains, and it remained unclear whether the viral genome associated with particular structures. As HSV-1 gene expression has been shown to involve viral chromatin [2], [3], we first focused on nuclear structures that are known to control cellular gene transcription through heterochromatin domains: the nuclear envelope, telomeres, centromeres, and pericentromeres. HSV-1 latent genomes were rarely found at the periphery of the nucleus, thus excluding a preferential association with the nuclear envelope. The association of the HSV-1 genome with telomeres and centromeres was assessed by dual-color DNA-FISH. No co-localization of the HSV-1 genome with telomeres was observed when assessing single or multiple patterns (Figure 3A). In mouse cells, the centromeres are positioned at the surface of pericentromeric aggregates (also called chromocenters), which are commonly detected by Hoechst staining [53] (Figure 3A and S2). By dual DNA-FISH, performed using specific probes to detect minor and major satellites, we confirmed that the organization observed in the cultured cells was similar to that in TG neurons, and that heterochromatic aggregates detected by DNA staining represent pericentromeres (Figure S2). Latent HSV-1 co-localized with centromeric repeats in 10.1±0.8% of the single-pattern neurons and in 39.4±8.8% of the multiple-pattern neurons (Figure 3B). In contrast, the frequency of association with pericentromeres remained low for both single- and multiple-pattern neurons (4.51±3.4% and 8.43±4.9%, respectively, n = 1,249 neurons in two mice). Within individual neuron nuclei, only a subset of HSV-1 FISH spots was associated with centromeres, showing that they are not the only residence sites of latent genomes. An immuno-FISH staining of the centromeric protein (CENP)-A, which is essential for the stability and functionality of the centromere [54] further confirmed the localization of a subset of HSV-1 genome onto the centromeres (Figure 3C–D). Additionally, these data demonstrate that the HSV-1-associated centromeric loci are likely to be functional centromeres. The association of HSV-1 genomes with centromeres did not appear to be an artifact caused by a strong HSV-1 signal in the multiple pattern for the following reasons: (i) the centromere and HSV-1 signals were largely co-localized (Figure 3A and 3D, bottom right image); (ii) the positioning of HSV-1 genomes adjacent to pericentromeres did not increase concomitantly with an increase in HSV-1 signal density (Figure 3B and S2); (iii) and co-detection of HSV-1 and telomeres did not result in signal co-localization, even though each cell contained twice as many telomeres as centromeres and telomeres are proximal to centromeres in acrocentric mouse chromosomes.
Because the single HSV-1 pattern did not frequently coincide with centromeres, and because a tight interplay between HSV-1 and PML-NBs exists in vitro, we developed an immuno-FISH approach to analyze whether PML-NBs could be involved in HSV-1 latency. In non-infected tissues, PML was detected by immuno-FISH in both non-neuronal cells and neurons. Neurons contained 1–10 PML spots, although a subpopulation of neurons did not display any detectable signal in the nucleus (Figure 4A). In mice at 28 d.p.i., a qualitative assessment of the number of PML-NBs in infected neurons did not reveal any obvious change, by comparison with uninfected neurons. In latently infected mice, PML protein invariably associated with single-pattern HSV-1 genomes, whereas it associated with HSV-1 genomes in only 61% of multiple-pattern neurons (n = 201 neurons). In the latter case, only some HSV-1 spots were associated with PML, revealing heterogeneity among the genomes regarding their association with PML.
To determine whether HSV-1 genomes associated with bona fide PML-NBs, immuno-FISH and 3D microscopy were used to detect two stable signature components of PML-NBs, ATRX and Daxx. Both ATRX and Daxx were found to be associated with single-pattern HSV-1 genomes (Figure 4B); in multiple-pattern genomes, ATRX and Daxx co-localized with at least one HSV-1 genome focus, consistent with the observed frequency of the association of these genomes with PML. A triple-labeling experiment confirmed that PML and Daxx (Figure 4C) or PML and ATRX (not shown) simultaneously associated with the HSV-1 genome. Careful inspection of PML-NBs associated with HSV-1 revealed that PML protein had a ring-like shape, with HSV-1 genome in its center (Figure 4D). The presence of HSV-1 DNA within PML-NBs was intriguing because PML-NBs have been generally found to be devoid of nucleic acids and to be localized adjacent to or within 2 µm of genomic loci [37], [55]. High-resolution 3D confocal microscopy confirmed that in the case of the HSV-1 latent genome, the DNA was clearly inside the PML ring (Figure 4D), and that PML was wrapped around the viral genome. This organization was also observed during the early phase of mouse infection (Figure 5C) and in vitro in cells infected with replication-defective HSV-1 [50]. Thus, our observations show that PML assembles around HSV-1 genomic DNA, forming an atypical DNA-containing PML-NB (DCP-NB).
PML-NB reorganization and co-localization with HSV-1 genomes were observed as early events in lytic infection in cultured cells [43], [45], [56], [57], raising the possibility that the association we observed during latency could be initiated early during the establishment of latency (the acute phase). In immuno-FISH analyzes performed on sections from mice sacrificed at 6 d.p.i., we observed that the PML protein signal within PML-NBs was stronger, and the PML-NBs were generally larger and more numerous in acute-phase tissues compared with latently infected and non-infected tissues (Figure 5A). Increases in the PML signal were observed in both neurons and accessory cells and, importantly, were restricted to infected TG (Figure 5A–B and S3; see Materials and methods for details), demonstrating that the increase in the PML signal resulted from the on-going infection. The increase in the PML signal could be attributable to the recruitment of nucleoplasmic PML (which accounts for 90% of nuclear PML; [36] into PML-NBs, or to an increase in the overall amount of PML. Western blotting of whole TG from mice at 6 d.p.i., showed increases in total PML and PML isoform levels in acutely infected TG compared with non-infected TG (see Materials and methods) and TG from non-infected mice (Figure 5B), demonstrating that the change in PML protein pattern results from an increase in total cellular PML protein and not only from a more efficient recruitment of nucleoplasmic PML into PML-NBs. These data support the stimulation of PML expression during the acute phase of HSV-1 infection, probably as a result of IFN pathway activation [41]. PML and HSV-1 formed 1–12 DCP-NBs per infected neuron nucleus, and most of them also contained ATRX, and Daxx (Figure 5C). Thus HSV-1 genome and PML patterns were significantly different from those observed in latently infected neurons. We conclude that acute infection provokes a PML response, leading to the formation of HSV-1 DCP-NBs, and that the association between PML and the viral genome is initiated during the very early stages of the latency process.
The above observations raised the possibility that the HSV-1/PML interaction may play a role in the formation of the latent HSV-1 patterns. To address this, we quantified HSV-1 latent genome patterns in latently infected PML-deficient mice. Both PML+/− and PML−/− mice displayed a significant decrease in the number of single-pattern neurons and a concomitant increase in the number of neurons with the super-multiple pattern (Figure 5D–E). These data show that PML protein and/or PML-NBs influence the intra-nuclear pattern adopted by the viral genome within latently infected neurons. Additionally, based on the number of viral genome foci detected within individual neurons, we conclude that in absence of PML/PML-NBs, the number of genome copy in latent TGs is higher, suggesting that PML/PML-NBs play a role in limiting the number of viral genomes that establish latency.
The above observations establish strong links between LAT expression, HSV-1 intra-nuclear distribution, and the association of the HSV-1 genome with PML-NBs and centromeres. This raised the possibility that the association of the HSV-1 genome with PML-NBs and centromeres may regulate LAT transcription. In support of this hypothesis, in single-pattern neurons, HSV-1 is systematically associated with PML-NBs, and LAT RNA is rarely present. To test whether association of genomes with PML NBs in multiple pattern cells had any effect on LAT expression in those cells, we performed a triple labeling experiment to simultaneously detect the HSV-1 genome, 2-kb LAT RNA, and PML/centromeres. Preliminary observations suggested that the 2-kb LAT signal was not correlated with the association of the HSV-1 genome with PML-NBs or centromeres. To confirm this, we traced LAT expression and the association of HSV-1 with PML and centromeres for each neuron across the entire TG of one mouse. The data clearly showed that there was no correlation (Table 2). This suggests that within a nucleus containing multiple copies of the HSV-1 genome, the association of some of the HSV-1 genomes with PML-NBs or centromeres does not have a dominant negative effect on the expression of LAT from the other copies of the viral genome.
An individual neuron contains a heterogeneous population of HSV-1 genomes (“free” or associated with a nuclear domain). Thus, the transcriptional status of these genomes may also be heterogeneous. To explore this possibility, we utilized the primary (nascent) 8.3-kb LAT transcript (Figure 1B) as a marker of the site of active transcription, in order to identify genomes that were being transcribed. Figure 6A illustrates the co-detection of HSV-1 genomes (red), the nascent 8.3-kb LAT transcript (blue), and the stable 2-kb LAT RNA (as a control). The nascent LAT RNA appeared as a set of large dots (1 to 7 per nucleus), each dot being associated with at least one HSV-1 genome spot (Figure 6). Such dotted pattern has been previously observed by ISH using peroxidase and alkaline phosphatase staining [58]. This suggests that in a single neuron, LAT can be transcribed from several copies of the HSV-1 genome. Notably, these neurons also contained several HSV-1 spots that were not associated with any 8.3-kb LAT RNA signal. Although we cannot exclude the possibility that these genomes are transcribed at a level below the sensitivity of our FISH method, the data suggest that they are not transcribed. Overall, these results show that only a fraction of the HSV-1 genomes within a single infected neuron are significantly transcribed and that the transcriptional status of HSV-1 genomes is highly heterogeneous in individual neurons.
We next analyzed whether LAT is actively transcribed from PML-NB-associated latent HSV-1 genomes. In mouse TG sections, we detected the HSV-1 genome, its associated nascent LAT RNA product, and PML-NBs by a triple-labeling approach. In the LAT-expressing neurons, the nascent 8.3-kb LAT RNA was never associated with viral genomes that co-localized with PML-NBs (Figure 6B, bottom panel). We paid particular attention to 8.3-kb LAT-positive neurons with the single and single+ patterns, and observed that the LAT positive neurons were all neurons with a single+ pattern, and that the genome that was transcribed was the one not associated with PML. The larger HSV-1 genome spot surrounded by PML protein was never associated with an 8.3-kb LAT RNA spot (Figure 6B, top and middle panels). These data further support the idea that PML-NBs repress transcription of the associated HSV-1 genome. However, in the 8.3-kb LAT-positive neurons with the multiple pattern, several non-transcribed HSV-1 genomes were not associated with PML-NBs. It is likely that 8.3-kb LAT is not transcribed from many of the genomes and that factors other than PML-NBs also regulate LAT transcription.
We extended the analysis to centromere-associated HSV-1 genomes. Similarly to the findings for PML-NBs, centromere-associated viral genomes were never adjacent to the nascent 8.3-kb LAT RNA, suggesting that centromeres may also inhibit LAT transcription (Figure 6C).
If PML were to inhibit LAT transcription from the HSV-1 genome, one would expect to see an increase in LAT expression in a PML-deficient background. The 2-kb LAT RNA-FISH analysis of latently infected PML+/− and PML−/− mice revealed that the percentage of 2-kb LAT-positive neurons was higher in PML−/− mice compared with wild-type and heterozygous mice (Figure 6D). The increase in the LAT-positive neuron percentage was not simply related to greater numbers of neurons with multiple/super-multiple pattern. Indeed, within this category of neurons, LAT expression was twofold higher in PML−/− animals (Figure 6E). These data demonstrate that PML/PML-NBs play a role in the regulation of LAT expression and support their transcriptional repressor activity.
Here, we report the structural and functional interactions of the genomes of a persistent virus, HSV-1, with the host-cell nuclear environment. Our data reveal two new features of the viral genomes that characterize the latency state. First, the intra-nuclear distribution of the latent genome is not random and correlates with viral gene expression, and second, the host-cell nuclear domains play a role in viral genome pattern acquisition and in the control of viral gene expression. Thus, the interaction between the viral genomes and host-cell nuclear components represents a new level of host–virus interaction, which is likely to participate in the process of latency and reactivation.
The ability to explore the cell host–virus interaction is of outmost importance in our understanding of persistent viral infections because regulation of the latent HSV-1 genome relies mainly on cellular components. A substantial benefit of the in situ immuno-DNA/RNA-FISH developed in this study lies in the simultaneous detection of the viral genome, virally encoded transcripts, and cellular components in the same cell. This will enable us to address important issues of cell host–virus interactions in tissues obtained from physiologically infected animal models but also in emerging in vitro HSV-1 latency models [59]. The FISH approach provides high-resolution individual cell data without sacrificing the access to a global view of the virus and of the host-cell population. This appears as a major advantage given that HSV-1 latency is highly heterogeneous. FISH and immuno-FISH will be essential assets to study latency and will complement the currently used biochemical and molecular approaches.
We confirmed that LAT was expressed in a fraction of infected neurons and that viral copy numbers varied among neurons. Based on the HSV-1 genome pattern and our estimate of genome copy number per FISH spot, the single and multiple patterns likely represent the low-copy and high-copy virus genome-containing neurons, respectively, identified by contextual analysis [19]. Additionally, we found that the HSV-1 latent genomes were heterogeneously distributed within neuron nuclei and preferentially associated with PML-NBs and centromeres. LAT expression is positively correlated with HSV-1 genome pattern and negatively correlated with its association with PML-NBs and centromeres, demonstrating that the intra-nuclear distribution of HSV-1 genomes is a major feature of the latency process. LAT detection was almost exclusively associated with the multiple genome pattern, demonstrating that LAT expression is restricted to neurons with high viral genome copy numbers. Single-cell contextual analysis has also revealed that a high genome copy number per neuron is associated with a higher probability of virus reactivation [14], suggesting that this parameter may be a key aspect of latent genome status. Importantly, the various copies of HSV-1 within a single nucleus are not transcriptionally equivalent, with LAT being transcribed only from a subset of genomes. This suggests that latent HSV-1 genomes are comparable to, and behave like, multi-allelic cellular genes, raising the possibility that only a subset of these genomes are susceptible to sustain full reactivation (i.e., to reach expression of lytic genes). The dotted pattern observed with the 8.3 kb LAT probe was previously reported [58], and was proposed to be sites of early processing of the LAT transcript. Our data confirm this hypothesis by demonstrating that these clouds of LAT primary RNA are associated with HSV-1 genomes.
PML-NBs are probably the most thoroughly studied nuclear domains in the context of virus infection for their involvement in the innate antiviral response and in the interferon (IFN) response pathway. Our data from the acute phase and from PML−/− mice support a role for PML-NBs in limiting the extent of viral replication during acute-phase, and thus the number of HSV-1 genomes that establish latency in each neuron. These data are consistent with the known role of PML-NBs, through the activity of several of their major components such as PML, Sp100, Daxx, ATRX, and small ubiquitin-like modifier (SUMO) protein, as repressors of HSV-1 onset of lytic infection in cultured cells [57], [60]–[63]. We provide a clear demonstration that PML expression increased in vivo in acutely infected mouse TG, and that the PML protein, through the formation of HSV-1-containing PML-NBs, associated with the HSV-1 genome during the early phase of latency. Additionally, we showed that the HSV-1 genomes remain associated with PML-NBs during latency in over 80% of infected neurons, suggesting that PML-NBs play an antiviral role influencing latency and probably reactivation.
PML-NBs have been proposed to create a specific local nuclear environment by concentrating proteins and hosting biochemical reactions within the PML shell. PML-NBs reorganize in response to various stressors, potentially to relocate their activity at selected nuclear sites [64]. The reorganization of the PML-NBs resulted in the formation of new PML-NBs around the HSV-1 genome at early stages of latency, thus altering the immediate nuclear environment of the incoming viral genome. Importantly, we showed that PML-NBs remain associated with viral genomes long after replicative infection has ceased, indicating that maintenance of this particular type of DNA-containing PML-NB (DCP-NB) requires neither on-going viral replication nor the associated antiviral and IFN signaling pathways. The pattern of both HSV-1 genome and PML-NBs are dramatically different between acute phase and latency, indicating that a profound remodeling of these patterns takes place during establishment of latency. Ongoing studies will provide pattern analysis at intermediate time between 6 d.p.i. and 28 d.p.i. The formation of DCP-NBs can be seen as a response to the presence of chromatinized foreign DNA [42], or more broadly, the presence of pathology-associated abnormal chromatin [65], [66]. Moreover, PML-NBs repress the synthesis of the LAT primary transcript through their association with HSV-1 genomes, from which microRNAs are produced [10]. The atypical assembly of a PML-NB around a genetic locus may thus be considered a distinct form of PML-NB controlling the expression of noncoding RNA in pathological situations. Indeed, PML-NBs assemble around pericentromeric satellite sequences and telomeres, two cellular loci known to give rise to noncoding RNA [65], [66].
We showed that HSV-1 genomes are also associated with host neuron centromeres during latency. Bishop and colleagues previously showed that foreign DNA delivered by polyomavirus-like particles was localized to centromeres [42]. Our data from a biologically relevant context and an in vivo model support the idea that centromeres represent docking sites for virus genomes. Centromeres and the adjacent pericentromeres are among the best-characterized nuclear domains that silence nearby genes [29]–[31]. Consistent with this, we showed that centromere-associated HSV-1 genomes did not express LAT RNA. We want to emphasize that the association with HSV-1 occurs at the centromere itself, which distinguishes the current set of data from most other published data related to associations of cellular genes with pericentromeres [67]–[75]. Only a subset of HSV-1 genomes within a nucleus is found associated with centromeres, indicating that this association is not the main mechanism repressing transcription of latent genomes. Of note, both PML-NBs and centromeres (because of their proximity with pericentromeres) are enriched in ATRX and Daxx. In addition, hDaxx has been shown to co-localize with centromeres in human cells [76]. This raises the possibility that both nuclear domains exert their repressive effect on HSV-1 transcription through common factors [62].
Interestingly, HSV-1 has developed strong “anti-centromere” activity through the combined activities of the viral E3 ubiquitin ligase ICP0 protein [77] and the proteasome. In cultured cells, ICP0 induces the degradation of at least 10 CENPs, which results in the alteration of centromeric chromatin and destabilization of the centromeres [78]–[81] (S. Gross and P. Lomonte, personal communication). The biology of HSV-1 does not favor ICP0-induced centromere destabilization, prompting the mitotic arrest of infected cells [82], [83]. Indeed, HSV-1 is able to replicate independently of the cell cycle [84], and the lytic cycle does not depend on cell arrest at the mitotic phase. This suggests that HSV-1 targets centromeres not to control their effect on chromosome segregation, but rather to control an activity more relevant of differentiated, non-dividing cells. On the other hand, it is suspected that ICP0, which does not bind DNA and is not a transcription factor per se [85], inhibits the activities of numerous repressive nuclear factors in order to favorably modify the nuclear environment to stimulate the virus replicative cycle, at both the onset of a new infection and during the course of reactivation [60], [86]–[93]. ICP0 is known to be essential for full reactivation of HSV-1 in latently infected quiescent cells [94]–[96]. We therefore propose that centromeres, although they seem to act as repressors during latency, may offer a favorable nuclear environment for transcriptional events during reactivation, providing their protein composition and structure are modified by ICP0. This agrees with data showing that centromere/pericentromere regions are sites of intense transcriptional activity following the exposure of cells to a variety of stressors such as heat shock, UV, and heavy metals, which potentially induce HSV-1 reactivation [97]–[99]. On-going work should provide evidence to support this hypothesis.
For animal experiments performed in France: all procedures involving experimental animals conformed to ethical issues from the Association for Research in Vision and Ophthalmology (ARVO) Statement for the use of animals in research, and were approved by the local Ethical Committee of UPR-3296-CNRS, in accordance with European Community Council Directive 86/609/EEC. All animals received unlimited access to food and water.
For animal experiments performed in the USA: animals were housed in American Association for Laboratory Animal Care-approved housing with unlimited access to food and water. All procedures involving animals were approved by the Children's Hospital Animal Care and Use Committee and were in compliance with the Guide for the Care and Use of Laboratory Animals.
Wild-type HSV-1 strains SC16 and 17syn+ were used. Stocks were generated in rabbit skin cell monolayers, and viral titers were determined as described previously [100]. Briefly, six-week-old inbred female BALB/c mice (Janvier Breeding, Le Genest Saint Ile, France), were inoculated with 106 PFU of the SC16 virus, injected into the upper-left lip of the mice. Mice were observed daily for clinical signs of ocular infection from 0 to 28 d.p.i. The sided inoculation of the lip results in an asymmetrical infection, which is characterized by an extremely low load of virus on the right TG compared to the left TG. Thus, data presented in this study were collected on the left TG, except in Figure 3A and 3D, as indicated [4]. Data presented in this study were collected from the left TG, except for those shown in Figure 3A and 3D. For the 17syn+/eye model, inoculation was performed as described previously [101]. Briefly, prior to inoculation, mice were anesthetized by intra-peritoneal injection of sodium pentobarbital (50 mg/kg of body weight). A 10 µL drop of inoculum containing 105 PFU of 17syn+ was placed onto each scarified corneal surface. This procedure results in ∼80% mice survival and 100% infected TG. PML wild-type, heterozygous, and knockout mice were obtained from the NCI Mouse Repository (NIH, http://mouse.ncifcrf.gov; strain, 129/Sv-Pmltm1Ppp) [102]. Genotypes were confirmed by PCR, according to the NCI Mouse Repository guidelines.
Primers:
Frozen sections of mouse TG were performed as previously described [100]. Mice were anesthetized at 6 or 28 d.p.i., and before tissue dissection, mice were perfused intra-cardially with a solution of 4% formaldehyde, 20% sucrose in 1× PBS. The whole head, or individual TG were prepared as previously described, and 10 µm frontal sections were collected in three parallel series, and stored at −80°C.
DNA-FISH probes were Cy3 labeled by nick-translation as described previously [103]. Briefly, cosmids 14, 28 and 56 [104] comprising a total of ∼90 kb of HSV-1 genome (see Figure 1A) were labeled by Nick translation (Roche Diagnostic) with dCTP-Cy3 (GE Healthcare), and stored in 100% formamide (Sigma-Aldrich). The DNA-FISH procedure was adapted from Solovei et al. [105].
Frozen sections stored at −80°C were thawed, rehydrated in 1× PBS and permeabilized in 0,5% Triton X-100. Heat based unmasking was performed in 100 mM citrate buffer, and sections were post-fixed using a standard methanol/acetic acid procedure, and dried for 10 min at RT. DNA denaturation of section and probe was performed for 5 min at 80°C, and hybridization was carried out overnight at 37°C. Hybridization mix contained 30 ng of each probe in 10% dextran, 1× denhardt, 2XSSC, 50% formamide. Sections were washed 3×10 min in 2XSSC and 3×10 min in 0.2XSSC at 37°C, and nuclei were stained with Hoechst 33258 or ToPro3 (Invitrogen). All sections were mounted under coverslip using Vectashield mounting medium (Vector Laboratories) and stored at +4°C until observation.
Frozen sections were treated as described for DNA-FISH up to the antigen-unmasking step. Tissues were then incubated for 24 h with the primary antibody (diluted at 1/100). After three washes, secondary antibody (1/200) was applied for 1 h. The secondary antibodies (Invitrogen) were either AlexaFluor-conjugated (PML, CENP-A), or HRP conjugated (ATRX and Daxx), which were subsequently detected by enzymatic amplification according to manufacturer's guideline (TSA, Invitrogen). Following immunostaining, the tissues were post-fixed in 1% PFA, and DNA-FISH was carried out from the methanol/acetic acid step onward.
RNA-FISH probe labeling and RNA-FISH procedures were performed as described previously [4]. Biotinylated single-strand RNA probes were prepared by in vitro transcription (Ambion) using plasmids pSLAT-2, pSLAT-4 and pSLAT-6 as template (see Figure 1) (Kind gift of S. Efstathiou, University of Cambridge, UK). Frozen sections were treated as described for DNA-FISH up to the antigen-unmasking step using solutions containing 2 mM the RNAse inhibitor Ribonucleoside vanadyl complex. The sections were pre-hybridized in 50% formamide/2× SSC and hybridized overnight at 65°C with 50 ng of RNA probe a 50% formamide buffer. Sections were washed in50% formamide/2× SSC at 65°C, and in 2× SSC at room temperature. Detection was performed using streptavidin-HRP conjugate, followed by TSA amplification (Invitrogen) with an AlexaFluor 350 conjugated substrate, according to the manufacturer's guidelines. The DNA-FISH procedure was performed starting from the methanol/acetic acid post-fixation step.
The following primary antibodies were used: anti-mouse PML (mAb3739, Millipore), anti-mouse CENP-A (rabbit mAb C51A7, Cell Signaling Technologies), anti-ATRX H-300 (Santa Cruz Biotechnology), anti-Daxx M-112 (Santa Cruz Biotechnology), anti-pan-HSV-1 (LSBio), and anti-mouse actin (Sigma-Aldrich). All secondary antibodies were Alexa Fluor-coupled and were raised in goats (Invitrogen). HRP-coupled secondary antibodies were provided with the TSA kit (Invitrogen).
Observations and most image collections were performed using an inverted Cell Observer microscope (Zeiss) with a Plan-Apochromat ×100 N.A. 1.4 objective and a CoolSnap HQ2 camera from Molecular Dynamics (Ropper Scientific). When indicated, images were collected on a Zeiss LSM 510 confocal microscope using a Plan-Apochromat ×63 N.A. 1.4 objective, except for those shown in Figure 2E, which were collected using a Zeiss LSM 780 microscope. Line scans, 3D projections, and surface rendering were performed using AIM and Zen software (Zeiss).
TGs were collected at 6 or 28 d.p.i. and snap-frozen. Frozen tissues were ground, thawed in lysis buffer (10 mM Tris-EDTA, pH 8.0) containing a protease inhibitor cocktail, and briefly sonicated. Protein extracts were homogenized using QiaShredders (Qiagen). Protein concentration was estimated by the Bradford method. Extracted proteins were analyzed by Western blotting using anti-mouse PML antibody (mAb3739, Millipore) [106].
|
10.1371/journal.pgen.1000097 | Mitochondrial Morphogenesis, Dendrite Development, and Synapse Formation in Cerebellum Require both Bcl-w and the Glutamate Receptor δ2 | Bcl-w belongs to the prosurvival group of the Bcl-2 family, while the glutamate receptor δ2 (Grid2) is an excitatory receptor that is specifically expressed in Purkinje cells, and required for Purkinje cell synapse formation. A recently published result as well as our own findings have shown that Bcl-w can physically interact with an autophagy protein, Beclin1, which in turn has been shown previously to form a protein complex with the intracellular domain of Grid2 and an adaptor protein, nPIST. This suggests that Bcl-w and Grid2 might interact genetically to regulate mitochondria, autophagy, and neuronal function. In this study, we investigated this genetic interaction of Bcl-w and Grid2 through analysis of single and double mutant mice of these two proteins using a combination of histological and behavior tests. It was found that Bcl-w does not control the cell number in mouse brain, but promotes what is likely to be the mitochondrial fission in Purkinje cell dendrites, and is required for synapse formation and motor learning in cerebellum, and that Grid2 has similar phenotypes. Mice carrying the double mutations of these two genes had synergistic effects including extremely long mitochondria in Purkinje cell dendrites, and strongly aberrant Purkinje cell dendrites, spines, and synapses, and severely ataxic behavior. Bcl-w and Grid2 mutations were not found to influence the basal autophagy that is required for Purkinje cell survival, thus resulting in these phenotypes. Our results demonstrate that Bcl-w and Grid2 are two critical proteins acting in distinct pathways to regulate mitochondrial morphogenesis and control Purkinje cell dendrite development and synapse formation. We propose that the mitochondrial fission occurring during neuronal growth might be critically important for dendrite development and synapse formation, and that it can be regulated coordinately by multiple pathways including Bcl-2 and glutamate receptor family members.
| A neuron cell is composed of cell body, axons, and dendrites. Dendritic spines on dendrites form synapses with axons of other neurons, establishing communication between neuron cells. Dendrite development and synapse formation are therefore important for neuronal function. Although many genes have been previously identified as affecting the development of dendrites and synapses, the apoptosis Bcl-2 family members have not yet been shown to regulate these processes. In this study, a Bcl-2 family survival member, Bcl-w, was found not to affect cell death, but to be required for synapse formation and motor learning in mouse cerebellum. Bcl-w also appears to control dendrite development as double null mutant mice of Bcl-w and the glutamate receptor δ2 (Grid2) have severe defects in Purkinje cell dendrites, spines, and synapses. In addition, Bcl-w and Grid2 act synergistically to promote what is likely to be mitochondrial fission in Purkinje cells. Neither the survival members of the Bcl-2 family nor the excitatory receptors have been demonstrated previously to regulate mitochondrial morphogenesis in brain. We conclude that neuronal dendrite development and synapse formation require perhaps mitochondrial fission that can be controlled by two critical pathways including Bcl-w and Grid2.
| Mitochondria have been shown to undergo morphological changes in many neurodegenerative and psychiatric diseases, suggesting their vital role in maintaining the normal function of neuron cells. One of the morphological changes in mitochondria is the length or size, which can be controlled by mitochondrial growth or mitochondrial fission/fusion cycles. Mitochondria are dynamic organelles that can undergo fission, fusion, branching, and change in subcellular distribution [1]–[3], resulting in the exchange of their genetic materials, alteration of their shape, and increase or decrease of their number [1]–[3]. This dynamic nature of mitochondria is also critically important for energy generation, calcium buffering, and control of apoptosis. Mitochondrial fission and fusion is normally a well-balanced event; when the fission is blocked, the length of mitochondria increases due to ongoing fusion, and mitochondrial fission sites persist as constriction sites due to the slowdown of fission, whiles when the fusion is inhibited, mitochondria usually appear fragmented [3]. Mitochondrial number increases during cellular division, growth, and differentiation via the fission process [4]. However, excessive fission can stimulate apoptosis [5], and cause neurodegenerative diseases [6]. In cultured healthy neurons, mitochondrial fission and fusion proteins have been shown to regulate the morphology and plasticity of dendritic spines and synapses [7]. In addition, glutamate [8] and synaptic activity [7] modulates the motility and fusion/fission balance of mitochondria and controls mitochondrial distribution in dendrites [7]. Several proteins have been identified in a variety of species to mediate mitochondrial fission or fusion process [2],[3], however, little is known about the signaling molecules that activate these processes.
Cerebellar Purkinje cells are characterized by large and highly branched dendritic arbors in the brain. Over 90% of Purkinje cell dendritic spines form excitatory synapses with granule cell parallel fiber axons, which relay information from pre-cerebellar nuclei to Purkinje cells. Grid2 is strongly expressed in Purkinje cells [9], and localizes specifically to Purkinje cell/ parallel fiber synapses [10],[11]. Analysis of Grid2 knockout mice [12], and Hotfoot mice carrying spontaneous loss-of-function mutations in Grid2 [13],[14] has demonstrated that these mice exhibit an impaired function on motor coordination and learning tasks, and have structural and functional defects in Purkinje cell/granule cell parallel fiber synapses and altered long term depression [12],[15],[16]. Physiologic studies of Grid2Lc, the Lurcher dominant mutation have established that the Grid2Lc mutation results in inward Ca2+/Na+ current and constitutive activation of the δ2 glutamate receptor, and also that the Grid2Lc receptor has similar channel properties to both NMDA [17] and AMPA receptors [18],[19]. Activation of Grid2Lc also induces autophagy and degeneration of Purkinje cells. This degeneration might be mediated through interaction of Grid2 with its downstream autophagy protein, Beclin1 [20]. Autophagy is a conserved mechanism for degradation of proteins and other subcellular constituents, and is often involved in cell and tissue remodeling or cell death [21]. Two recent reports demonstrated that Purkinje cells also degenerate without the presence of the basal level of autophagy [22],[23].
Bcl-2 family members have been most extensively studied in the context of apoptotic cell death [24]. The Bcl-2 family was divided into the pro-survival members that protect cells from being killed, and the pro-death members that kill cells. Bcl-w belongs to the pro-survival group of the Bcl-2 family that includes Bcl-2, Bcl-XL, A1, and CED-9 [25],[26]. These proteins function to protect cells from apoptosis by binding to the outer membrane of mitochondria through their C-terminal hydrophobic domain, thereby preventing the release of several apoptosis proteins from mitochondria into the cytoplasm. They include the caspase regulatory proteins and proteins that lead to DNA fragmentation and chromosome condensation [27]. Bcl-w is widely expressed in a variety of tissues, but predominantly in adult brain and spinal cord [28]. The expression of Bcl-w in brain increases during the postnatal development and is maintained at high levels in the adult brain including cerebellum Purkinje cells, where it is localized to Purkinje cell soma [29] and dendrites (Lab Vision Corporation), whereas Bcl-XL, the only other pro-survival member that is expressed in adult brain had much lower level of expression [29]. Bcl-w−/− mice are smaller during the early postnatal development, but viable and normal in appearance as adults. Both apoptotic and non-apoptotic cell death have been observed in the testes of Bcl-w−/− mice [30],[31].
A recent report as well as our own findings demonstrated that several other survival members of the Bcl-2 family including Bcl-w could also bind to the autophagy protein, Beclin1 [32]–[34]. Beclin1 has been shown previously to form a protein complex with an adaptor protein, nPIST, and the intracellular domain of Grid2 [20]. Thus, Bcl-w might interact genetically with Grid2 to regulate mitochondrial, autophagy, and neuronal function.
In this study, we aim to understand how Bcl-w and Grid2 interact genetically to regulate mitochondria, autophagy, and neuronal function using Bcl-w and Grid2 null mutant mice. We show that the survival member of the Bcl-2 family member, Bcl-w does not control the cell number in brain, but promotes what is likely to be the mitochondrial fission in Purkinje cell dendrites, and is required for the Purkinje cells/parallel fibers synapses and motor learning. We demonstrated that the excitatory receptor Grid2 could regulate mitochondrial length, and the mutation of this protein shares the similar phenotypes in cerebella with the loss-of-function of Bcl-w. Comparative analyses of single and double mutant mice of Bcl-w and Grid2 further indicate that these molecules act synergistically to regulate mitochondrial length and to control the development of Purkinje cell dendrites, dendritic spines, and synapse formation. We further show that no evidence of alteration of autophagy in single and double mutant mice was observed, and the potential upregulation of Beclin1 in Bcl-w−/− mice and overexpression of Beclin1 was not sufficient to activate autophagy. We have thus identified Bcl-w and Grid2 as two critical proteins acting in distinct pathways to control mitochondrial morphogenesis and Purkinje cell development in the mouse cerebellum.
Since Bcl-w binds to Beclin1, which in turn can form a protein complex with nPIST and the intracellular domain of Grid2 [20], the possibility arises that Bcl-w may function downstream of Grid2. We thus examined if Bcl-w−/− [30],[31] and Grid2ho−4J(−/−) mice [14],[35],[36] that carry spontaneous null mutation of the Grid2 gene share similar phenotypes. Since Bcl-w binds to the outer membrane of mitochondria to regulate apoptotic activity [26], we examined both cell numbers (see below) and the morphology of mitochondria in Bcl-w−/− and Grid2ho−4J(−/−) mice by electron microscopy (EM). Purkinje cells were focused on because Grid2 is only expressed in Purkinje cells [9]–[11], and Bcl-w also had strong expression in these cells in adult brain [28],[29].
In these EM micrograph, profiles of mitochondria collected from longitudinal sections of dendritic tracks in electron micrographs appear lengthened in both Bcl-w−/− and Grid2ho−4J(−/−) mice (Figure 1A). The lengths of mitochondria were thus measured and quantified. Palay and Chan-Palay [37] have demonstrated using EM method that the mitochondrial lengths in Purkinje cells of wild type mice are ∼0.1–0.6 μ. In the present study, wild type mice yielded an average value ∼0.7 – 0.8 μ, with about two third of mitochondria measuring between 0.1–0.8 μ (Figure 1B). This is similar to the previous electron micrographic estimates [37]. In Bcl-w−/− mice, the average length of mitochondria was increased to ∼1.4–1.5 μ (Figure 1A, B), similarly to that in Grid2ho−4J(−/−) mice, ∼1.3–1.6 μ. Both numbers are significantly different from that obtained in wild type mice (Figure 1B). In addition, it was notified upon detailed examination of the micrographs that mitochondria in both Bcl-w−/− and Grid2ho−4J(−/−) mice often contained points where they became constricted (Figure 1A). This observation suggests that the lengthened mitochondria might be due to the inhibition or slowdown of mitochondrial fission process.
To understand if Bcl-w and Grid2 act in the same or separate pathways, we generate double mutant mice of Bcl-w and Grid2. The reason for this was that if mitochondrial and synaptic phenotypes in Bcl-w−/− Grid2 ho−4J(−/−) double null mutant mice was similar to that in either of the single knockouts, then it can be concluded that Grid2 and Bcl-w function in the same pathway; whereas a finding be concluded that the additive or synergistic phenotypes of Bcl-w and Grid2 mutants would suggest that they act instead in distinct pathways. EM analysis of mitochondria in Purkinje cell dendrites of Bcl-w−/−Grid2 ho−4J(−/−) mice indicated that the average value ∼1.8–1.9 μ (Figure 1A, B) was significantly longer than those mitochondrial profiles obtained from the single mutant mice (Figure 1A, B). In addition, mitochondrial profiles from Bcl-w−/−Grid2 ho−4J(−/−) mice contained frequent thinning and constriction sites (Figure 1A; blue arrows), and in some cases their cristae appeared slightly dilated (Figure 1A), indicating perhaps much slower mitochondrial fission compared to single and wild type mice.
The mitochondrial length estimated in Bcl-w−/−Grid2 ho−4J(−/−) mice is likely to be much underestimated length because very long mitochondria transit out of plate on very thin EM sections. In order to view the morphology of large mitochondria in Purkinje cells, we thus made thicker, 0.5 μ semi-thin plastic sections in distal Purkinje cell dendrites of wild type, single, and double mutant mice (Figure 1C). Intriguingly, many extremely long mitochondria (often >10 μ), which can extend for much of the visible length of Purkinje cell dendrites were frequently found in Bcl-w−/− Grid2 ho−4J(−/−) double mutant mice (Figure 1C). In addition, small mitochondria in dendrites that contain extremely long mitochondria seem depleted. However, these were rarely found in single, and not at all in wild type mice. Since the mitochondrial length in double mutant mice seem longer than the sum of that in the single mutant mice in semi-thin sections, this supports strongly that Bcl-w and Grid2 genes interact synergistically rather than additively to control the mitochondrial length, and that the severely increased mitochondrial length in Bcl-w−/− Grid2 ho−4J(−/−) double mutant mice might result from significant slowdown of the mitochondrial fission in Purkinje cell dendrites.
We next examined if Purkinje cell number was altered in Bcl-w−/−, Grid2ho−4J(−/−) and Bcl-w−/−Grid2 ho−4J(−/−) mice. The rational for this experiment is that cell death has been observed in testes of Bcl-w knockout mice [30],[31] and in Purkinje cells of Grid2Lc mutant mice [17], and that mitochondrial fission can stimulate apoptosis and cell degeneration. In these studies, we found that Bcl-w−/− mice brains appeared grossly normal, and no significant difference in Purkinje cell number (Figure 2A, B), or obvious changes in neuron number in other brain regions compared to wild type mice were observed (data not shown). A similar result was obtained in Grid2ho−4J(−/−) mice, agreeing with previous studies of Grid2 mutant mice [38]. Despite the fact that cerebella from the Bcl-w−/− Grid2 ho−4J(−/−) animals were obviously smaller and contained an overcrowded Purkinje cell monolayer (Figure 2A), normal Purkinje cell numbers were found in the double mutant mice (Figure 2B). This result suggests that Bcl-w and Grid2 promote perhaps mitochondrial fission in non-degenerating Purkinje cells.
Previous EM studies of Grid2 null mutant mice have revealed a large number of naked Purkinje cell dendritic spines, and mismatched connections between the pre- and postsynaptic active zones of Purkinje cell/parallel fiber synapses [12],[35]. Interestingly, our EM analysis of these synapses in Bcl-w−/− mice also indicated a large number (∼44% of total spines) of naked spines, and several synaptic defects including mismatched connections, shortened active zones, and thickened postsynaptic densities (Figure 3A; Table 1). Although the average number of naked spines did not change between Grid2ho−4J(−/−) (∼50% –70%) and Bcl-w−/−Grid2ho−4J(−/−) mutant animals (∼65%) (Table 1), there were much fewer synapses in the double mutant mice due to the significantly reduced number of dendritic arbors (see below; Figure 4).
Using the accelerating Rotarod behavioral test [12],[15], we obtained evidence that the synaptic defects in the Bcl-w−/− mice observed in EM studies may result in deficits in cerebellar motor learning function in these mice (Figure 3B). We found that whereas wild type mice improved their performance with experience on the rotating bar, Bcl-w−/− mice consistently failed to improve in performance throughout the trials (Figure 3B). Interestingly, a similar result has been previously reported in the loss-of-function of Grid2 mice [12],[35]. To rule out potential neuromuscular abnormalities as the cause of this phenotype, a hanging wire test [39] was performed on Bcl-w−/− mice. No obvious differences in the retention time between Bcl-w−/− and wild type mice were observed (data not shown), suggesting that the lack of motor learning evident in Bcl-w−/− mice in the Rotarod assay is due to defects in cerebellar function.
In summary, these results demonstrated that both Bcl-w−/− and Grid2ho−4J(−/−) mice had the similar phenotypes including significantly lengthened mitochondria in Purkinje cell dendrites, fewer and malformed Purkinje cell/parallel fiber synapses, and motor learning defects.
Visually, Bcl-w−/−Grid2 ho−4J (−/−) double mutant mice were immediately distinguishable from wild type and the single mutant animals by the fact that they were smaller in size, moved very little, and when prodded moved with an extremely ataxic gait. These motor difficulties were so severe that these mice could not be properly tested in the Rotarod assay (data not shown).
To examine these double mutant mice for histological abnormalities, the Golgi impregnation technique was used to further visualize the architecture of Purkinje cell dendrites and the morphology of dendritic spines (Figure 4A). The Strahler method of ordering (Figure 4B) was subsequently applied to obtain quantitative estimates on the impact of these mutations on the complexity of dendritic trees [40],[41]. In the Strahler method of ordering, each dendritic arbor is assigned as “order” such as primary, secondary, tertiary, etc. Dendritic arbors in each order are then subsequently quantified (Figure 4B). An analysis of the data using the Strahler method demonstrated that the single mutant mice did not have significant differences in each of six Strahler orders compared with that of wild type mice. By contrast in the double knockout mice, Purkinje cell dendritic arbors were reduced significantly that there were four or five Strahler orders compared to six orders in wild type and single knockout mice. Furthermore, the number of branches in each order was also reduced significantly in the double mutant mice in comparison to the single knockout and wild type mice (Figure 4B; Table 2).
Additionally, we also found that ∼45% Purkinje cells analyzed in the double mutant mice contained two dendritic branches rather than the single primary branch that extended from the Purkinje cell soma. This compared to ∼5% Purkinje cells analyzed in wild type and the single mutant animals (Table 3).
Examination of Purkinje cell dendritic spines using Golgi impregnation in 4–6 mice (Figure 5A–D) for each genotype revealed that Purkinje cell dendritic spines in wild type, Bcl-w−/−, and Grid2ho−4J(−/−) mice all had the characteristic door knob-shaped structure and spacing expected, whiles in Bcl-w−/−Grid ho−4J(−/−) mice they were crowded onto dendrites, appeared significantly shorter, and often branched or lacked a clearly distinguishable spine head or neck (Figure 5A, B). Similar spine defects were found in EM studies in an additional three double mutant mice.
To quantify the difference in spine length between single and double mutant mice, we measured spine profiles that have their necks connected to dendrites (Figure 5C). The wild type Purkinje cell spine lengths determined by this method (0.86±0.33 μ; Figure 5D) were agreed very well with previous measurements of mouse Purkinje cell dendritic spine length (0.87±0.21 μ) obtained using confocal microscopy of Lucifer yellow injected Purkinje cells [42]. In this study, wild type and single mutant mice had similar length of spines, although the spine length decreased significantly by ∼25% in Bcl-w−/−Grid2 ho−4J (−/−) Purkinje cells (Figure 5C, D).
In summary, these results demonstrated that normal Purkinje cell dendrite development and synapse formation require both Bcl-w and Grid2. Since both dendritic arbor number and spine length in double mutant mice are much more severely affected than the sum of these defects in single mutant mice, we conclude that the interaction between Bcl-w and Grid2 genes in regulation of dendrite and spine development is synergistic.
Two recent reports demonstrated that a basal level of autophagy was required for preventing the accumulation of protein aggregates and inclusion bodies and the survival of Purkinje cells [22],[23]. Since Bcl-w can interact physically with Beclin1 and inhibit starvation-induced autophagy (unpublished results), and the dominant Grid2Lc mutation can induce autophagy in Purkinje cells, it is possible that Bcl-w−/− and Grid2ho−4J(−/−) mutations might potentially alter autophagy in Purkinje cells and result in observed phenotypes in these cells. We thus examined anatomic evidences of autophay in Purkinje cell EM sections of single and double mutant mice.
It has been shown previously that morphologic evidence for the activation of autophagy indicated by the presence of autophagosomes was readily apparent in Grid2Lc Purkinje cells [20]. In contrast, inhibition of basal autophagy in Purkinje cells can result in the accumulation of inclusion bodies or protein aggregates [22],[23]. However, in a careful examination of the Purkinje cell cytoplasm in wild type and single and double mutant animals in electron micrograph, no morphological evidence indicative of alterations in autophagy was observed (data not shown), suggesting that autophagy is unlikely to be the cause of Purkinje cell phenotypes in mutant mice.
During mitochondrial fission process, the mitochondrial fission protein complexes localize on the fission sites, which appear as constriction sites during ongoing mitochondrial proliferation [2],[3]. When the mitochondrial fission is blocked, the constriction sites persist and can be identified easily under the electron microscope [43]. Both the growth of mitochondria and the mitochondrial fission/fusion processes determine the final size or length of the mitochondria in cells. However, it should be notified that mitochondrial growth alone does not generate constriction sites. The frequently observed constriction sites in mitochondria of double mutant mice strongly support that the lengthened mitochondria in single and double mutant mice are due to the slowdown of mitochondrial fission process. Since small mitochondria are seemingly depleted in dendrites that contain extremely long mitochondria in the semi-thin section of double mutant mice, this also supports that slow mitochondrial fission led to the decreased number of mitochondria. However, we cannot rule out that these mitochondrial phenotypes were due to enhanced fusion process.
The survival members of the Bcl-2 family have not been previously reported to regulate mitochondrial length or mitochondrial fission/ fusion in mammalian cells. However, the pro-death members of Bcl-2 family, Bax and Bak have been demonstrated to regulate mitochondrial fission and fusion in both apoptotic and living healthy cells [43]–[45]. For example, in C. elegans, overexpression of EGL-1 can induce mitochondrial fission and apoptosis [46]. In mammals, Bax or/and Bak promote mitochondrial fission in apoptotic cells through regulating mitochondrial fission proteins directly [43] or indirectly [47]. In these cells, Bax could also inhibit mitochondrial fusion [44]. However, in living cells Bax and Bak act oppositely as they function to promote mitochondrial fusion [45]. We show in this study that the survival member, Bcl-w has similar characteristics. Thus, Bcl-w promotes mitochondrial fission in Purkinje cells, whiles in testis it protects cells that should normally die in apoptosis.
The excitatory receptors have not been previously reported to regulate mitochondrial length or dynamics. Previous studies though have demonstrated in neuronal culture system that synaptic activity can stimulate mitochondrial fission and clustering to the dendritic spines [7]. In vivo results for this regulation are lacking, however. Our results implicate that the excitatory receptor Grid2 regulates mitochondrial morphology in addition to its previously found regulation of channel activity and other functions.
The in vivo analysis of mitochondrial localization in Purkinje cells in the current study [37] indicates that mitochondria are normally present in dendrites, but rarely inside dendritic spines. Thus, the actions of Grid2 from synapses on mitochondria may be indirect because Grid2 is localized in Purkinje cell/parallel fiber synapses. The synergistic effect in Bcl-w and Grid2 double mutant mice on mitochondrial length rules out the possibility that Grid2 promotes mitochondrial fission mainly through regulating Bcl-w, and suggests that other pathways might be responsible for Grid2 in regulation of mitochondrial length or fission/fusion. Our studies in the dominant Grid2Lc mutation demonstrated the extensive mitochondrial fragmentation in cytoplasm of Purkinje cell of Lurcher mice during the postnatal development (unpublished results). This suggests that Grid2 might control mitochondrial length through the mitochondrial fission process by regulating calcium influx. Indeed, calcium has been shown in several studies to stimulate mitochondrial fission by regulating the activities of dynamin and the dynamin-like large GTPase, Drp-1 [6]. Thus, Grid2 is likely to function to promote mitochondrial fission through its channel activity.
In this study, we observed fewer Purkinje cells/parallel fiber synapses, an increased ratio of naked spines, and motor learning defect in the double mutant mice. This may be correlated with slowdown of mitochondrial fission in Bcl-w−/− and Grid2 ho−4J(−/−) mice, as suggested by our studies. A more direct correlation between mitochondrial fission or fusion and the number of spines and synapses has been demonstrated in primary neuronal culture; overexpression of mitochondrial fission protein Drp-1 in these cells resulted in increased number of mitochondria, correlated with increased number of spine and synapse. In contrast, the expression of mitochondrial fusion protein, OPA1 or dominant negative version of Drp-1 has been reported to lead to fewer numbers of spines and synapses [7]. These results implicate that mitochondrial fission in healthy cells might serve as a means to increase the number of mitochondria to meet energy demands during neuronal growth, or neuronal plasticity, and is likely to be different in mechanism from the excessive mitochondrial fission observed during apoptosis and neurodegeneration. Indeed, the mitochondrial fission during apoptosis can result in the loss of mitochondrial DNAs and lower the function of mitochondria [6]. However, in living healthy human cells, Benard et al. demonstrated recently that when mitochondrial fragmentation was inhibited, a strong inhibition of mitochondrial energy production was observed [48].
Bcl-w is localized in Purkinje cell dendrites and acts on mitochondria, the synaptic defects in Bcl-w−/− mice are thus likely to be the consequence, not the cause of the mitochondrial morphogenesis defects. Since mitochondrial fission resulted from the neuronal excitation is linked to the danger of degeneration, it is intriguing that the survival member, Bcl-w could promote mitochondrial fission, and has protective function to cells as well.
In summary, the results in the current study suggest that the Bcl-2 family member, Bcl-w, and the excitatory receptor Grid2 can regulate the mitochondrial fission and thus mitochondrial length in dendrites. Altered mitochondrial length in mutant mice of these genes in turn results in abnormalities in synapse formation in the mice.
The Bcl-2 family has not been shown previously to regulate neuronal dendrite development; its effect on neuronal growth has been only associated with cell death. In comparison, the NMDA receptor, one of the glutamate receptor family members has been demonstrated to regulate the activity-dependent dendrite development [49]. However, it is not known if this function of the NMDA receptor has anything to do with mitochondrial morphogenesis or Bcl-2 family members. In this study, we demonstrate that the normal development of Purkinje cell dendrite, dendritic spine, and synapse formation requires both Bcl-w and Grid2, and their regulation of mitochondrial morphogenesis.
Mitochondrial proliferation is a biological process that is associated with cellular division and growth. It takes normally three weeks for Purkinje cells to grow from cell bodies into fully-grown trees with extensive synaptic connections [50]. During this period of time, mitochondrial number also increases significantly. This mitochondrial growing process cannot be exhibited well in cultured Purkinje cells that contain very short and little branched dendritic arbors, unfortunately [51]. The significantly inhibition of mitochondrial fission can thus result in decreased small mitochondrial number and the large size of fused mitochondria that result in reduced mitochondrial motility in Purkinje cell dendrites. These could place an intrinsic limitation on the local energy production and calcium buffering in dendrites, resulting in a failure of perhaps neuronal development and function such as the dynamic growth and branching of dendrites, the development and plasticity of dendritic spines and synapses, channel activities, and the formation of the postsynaptic density, thus leading to the severe morphological defects observed in the double mutant Purkinje cells.
The abundance of mitochondrial fission during the Purkinje cell growth is also balanced or controlled by mitochondrial fusion. A recent paper demonstrated that the absence of mitochondrial fusion protein Mfn2 during Purkinje cell development resulted in excessive mitochondrial fragmentation and Purkinje cell degeneration, suggesting that mitochondrial fusion is required to prevent cells from degeneration [51]. Similarly, the reason that we did not observe any Purkinje cell death in Bcl-w−/−Grid2 ho−4J(−/−) mice in spite of the extensive loss of dendrites, spines, and synapses is likely due to the protective effect by extensively fused mitochondria in these cells. Mitochondrial fusion has also been shown in cultured cells to protect cells from cell death [52],[53].
The early developmental defects in Purkinje cell primary branches in Bcl-w−/−Grid2 ho−4J(−/−) mice indicate that dendritic defects, at least initially, are caused intrinsically, not due to the granule cell parallel fiber innervations because these innervations occur later than the emergence of Purkinje cell primary branches [50].
It has been hypothesized that mitochondrial respiration and metabolism may be spatially and temporally regulated by mitochondrial morphology and location that can be integrated to multiple pathways of cellular function [54]. The regulation of mitochondrial length that can result from mitochondrial fission or fusion thus might participate in other pathways that control dendrite and spine morphology and synapse formation, such as development, diseases, and in response to many extrinsic factors such as neuronal activity, hormones [55],[56], and chronic stress [57]. Since we have only observed the mitochondrial morphology changes in fixed tissues, we hope that we can also demonstrate that Bcl-w and Grid2 can affect the mitochondrial fission or fusion in a real time system. This system can also be used to understand the mechanism for Bcl-w and Grid2 and their family members to regulate mitochondrial morphology or the mitochondrial fission and fusion cycle. The studies on mitochondrial fission or fusion will yield important knowledge for our understanding of the development and function of central nervous system.
Bcl-w knockout mice were obtained from Dr. Grant Macgregor’s lab from Emory University (currently at University of California, Irvine). The generation and typing of these mice were described previously [30]. DBA/2J-Grid2ho−4J(−/−) mice were purchased from Jackson lab. These mice carry spontaneous null mutation of the Grid2 gene with exons 5–8 deleted, resulting in a 170 amino acid loss in the N-terminal LIVBP-like domain [14],[35],[36]. To obtain Bcl-w−/−Grid2ho−4J(−/−) double knockout mice, the male Grid2ho−4J(−/−) mice were crossed into the female Bcl-w−/− mice to obtain Bcl-w+/−Grid2ho−4J(+/−) mice in F1 generation. Both male and female Bcl-w+/−Grid2ho−4J(+/−) mice from F1 generation were selected and crossed with each other to obtain F2 generation mice. Pups in F2 generation demonstrating the “hotfoot” ataxic phenotypes were identified as homozygous Grid2. Molecular genotyping [30] was applied to distinguish Grid2ho−4J(−/−), Bcl-w+/−Grid2ho−4J(−/−), and Bcl-w−/−Grid2ho−4J(−/−) animals. Bcl-w−/−Grid2ho−4J(+/−) mice do not have an obvious ataxia phenotype because mice obtained from the cross using Bcl-w+/−Grid2ho−4J(+/−) male and Bcl-w−/− female mice did not show the obvious ataxic phenotypes.
Mice were perfused with 2.5% glutaraldehyde, and cerebella were sliced sagitally, and each slice was then diced into pieces containing 2–3 folia. The tissue pieces were post-fixed in 1% osmium, treated with 0.5% aqueous uranyl acetate, and then dehydrated through graded alcohol (70, 95, and 100%). After the treatment with propylene oxide, tissue pieces were embedded in a manner allowing sectioning in the sagital plane in Ducupan (Fluka). The blocks were cured in a 60°C degree oven for 2–3 days.
Blocks were cut with a glass knife to get Semi-thin sections of 0.5 micron. The sections were then stained with 0.25% toluidine blue in 1% sodium borate, and evaluated at the light microscope (LM) level to select the tissue orientation of sagital and longitudinal sections. Photographs of mitochondria were taken in the molecular layer approximately 1/3 of the distance between the pia and Purkinje cell monolayer using a 100X oil lens in a Zeiss Axioplan light microscope and MetaVue acquisition software (Universal Imaging). Four wild type, four Bcl-w−/−, three Grid2ho−4J(−/−), and three Bcl-w−/−Grid2ho−4J(−/−) were examined.
To obtain the EM pictures, the sagital and longitudinal block faces were trimmed, and ultra-thin silver sections were cut with a Reichert-Jung Ultracut E ultramicrotome with a Dupont diamond knife, and collected on copper grids, stained with saturated aqueous uranyl acetate, and lead citrate before examination in Jeol 100 cx electron microscope operated at 80 Kv. EM photographs were taken on dendritic tracks randomly in the molecular layer approximately 1/3 of the distance between the pia and Purkinje cell monolayer in sections collected from several different blocks. Mitochondria profiles were traced and measured from the longitudinal dendritic tracks of each set of photographs.
For morphometric analysis of synapses, only one section was collected on each grid. After establishing the orientation, locating the pia and Purkinje cell layer, images were recorded from regions 1/3 down from the pia at primary magnification of 6,600X x. The print magnification was 16,500X. For locating proximal spines on the Purkinje cell primary branches, the block face was reduced allowing more sections to be collected per grid.
The male wild type mice were crossed to the female Bcl-w−/− mice to obtain Bcl-w+/− mice, which were subsequently inbred with each other to obtain littermates of wild type, Bcl-w+/−, and Bcl-w−/− mice. Six to eight weeks old of these littermates were tested on accelerating Rotarod with 0.1 round/second as starting speed and 0.4 round/second2 as accelerating speed. Retention time were begin to be recorded when the mice were placed on the rotating bar and acceleration was applied, and stopped when they failed to run on the rotating bar. Each mouse was given three trials per day for five constitutive days.
The hanging wire test was performed by placing mouse on the top of a wire cage lid, and after mouse grip the wires, lid was turned upside down [39]. The retention time for the mouse to hold wire was recorded. Four Bcl-w−/− and wild type mice were tested, respectively.
Mice were intracardially perfused with 4% paraformaldehyde in PBS, and their brains were subsequently dissected and post-fixed overnight. Brains were then dehydrated with 70, 95, and 100% ethanol, and treated in organic solvent, butanol for three days before replacing it with paraffin. 10 μ sections were obtained from paraffin-embedded brain. These sections were then treated with xylene to remove wax and rehydrated before staining them in Cresyl violet and being mounted on slides. Sections obtained in the region close to midline were selected and counted for Purkinje cells on all folia.
Golgi stain of mouse brain was obtained using FD Rapid GolgiStain Kit (FD NeuroTechnologies, Inc.) according to the manufacture’s instruction. The images of Purkinje cells were collected using the DIC microscope (Zeiss) with 10X object lens (Figure 4A, upper panel) and with 5X lens (Figure 4A, lower panel). To analyze Purkinje dendrite branches using the Strahler method of ordering, a Z-stack of Purkinje cell images was collected using MetaVue acquisition software (Universal Imaging), and 20X water lens, and used to quantify Purkinje cell dendritic branches. The spines and dendrites were photographed from combined images from Z-stack (Figure 5A). The spines on dendritic branches of the first Strahler order (Figure 5B) were photographed using 100X oil lens. The number of mice examined was indicated in Table 1. |
10.1371/journal.pgen.1001043 | Breast Cancer DNA Methylation Profiles Are Associated with Tumor Size and Alcohol and Folate Intake | Although tumor size and lymph node involvement are the current cornerstones of breast cancer prognosis, they have not been extensively explored in relation to tumor methylation attributes in conjunction with other tumor and patient dietary and hormonal characteristics. Using primary breast tumors from 162 (AJCC stage I–IV) women from the Kaiser Division of Research Pathways Study and the Illumina GoldenGate methylation bead-array platform, we measured 1,413 autosomal CpG loci associated with 773 cancer-related genes and validated select CpG loci with Sequenom EpiTYPER. Tumor grade, size, estrogen and progesterone receptor status, and triple negative status were significantly (Q-values <0.05) associated with altered methylation of 209, 74, 183, 69, and 130 loci, respectively. Unsupervised clustering, using a recursively partitioned mixture model (RPMM), of all autosomal CpG loci revealed eight distinct methylation classes. Methylation class membership was significantly associated with patient race (P<0.02) and tumor size (P<0.001) in univariate tests. Using multinomial logistic regression to adjust for potential confounders, patient age and tumor size, as well as known disease risk factors of alcohol intake and total dietary folate, were all significantly (P<0.0001) associated with methylation class membership. Breast cancer prognostic characteristics and risk-related exposures appear to be associated with gene-specific tumor methylation, as well as overall methylation patterns.
| The current standard prognostic indicator for breast cancer is tumor-node-metastasis staging; though, as population-based studies and clinical trials are conducted, molecular characterization of disease is beginning to allow improved markers of prognosis and assist clinicians in choosing the most appropriate therapies. We investigated DNA methylation profiles in over 160 well annotated breast tumor samples and found significant relationships with standard and other known predictors of prognosis, as well as established risk factors for disease: alcohol intake and dietary folate. Recently the United States National Cancer Institute Cancer Biomarkers Research Group articulated a need for a “Strategic Approach to Validating Methylated Genes as Biomarkers for Breast Cancer,” and our work is extremely responsive to this call for a national strategy. Recognizing the increasing use of pre-operative chemotherapy for patients with operable, early-stage disease, there is added complexity in breast cancer staging. Since chemotherapy can considerably decrease tumor size, it is still unclear whether pre-operative or post-operative stage best informs prognosis and treatment decisions for patients electing pre-operative chemotherapy. However, our data clearly illustrate the promise of tumor DNA methylation for augmenting tumor staging and can be attained with minimal tissue in a pre-operative context.
| Breast cancer is the most common non-skin cancer among American women. The American Cancer Society's estimates indicate approximately 1.3 million new cases of invasive breast cancer were diagnosed globally in 2007; and nearly 500,000 women died from the disease [1]. Currently, there are over 2.5 million breast cancer survivors in the US, and an estimated $8.1 billion dollars is spent each year on treatment of breast cancer [2].
The principal prognostic indicator currently in clinical use for breast cancer is the tumor-node-metastasis (TNM) stage [3], [4]. Morphological attributes of malignant tumors that influence disease prognosis are the size of the primary tumor (T), presence and extent of regional lymph node involvement (N) and presence of distant metastases (M). Molecular attributes of tumors are also considered in clinical decision-making; loss of hormone receptor expression [5] and increased expression of ERBB2 [6] have each been associated with poor prognosis. Although numerous recent studies have demonstrated that alterations of DNA methylation in breast cancers are common and may be important etiologic and prognostic markers [7]-[14], large gaps in our knowledge remain. There is a notable lack of studies examining tumor DNA methylation in relation to breast cancer risk factors such as diet or reproductive factors in conjunction with other important tumor markers. Patient exposures such as alcohol and folate intake have potentially strong mechanistic links to epigenetic dysregulation [15]. In addition, recent work in-vitro and in animal models suggest that long term exposure to estrogen may lead to epigenetic effects and altered profiles of DNA methylation [16], [17]. To explore associations of tumor methylation with important tumor and patient characteristics, we analyzed tumors from breast cancer patients in the Kaiser Permanente Division of Research Pathways Study using a large scale methylation array.
Table 1 shows the patient demographic, hormonal, dietary and tumor characteristics for the 162 women overall (and stratified by menopausal status in Table S1). Results of unsupervised hierarchical clustering of the 750 most variable CpG loci indicate the epigenetic heterogeneity of these tumors (Figure 1).
In array-wide locus-by-locus analysis the strongest associations of methylation of individual loci (Q-values <0.05) were observed for tumor grade (loci n = 209), tumor size (loci n = 74), estrogen receptor status (loci n = 183), progesterone receptor status (loci n = 69), and triple negative status (tumors negative for both estrogen and progesterone receptors as well as ERBB2; loci n = 130; Table S2). Together with tumor size, patient lymph node status is used in tumor staging. Among five CpG loci whose methylation was significantly associated (Q<0.05) with lymph node status, four (two in COL1A2, and one each in LOX and P2RX7) were also associated with tumor size (Q<0.05). Additionally, there was a trend of increased methylation associated with increased tumor size: for all 74 CpG loci that were significantly associated with tumor size (Q<0.05) methylation increased with larger tumor size. Similarly, all five CpGs associated with disease-positive lymph nodes had increased methylation in tumors in women with disease-positive lymph nodes. Details of locus-by-locus analyses for tumor grade, size, hormone receptor, and triple negative status (loci with Q<0.05) are given in Table S3.
Methylation array validation was performed at CpGs with highly ranked associations from locus-by-locus analysis. The array CpG whose methylation was most significantly increased with increasing tumor stage was in the FES gene (Table S3) and array methylation was significantly correlated with Sequenom methylation (rho = 0.68, P = 1.1E-12, n = 85; Figure 2A). Promoter CpGs in P2RX7 and HSD17B12 had significantly increased methylation (Q<0.0001, and Q = 0.01 respectively) with increasing tumor size (Table S3) and array methylation at these CpGs were significantly correlated with Sequenom methylation (P2RX7; rho = 0.65, P = 8.6E-12, n = 88; HSD17B12; rho = 0.34, P = 5.4E-05, n = 137; Figure 2B and 2C). A promoter CpG in GSTM2 had significantly increased methylation with increasing tumor grade (Table S3) and array methylation was significantly correlated with Sequenom methylation (rho = 0.83, P<2.2E-16, n = 140; Figure 2D). Additionally, in all cases, Sequenom methylation values were significantly associated with respective covariates; tumor stage with FES methylation (P = 0.05), tumor size with P2RX7 (P<0.005) and HSD17B12 methylation (P<0.02), and tumor grade with GSTM2 methylation (P<0.001). Furthermore, relative mRNA expression of GSTM2 was significantly decreased among tumors with high array methylation at both CpGs associated with tumor grade (P<0.001 and P<0.03, Figure S1).
In order to explore overall methylation profiles of these tumors and their potential relationships with patient demographic, tumor and exposure characteristics we applied a modified model-based form of unsupervised clustering known as recursively partitioned mixture modeling (RPMM) [18]. The RPMM resulted in the eight methylation classes (average methylation profiles shown in Figure 3). Patient race was significantly associated with methylation class membership (P = 0.015, Table 2), with the majority of African Americans (54%) residing in class 2, and 40% of Hispanic cases residing in class 4. An association between methylation class membership and alcohol consumption approached statistical significance (P = 0.07, ever vs. never drinker, Table 2). Both supplemental folic acid intake (µg/day) and total dietary folate (µg/day) had associations with methylation class membership that approached statistical significance (P = 0.06 and P = 0.08 respectively; Table 2). For both folate variables, cases in methylation class 4 had the lowest intake and cases in methylation class 6 had the highest intake. Of the tumor characteristic variables, only tumor size was significantly associated with overall methylation profile (P = 0.0006, Table 2).
Associations between alcohol intake and dietary folate and methylation class membership approached statistical significance. While methylation of only one CpG locus (in IL17RB) was significantly associated with folate intake in locus-by-locus tests (Q<0.05), regression coefficients from univariate locus-by-locus analysis plotted against their respective P-values revealed trends in the pattern of methylation for both alcohol and folate intake. Figure 4A illustrates the strong trend for patients with increasing alcohol intake to have negative regression coefficients, indicative of decreased methylation. In contrast, the trend for patients with increasing total dietary folate shows a strong shift to positive regression coefficients, indicative of increased methylation (Figure 4B).
The relationships between methylation classes and several covariates of interest were then modeled together using multinomial logistic regression in order to adjust for other factors in the model. Patient age, alcohol consumption, total dietary folate, and tumor size were each strongly associated with methylation class membership when controlling for all modeled variables (all Wald P-values <0.0001) and complete model details are given in Table S4. Figure 5 displays an illustration of the model results for covariates significantly associated with methylation classes. As alcohol consumption increased, there was an increased probability of cases residing in methylation classes 3 and 8, and a concomitant decrease in the probability of cases residing in classes 2 and 4 (Figure 5B). Increasing total dietary folate intake imparted a striking increase in the probability of membership in class 6, and a decreased probability of class membership in classes 1, 3, 4, and 7 (Figure 5C). The strong association between tumor size and methylation class membership remained after controlling for potential confounders, with the probability of patients being in class 2 increasing from about 20% to about 60% across the span of tumor size from 0 mm to 80+mm (Figure 5D). Accompanying this trend for tumor size were simultaneous decreases in the probability of cases with increasingly large tumors residing in classes 1 and 5–8, while tumor size had less influence on the probability for residing in classes 3 or 4 (Figure 5D).
Although neither estrogen nor progesterone receptor status were significantly associated with RPMM methylation profiles, large numbers of specific CpG loci had significant methylation associations with these tumor characteristics in locus-by-locus analysis (Table S2 and Table S3). Compared to the overall population of women diagnosed with breast cancer in the Kiaser Permanente Northern California cancer registry from 200–2009, this surgical cohort has a higher prevalence of hormone receptor positivity (78% overall vs. 88% here), particularly among pre-menopausal women's tumors (74% overall vs. 95% here). We therefore stratified on menopausal status, running RPMM on methylation data from post menopausal patients' tumors only (n = 117). This model resulted in eleven methylation classes (Figure S2) and methylation class membership was significantly associated with estrogen receptor status (P<0.03), and the association for triple negative tumors approached significance (P = 0.07) detailed results available in Table S5.
It is becoming increasingly common to include data on molecular alterations from patient tumor samples into routine clinical practice as a means of improving prognosis and evaluating the predictive power of alterations of interest. As technology improves and population-based studies and clinical trials are conducted, medicine is being ushered into a new era of molecular characterization of disease. Tumor-node-metastasis (TNM) stage is the current prognostic indicator for breast cancer, though several clinical trials are currently under way to investigate the utility of molecular markers [19], and as more patients elect neoadjuvant therapy (specifically pre-operative chemotherapy), improved clinical staging and additional staging tools are poised to have great impact. Most current studies and one commercially available tool (Oncotype DX) are focused on gene expression markers, though the inherent instability of mRNA may make implementation of these strategies challenging outside of major surgical centers or centralized commercial laboratories. In contrast, DNA methylation is a stable mechanism of control of transcription, and the stability of DNA makes it an attractive target for accurate and reproducible assessment. Here we reported that tumor size, a cornerstone of breast cancer prognosis, is associated with tumor DNA methylation profile. In addition, we found that alcohol and folate intake, exposures related to disease risk, are independently associated with tumor DNA methylation profiles. This work sheds light on the relationship between important etiologic exposures and molecular subclasses of disease, extends the evidence for the utility of molecular characterization in tumor staging, and can be accomplished with minimal tissue in a pre-operative context.
The recently updated American Joint Committee on Cancer (AJCC) staging manual for breast cancer does not include additional molecular markers, though the committee acknowledged their consideration of markers such as hormone receptor status and stated that TNM staging “may play increasingly less important roles than understanding the biology of the cancer” [4]. Examining TNM variables we found that overall DNA methylation profile and methylation alterations in dozens of individual CpG loci were significantly associated with tumor size (all increased methylation). In contrast, methylation alterations of only five CpG loci (two in COL1A2, and one each in FAS, LOX, and P2RX7) were significantly associated with disease-positive lymph nodes. However, methylation of four of five lymph-node-positive associated CpGs (excepting FAS) were also significantly associated with tumor size, suggesting that these phenotypes are mechanistically related, and at least in part manifest via epigenetic alterations. As FAS encodes a TNF-receptor involved in regulating apoptosis it is not surprising that methylation-induced silencing of this receptor is associated with disease-positive lymph node status. In addition, hypermethylation of COL1A2 (collagen type I, alpha 2) has been associated with both proliferation and migration activity in bladder cancer [20], LOX is involved in the control of normal collagen deposition [21], and P2RX7 loss has been linked to morphologic changes in stroma related to altered collagen fibril alignment [22]. Collectively these data suggest that perturbations in collagen and collagen-related genes promote tumor growth and invasion, perhaps by altering the architecture of connective tissues in the tumor microenvironment. In support of this hypothesis, recent work in a mouse model has shown that altered mammary stromal tissue collagen expression significantly increases tumor formation and invasiveness potential [23]. Additionally, Chernov et al. showned that epigenetic alterations in collagen and collagen-related genes allows the deposition of an invasion-promoting collagen matrix in both breast and brain tumor cell lines [24].
The primary objective of TNM staging is to provide a standard prognosis nomenclature for patient care [4], and our results suggest that methylation markers may be a robust proxy for tumor size. Importantly, broader application of neoadjuvant therapy complicates breast cancer staging since chemotherapy can considerably decrease tumor size prior to surgical treatment, and it is still unclear whether clinical or pathologic stage best informs prognosis and treatment decisions [19]. The AJCC has added methodology (yc or ypTNM) for differentiating clinical and pathologic staging; in part, this is from recognition of the increasing use of neoadjuvant therapy for patients with operable, early stage disease [4], [25], [26]. Our data illustrate the promise of tumor DNA methylation for augmenting tumor staging. However, additional study of the relationship between tumor methylation and size in both pretreatment and postoperative samples is necessary. Specifically, the value of methylation to act as an additional marker of size in the neoadjuvant setting should be evaluated in future studies that compare both imaging and pathologically based size determination.
In order to evaluate the predictive power of DNA methylation profiles and individual loci for disease prognosis and recurrence, these patients continue to be followed for these events. Associations between DNA methylation and patient survival have been reported for individual genes such as GSTP1 and PITX [7], [8], [10], though overall DNA methylation profiles, or patterns of methylation at selected CpG loci or genes, may improve predictive power. Well recognized molecular subtypes of breast cancer such as hormone receptor negative and ERBB2 over-expressing tumors are known to be associated with reduced survival [27], and it will be necessary to extensively examine methylation markers stratified by commonly used molecular tumor markers. However, we did not find significant associations between ERBB2 status and CpG methylation in our analysis. Nonetheless, other well recognized molecular subtype markers; estrogen receptor, progesterone receptor, and triple negative status were among the covariates with the highest number of significant CpGs from array-wide locus-by-locus analysis. However, hormone receptor status and triple negativity were not associated with methylation profile when modeling all cases. Premenopausal patients' tumors in our surgical cohort had a higher prevalence of hormone receptor positivity compared to the overall population of premenopausal patients diagnosed with breast cancer. In order to address the potential bias this introduced we modeled the methylation profiles of postmenopausal patients' tumors separately and found a significant association between estrogen receptor status and methylation class. Additional study will be needed to better understand the role of hormone receptor and growth factor receptor expression in these tumors as they relate to methylation profile in the context of a patient's menopausal status.
We found significant, independent associations between both alcohol and folate intake and overall tumor DNA methylation profiles when controlling for potential confounders. Folate is a B vitamin that donates its methyl group for homocysteine remethylation to methionine as part of one-carbon metabolism. In turn, methionine is the methyl donor for DNA methylation via S-adenosyl methionine. However, alcohol is known to interfere with folate absorption in the intestine and hepatic release of folate, and hence, supply to tissues [28]. In fact, strong evidence of an etiologic role for alcohol in breast cancer has been reported in multiple meta-analyses of prospective and case-control studies with an excess risk for each alcoholic drink per day of about 10% [29], [30]. In contrast, meta-analysis of prospective studies has not provided clear support for an overall protective association between folate intake and breast cancer risk [31]. Yet, meta-analysis of case control studies of dietary folate, including results from the Shanghai Breast Cancer Study (whose participants are not regular alcohol drinkers) generally support a protective role for folate [31], [32].
While there have been numerous studies of alcohol and folate in relation to risk of breast cancer, investigations of the relationship between these exposures and epigenetic alterations in tumors themselves are scarce. Tao et al. reported that the prevalence of breast tumor methylation at CDKN2A, CDH1, and RARB did not differ by folate intake or lifetime alcohol consumption in genotype strata of one-carbon metabolism enzymes methylenetetrahydrofolate reductase (MTHFR) and methionine synthase (MTR) [33]. Consistent with these findings (and perhaps the lack of similar null results in the literature), we too did not find associations between alcohol or folate and methylation of CpG loci in CDKN2A, CDH1, and RARB. Further, after correcting for multiple comparisons, no CpG loci had significant alcohol-related methylation, and only one CpG locus (in the IL17RB promoter) was associated with folate intake. Alone, these results suggested that folate and alcohol intake do not influence tumor DNA methylation. However, plots of regression coefficients indicated strong independent trends for increased folate and reduced alcohol intake associations with increased CpG methylation. Since global, low-level effects of alcohol and folate intake on CpG methylation may not be detectable at individual CpGs in a genome-wide context, we examined the global relationships between alcohol or folate intake and DNA methylation using RPMM methylation classes. Modeling both exposures together revealed highly significant, independent associations between alcohol and folate and DNA methylation profile. Another human cancer for which alcohol is an important etiologic factor is head and neck squamous cell carcinoma, and previous work from our group demonstrated a similar relationship between DNA methylation profiles of these tumors and alcohol consumption [34]. Taken together with the weak mutagenic potential of alcohol [35], these results suggest that a major carcinogenic mechanism of action of alcohol is interference with epigenetic regulation through disruption of one-carbon metabolism.
In summary, we found tumor DNA methylation associated with tumor characteristics predictive of prognosis, and DNA methylation and patient exposures known to be related to disease risk. Additional study is needed to determine the prognostic value of DNA methylation markers. However, the potential clinical utility of tumor-size-related DNA methylation is apparent.
The Pathways Study is a prospective cohort study of breast cancer survival actively recruiting women diagnosed with invasive breast cancer from the Kaiser Permanente Northern California (KPNC) patient population since January 2006. Further study details are provided elsewhere [36]. Written informed consent is obtained from all participants before they are enrolled in the study. The study was approved by the IRB of KPNC and all collaborating sites.
During the in-person baseline interview, participants were asked detailed information on family history of cancer and reproductive history, including: age at first full-term pregnancy, number of biological children, breastfeeding, and menopausal status. Additional information was collected on smoking, alcohol use, hormone use (oral contraceptives, hormone replacement therapy), and demographics (age at breast cancer diagnosis, race/ethnicity, household income, education). Self-reported height and weight around diagnosis was obtained to calculate body mass index (BMI, kg/m2). Any missing values were supplemented by concurrent information from KPNC electronic medical records.
Data on estrogen and progesterone receptor status and ERBB2 expression were obtained from the KPNC Cancer Registry [37]. Tumor size was measured in a uniform manner by participating study pathologists. Data are collected, coded, and added to the KPNC registry approximately four months post-diagnosis to allow for the completion of treatment. For all breast surgical specimens, hormone receptor status and ERBB2 expression is routinely determined by IHC at the KPNC regional IHC lab, and if the IHC staining for ERBB2 expression is equivocal (less than 30% strong staining, but more than 10% weak staining), by fluorescence in situ hybridization at the KPNC regional cytogenetics lab.
162 tumor specimens from the initial diagnostic biopsy were obtained from the KPNC tumor biorepository for methylation analysis. All tumor specimens were from patients who did not receive neoadjuvant chemotherapy.
FFPE tissue DNA was extracted using the QIAamp DNA mini kit according to the manufacturer's protocol (Qiagen, Valencia, CA). DNA was treated with sodium bisulfite to convert unmethylated cytosines to uracil using the EZ DNA Methylation Kit (Zymo Research, Orange, CA) according to the manufacturer's protocol. Illumina GoldenGate methylation bead arrays were used to simultaneously interrogate 1505 CpG loci associated with 803 cancer-related genes. Bead arrays have a similar sensitivity as quantitative methylation-specific PCR and were run at the UCSF Institute for Human Genetics, Genomics Core Facility according to the manufacturer's protocol and as described by Bibikova et al [38]. GoldenGate array methylation data are publicly available on the Gene Expression Omnibus archive, accession GSE22290.
Array methylation was validated with Sequenom EpiTYPER base-specific cleavage and MALDI-TOF MS of bisulfite treated DNA [39]. EpiTYPER assays were designed for CpG loci both with significant associations between methylation and tumor characteristic variables as well as a high standard deviation of methylation values across samples. One assay (for COL1A2) failed the design process. Samples were processed at the UCSF Institute for Human Genetics, Genomics Core Facility. Briefly, PCR with primers located on either side of the CpG sites of interest are transcribed into an RNA transcript and cleaved base specifically. The cleavage products are analyzed by MALDI-TOF MS, and a characteristic mass signal pattern that distinguishes methyl-cytosine from thymine is obtained.
Messenger RNA expression was measured using RT-PCR with preamplification using a validated approach [40]. RNA extraction was performed using the RecoverAll (Ambion), with a 16 hour tissue digestion and yields were determined using a Nanodrop spectrophotometer. Samples were concentration-normalized and reverse-transcribed with iScript cDNA synthesis kit (BioRad). Following cDNA synthesis, we performed linear, gene specific preamplification of samples and controls using the TaqMan preamp protocol (Applied Biosystems). Relative expression was measured using a HT7900 real time PCR instrument (Applied Biosystems).
|
10.1371/journal.pgen.1004391 | PINK1-Mediated Phosphorylation of Parkin Boosts Parkin Activity in Drosophila | Two genes linked to early onset Parkinson's disease, PINK1 and Parkin, encode a protein kinase and a ubiquitin-ligase, respectively. Both enzymes have been suggested to support mitochondrial quality control. We have reported that Parkin is phosphorylated at Ser65 within the ubiquitin-like domain by PINK1 in mammalian cultured cells. However, it remains unclear whether Parkin phosphorylation is involved in mitochondrial maintenance and activity of dopaminergic neurons in vivo. Here, we examined the effects of Parkin phosphorylation in Drosophila, in which the phosphorylation residue is conserved at Ser94. Morphological changes of mitochondria caused by the ectopic expression of wild-type Parkin in muscle tissue and brain dopaminergic neurons disappeared in the absence of PINK1. In contrast, phosphomimetic Parkin accelerated mitochondrial fragmentation or aggregation and the degradation of mitochondrial proteins regardless of PINK1 activity, suggesting that the phosphorylation of Parkin boosts its ubiquitin-ligase activity. A non-phosphorylated form of Parkin fully rescued the muscular mitochondrial degeneration due to the loss of PINK1 activity, whereas the introduction of the non-phosphorylated Parkin mutant in Parkin-null flies led to the emergence of abnormally fused mitochondria in the muscle tissue. Manipulating the Parkin phosphorylation status affected spontaneous dopamine release in the nerve terminals of dopaminergic neurons, the survivability of dopaminergic neurons and flight activity. Our data reveal that Parkin phosphorylation regulates not only mitochondrial function but also the neuronal activity of dopaminergic neurons in vivo, suggesting that the appropriate regulation of Parkin phosphorylation is important for muscular and dopaminergic functions.
| Parkinson's disease is a neurodegenerative disorder caused by degeneration of the midbrain dopaminergic system in addition to other nervous systems. PINK1 and parkin, which encode protein kinase and ubiquitin-ligase, respectively, were identified as the genes responsible for the autosomal recessive form of juvenile Parkinson's disease. These two enzymes are involved in mitochondrial maintenance. Although we previously found that Parkin is phosphorylated by PINK1 in mammalian cultured cells, the physiological significance of this interaction in vivo remained unclear. Here, we describe that the phosphorylation of Parkin altered mitochondrial morphology and function in muscle tissue through the degradation of mitochondrial GTPase proteins (such as Mitofusin and Miro) and a mitochondrial respiratory complex I subunit by increasing its ubiquitin-ligase activity. We also found that the dopaminergic expression of both constitutively phosphorylated and non-phosphorylated forms of Parkin affects the flight activity and shortens the lifespan of flies, suggesting that the appropriate phosphorylation of Parkin is important for both dopaminergic activity and the survival of dopaminergic neurons.
| Mutations of the Parkin and PINK1 genes cause selective degeneration of midbrain dopaminergic neurons in early-onset Parkinson's disease (PD) [1], [2]. The Parkin and PINK1 genes encode a cytosolic ubiquitin-ligase [3]–[5] and a mitochondrial serine/threonine kinase [6], respectively. Loss of the Parkin or PINK1 genes in Drosophila results in degeneration of mitochondria with high energy demands, such as those in muscle and sperm cells [7], [8], and Drosophila epistasis analysis has revealed that PINK1 acts upstream of Parkin [9]–[11].
Cell biological studies have demonstrated that Parkin, in cooperation with PINK1, clears damaged mitochondria by utilizing autophagy machinery in a process known as mitophagy [12]–[16]. Reduction of the mitochondrial membrane potential (ΔΨm) leads to accumulation and activation of PINK1 in the mitochondria [15], [17], which recruits Parkin from the cytosol to the mitochondria and activates the ubiquitin-ligase activity of Parkin. Parkin translocates to the mitochondria, where it ubiquitinates and degrades mitochondrial proteins, such as Mitofusin (Mfn) [18], [19] and Miro [20], [21], via the proteasome pathway. These events are thought to reorganize the mitochondrial network and stimulate the recruitment of autophagy machinery. PINK1-dependent recruitment of Parkin to the mitochondria is believed to be the first step of mitophagy [12]–[16]. We have reported that PINK1 phosphorylates Parkin at Ser65 in the ubiquitin-like (Ubl) domain during Parkin translocation, which appears to regulate mitophagy in cultured cells [22]. However, it remains unclear whether Parkin phosphorylation by PINK1 contributes to mitochondrial maintenance and activity of dopaminergic (DA) neurons in vivo.
To address this issue, we generated transgenic flies harboring phospho-mutant forms of Drosophila Parkin, the phosphorylation site of which is conserved. Transgenic expression of phospho-mutant forms of Parkin in PINK1 or Parkin mutant flies suggests that Parkin phosphorylation by PINK1 enhances the ubiquitin-ligase (E3) activity of Parkin. Our data also provide evidence that overactivation of Parkin by constitutive phosphorylation could lead to tissue dysfunction caused by mitochondrial degeneration whereas absence of Parkin phosphorylation affects DA neuronal activity, leading to the hypothesis that PINK1 is responsible for fine-tuning Parkin activity.
We and Kondapalli et al. have reported that Ser65 in the human Parkin Ubl domain is phosphorylated by PINK1, which is activated by reduction of ΔΨm in cultured cells [22], [23]. The amino acid sequence of the phosphorylation site of Drosophila Parkin appears to be conserved [22]. Phos-tag western blotting of Drosophila Parkin revealed bands representing PINK1-dependent phosphorylation of Parkin when wild-type (WT) Parkin and Drosophila PINK1 were co-transfected into Drosophila S2 cells. Introduction of a non-phosphomutated Ser94Ala (SA, corresponding to Ser65Ala in humans) Parkin abolished the phosphorylation bands (Figure 1A, right). The phosphorylation shifts did not occur when ΔΨm was simply disrupted, most likely because of the detection limit of Parkin phosphorylation under this experimental condition (Figure 1A). Next, we generated flies harboring transgenes encoding WT Parkin or non-phospho SA or phospho-mimetic Ser94Glu (SE) mutants. Using the ubiquitous Da or eye-specific GMR driver of the GAL4-UAS system, we chose at least two independent lines expressing Parkin protein at similar levels in each genotype, and we observed a ∼9-fold increase in Parkin expression relative to endogenous Parkin (data not shown). Because different lines of the same genotype showed similar results, we have presented representative data from each genotype.
Mitochondria of the indirect flight muscles (IFMs) are prominently affected in flies lacking Parkin [7] or PINK1 activity [9]–[11]. Visualization of mitochondria using mitochondrially targeted GFP (mitoGFP) or transmission electron microscopy analysis revealed that expression of WT Parkin in IFMs shortened the mitochondria in the direction of the long axis compared with a normal control as reported previously [24], and Parkin SE expression resulted in the over-fragmentation of mitochondria (Figure 1B, 1D). In contrast, SA Parkin expression had minor effects on mitochondrial length, although the mitochondrial morphology was less uniform (Figure 1B, 1D).
The abnormally large, fused mitochondria observed in PINK1 knockdown or null flies completely disappeared after introduction of WT or SA Parkin (Figure 1C). Intriguingly, the length of the muscular mitochondria, which was reduced by WT Parkin but not SA Parkin expression in the wild-type genetic background, was increased in the PINK1 knockdown background (Figure 1D). However, mitochondria of PINK1-deficient flies expressing SE Parkin were over-fragmented, as observed in flies expressing SE Parkin with endogenous PINK1 activity (i.e., in a wild-type genetic background; Figure 1D). We next examined whether the lack of Parkin phosphorylation alters mitochondrial function when endogenous Parkin activity is removed (Figure 1E). Spotty, large mitochondria labeled with mitoGFP were observed in Parkin-deficient flies (Figure 1E, upper), and the wild-type phenotype was recovered by WT Parkin expression (Figure 1E, middle). In contrast, some abnormally fused mitochondria similar to those in PINK1 mutant flies were observed when SA Parkin was expressed in Parkin-deficient flies (Figure 1E, lower). Similarly, SA Parkin failed to suppress the mitochondrial elongation of Parkin-deficient flies (Figure 1F).
Consistent with the results obtained for mitochondrial morphology, as shown in Figure 1, levels of the Drosophila mitochondrial outer membrane proteins Mfn and Miro, which regulate mitochondrial morphology and motility and are ubiquitination substrates of Parkin [18], [20], [21], [25], [26], were reduced in SE Parkin-expressing muscles (Figure 2A, 2B). Levels of the mitochondrial complex I subunit NDUFS3 were also reduced in SE Parkin-expressing muscles (Figure 2C). WT, but not SA, Parkin expression reduced Mfn levels to a milder extent than SE Parkin expression. These results suggest that SE Parkin has more potent E3 activity than WT Parkin, whereas SA Parkin has less activity than WT Parkin.
Expression of an unrelated protein, β-galactosidase (LacZ), in PINK1-deficient flies failed to rescue the mitochondrial phenotype, and reduction of NDUFS3 levels was observed (Figure 2D, 2F). Introduction of WT or SA Parkin in PINK1-deficient flies maintained the NDUFS3 level at the level of the control (Figure 2D, 2F), and the levels of Miro were comparable to those of the control. Mfn levels tended to decrease, although the changes were not statistically significant (Figure 2D, 2E). Expression of SE Parkin in the absence of PINK1 produced results similar to those observed when SE Parkin was expressed in the presence of PINK1 (Figure 2D–F compared with Figure 2A–C). The amounts of Mfn and Miro tended to increase in aged PINK1-deficient flies; however, these increases did not reach statistical significance when compared with levels in normal flies expressing β-galactosidase (PINK1-/-; LacZ vs. control in Figure S1). The effects of ectopic expression of Parkin in aged PINK1-deficient flies were similar to those in young flies (Figure 2 compared with Figure S1) although mild reduction of NDUFS3 levels was observed in PINK1-deficient flies expressing SE Parkin with age (Figure S2). Inducible expression of three kinds of Parkin for 3 days after eclosion at various levels also supported the finding that E3 activities of SE Parkin and SA Parkin against Mfn are increased and decreased, respectively (Figure 2G, 2H). In contrast, NDUFS3 levels were not affected in this short duration of Parkin expression (Figure 2G). There was not a good correlation between a mitochondrial fission factor Drp1 level and Parkin E3 activity, suggesting that the morphological change of mitochondria by Parkin is not due to increased expression of Drp1 (Figure 2G). These biochemical results suggest that SA Parkin has lower E3 activity than WT in the presence of endogenous PINK1, whereas SE Parkin has potent E3 activity regardless of PINK1 expression. The results also indicate that an appropriate level of Parkin phosphorylation is required for preservation of mitochondrial complex I and Parkin substrates.
Functional disturbance of the mitochondria was estimated by measurement of ATP levels in muscle tissues. Age-dependent reduction of ATP content was observed in muscle tissues expressing SE Parkin (Figure 3A, 3B) and in tissues of PINK1 null flies (PINK1-/-; LacZ, Figure 3B). The ATP reduction by SE Parkin expression was further exacerbated in the PINK1 null genetic background (p<0.05, SE Parkin vs. PINK1-/-; SE Parkin, Figure 3B). Expression of WT or SA Parkin returned the ATP levels of 40-day-old PINK1 null flies to normal control levels (not significant, PINK1-/-; WT or SA Parkin vs. PINK1+/+; LacZ, Figure 3B). We observed that the thorax muscles of PINK1-deficient flies (PINK1-/-; LacZ) and SE Parkin flies with or without PINK1 (i.e., PINK1+/+; SE Parkin and PINK1-/-; SE Parkin) became very fragile when we dissected the tissues. Aged SE Parkin flies presented a slight tendency to lose soluble tissue proteins (Figure 3C), and a significant reduction in tissue proteins was detected in aged PINK1 null flies expressing SE Parkin (Figure 3D). Because PINK1 regulates calcium efflux from the mitochondria via the mitochondrial Na+/Ca2+ exchanger [27], and is involved in actin dynamics through TORC2-Tricornered kinase pathway [28], Parkin-independent functions of PINK1 might partially contribute to the prevention of ATP shortage and tissue protein loss. Similar results with respect to ATP and protein levels were obtained when we expressed WT and mutant forms of Parkin in PINK1 knockdown flies, indicating that our PINK1 knockdown line faithfully recapitulates the PINK1-null phenotype (Figure S3). Consistent with the results of ATP production, SE Parkin expression impaired the respiratory complex I activity and failed to rescue the complex I dysfunction by loss of PINK1 (Figure 3E, F). The activity of citrate synthase, a key enzyme in the Krebs cycle, was reduced in PINK1 null flies (p<0.01, PINK1-/-; LacZ vs. PINK1+/+; LacZ), which was rescued by WT and SA Parkin, but not SE Parkin (Figure 3E, F).
Muscular expression of both WT and SA Parkin improved the age-dependent defect in climbing ability, although WT Parkin expression resulted in a somewhat worse performance than SA in 7-day-old flies (Figure 4A). SE Parkin expression worsened climbing ability compared with a LacZ control. When expressed in the PINK1 null background, both WT and SA Parkin suppressed the motor defects of PINK1-deficient flies (Figure 4B). However, SE Parkin failed to rescue motor behavior after 7-day-old trial (Figure 4B). These data show that constitutive phosphorylation of Parkin is deleterious to mitochondrial function and motor activity.
Mitochondrial degeneration caused by Parkin loss also caused a reduction in ATP content in thoracic muscle tissues. We ubiquitously expressed WT Parkin and mutant forms of Parkin in Parkin-null flies using the Da driver. Ubiquitous expression of SE Parkin caused lethality, even in the Parkin-null genetic background. Introduction of WT or SA Parkin greatly improved ATP production to a level even higher than that of the normal control (Figure S4A). Similarly, the loss of tissue proteins resulting from Parkin loss was suppressed by expression of WT or SA Parkin (Figure S4B). Mfn and Miro were obviously accumulated in 30-day-old Parkin null flies, and the accumulation was prevented by expression of WT or SA Parkin (Figure S4C, S4D). Protein levels of the respiratory complex NDUFS3 were reduced, whereas levels of ATP5A were unchanged in Parkin-null flies (Figure S4C, S4E). Mitochondrial Hsp60 was extensively accumulated in Parkin null flies, most likely as part of a compensatory mechanism against mitochondrial stress (Figure S4C, S4F) [29]. The alterations in NDUFS3 and Hsp60 levels were also suppressed by both WT and SA Parkin. These results indicate that Parkin phosphorylation is not necessary for mitochondrial function, although an increase in fused mitochondria was observed in Parkin null flies expressing SA Parkin (Figure 1F).
Abnormal wing postures and dents in the thorax caused by degeneration of mitochondria in IFMs are noticeable phenotypes of flies lacking PINK1 or Parkin activity. Expression of both WT and SA Parkin completely suppressed these phenotypes (Figure 5A). When endogenous Parkin activity was abolished, SA Parkin suppressed the formation of abnormal wing and thorax phenotypes to a lesser extent than WT Parkin (Figure 5B).
Ubiquitous expression of SE Parkin using the Da driver caused lethality, and muscle-specific expression of SE Parkin using the MHC driver produced an abnormal wing posture similar to that of flies lacking PINK1 or Parkin (Figure 5C, 5D). Although the abnormal wing phenotype was observed even in young adult flies, elimination of PINK1 activity counteracted it, and this effect was weakened by aging (Figure 5C, 5D). Histochemical analyses revealed that expression of SE Parkin in PINK1-deficient flies preserved the internal structure of mitochondria (Figure 1D), and the defects in mitochondrial membrane integrity in PINK1-knockdown flies appeared to be suppressed by the expression of SE as well as WT Parkin throughout life despite the drooped wing phenotype that occurs with age (Figure 5E).
We next focused on the effects of Parkin phosphorylation in DA neurons, which are affected during PD pathogenesis in humans. Using the DA neuron-specific TH driver, WT Parkin and its phosphomutants were expressed in the DA neurons of the adult fly brain. The mitochondrial morphology of the DA neurons in 5-day-old adult flies was analyzed by visualizing mitochondria using mitoGFP. The mitochondria formed tubular networks with several small spherical bodies within the cell bodies of tyrosine hydroxylase (TH)-positive DA neurons of normal control flies (Figure 6A), and many mitochondrial signals were observed outside the cell bodies, which likely represented axonal and dendritic mitochondria transported from the cell bodies of DA neurons (Figure 6A′). Expression of WT Parkin caused a reduction in the number of tubular mitochondria in the cell bodies (Figure 6B), which is consistent with the previous finding that overexpression of PINK1 or Parkin promotes spherical clustering of mitochondria in DA neurons [30], and led to the disappearance of mitochondria outside the cell bodies (Figure 6B′), suggesting that Miro-dependent mitochondrial transport was disturbed. Mitochondrial distribution and morphology in SA Parkin-expressing TH-positive neurons were similar to those in the normal control (Figure 6C, 6C′). Expression of SE Parkin further enhanced the effects of WT Parkin, whereby a single large aggregate of mitochondria appeared in each cell body, and the peripheral mitochondria disappeared (Figure 6D, 6D′).
Given that Parkin is phosphorylated by PINK1, changes in mitochondrial morphology and distribution should be affected in the absence of PINK1. We next examined the mitochondrial morphology of TH-positive neurons in the PINK1-null genetic background (Figure 6E-H′), and we compared this morphology with that in the wild-type genetic background (Figure 6A-D′). As previously reported, large aggregates of mitochondria were frequently observed in the cell bodies of flies lacking PINK1 activity (Figure 6E) [10], [30], and mitochondria outside the cell bodies were also observed (Figure 6E′). Introduction of SA and WT Parkin in the absence of PINK1 rescued the phenotype of mitochondrial aggregation and restored the normal mitochondrial morphology (Figure 6F-G′ compared with Figure 6E, 6E′). In TH-positive neurons expressing WT Parkin with or without PINK1, the lack of PINK1 suppressed the mitochondrial aggregation and distribution defects caused by ectopic expression of WT Parkin (Figure 6F, 6F′ compared with Figure 6B, 6B′). In sharp contrast, PINK1 activity did not affect the morphological changes caused by SA Parkin (Figure 6C, 6C′ compared with Figure 6G, 6G′). The effects of SE Parkin expression were similar regardless of PINK1 activity (Figure 6H, 6H′ compared with Figure 6D, 6D′). Taken together, these results indicate that Parkin activity is regulated by PINK1-mediated phosphorylation of the Ubl domain, even under physiological conditions under which PINK1 is not thought to be activated. The total length of tubular mitochondria and the average size of aggregated mitochondria greater than 3 µm2 in the cell bodies of TH-positive neurons of each genotype are summarized in Figure 6I and 6J.
We next estimated the presynaptic activity of DA neurons expressing mutant forms of Parkin using VMAT-pHluorin, a pH-sensitive form of GFP-conjugated VMAT, to visualize the release of DA [31]. The total expression levels of VMAT-pHluorin were estimated using whole brain samples treated with fixative solution to disrupt the acidic conditions of the synaptic vesicle lumen. In LacZ-expressing DA neurons, VMAT-pHluorin signals were observed in association with DA neuron terminals in the fly brain, including mushroom bodies and the fan-shaped body, as previously reported (Figure 7A) [31]. The localization signals of VMAT-pHluorin in DA neurons expressing WT and SA Parkin were similar to those of LacZ (Figure 7A). In contrast, the VMAT-pHluorin signal was reduced in the SE Parkin-expressing DA neurons (Figure 7A). We next estimated the spontaneous vesicle fusion occurring at the DA neuron terminals using brain tissue cultures expressing VMAT-pHluorin (Figure 7B). Fluorescence recovery after photobleaching (FRAP) was analyzed to estimate the spontaneous neuronal activity. In LacZ and WT Parkin-expressing DA neuron terminals, the fluorescence intensity was recovered to 14–17% of baseline in 9 min. In contrast, the fluorescence recovery in the nerve terminals of DA neurons of SA and SE Parkin flies was reduced compared with that of WT Parkin flies, with only 5% and 10% being recovered, respectively.
The changes in expression of VMAT-pHluorin and dopamine release of the DA nerve terminals in the presence of phospho-mutant forms of Parkin prompted us to test whether PD-associated behaviors could be affected by the status of Parkin phosphorylation. Taking advantage of a startle-induced negative geotaxis, climbing behavior is often investigated in Drosophila PD models. However, we found that Parkin expression in DA neurons or throughout the body causes restless behavior, leading the results to not accurately reflect the motor activity. In contrast, flying ability was markedly impaired when SA and SE but not WT Parkin were expressed in DA neurons (Figure 7C and Movie S1), implying that the phosphorylation of Parkin might regulate DA neuronal activity for the motor coordination of flight behavior.
Ectopic expression of WT and SA Parkin using the TH driver resulted in an age-dependent loss of DA neurons (Figure 7D). SE Parkin had a more toxic effect; the loss of neurons was detected even in young adult 5-day-old flies (Figure 7D). DA neuronal expression of WT Parkin was less toxic in the absence of PINK1 activity, and SE Parkin expression without PINK1 had a strong neurotoxicity similar to the effect observed with PINK1 (Figure 7E). However, SA Parkin showed more neurotoxicity than did WT Parkin in the absence of PINK1, implying a more complicated molecular mechanism of Parkin regulation in neurons than in muscle tissue (see Discussion).
Finally we investigated whether phospho-mutant forms of Parkin affect lifespan. Loss of PINK1 shortened the lifespan of Drosophila; however, this effect was reversed by muscle-specific expression of WT Parkin. In this setting, both SA and SE Parkin fully restored longevity (Figure 7F). In contrast, DA neuronal expression of both SA and SE Parkin shortened the lifespan compared with that of LacZ and WT Parkin (Figure 7G). These results suggest that the appropriate phosphorylation of Parkin is important for neuronal activity and survival. We summarize our findings using fly models in Table 1.
We have found that Ser65 in the Ubl domain of human Parkin is phosphorylated by PINK1 upon the reduction of ΔΨm in cultured cells, which appears to be required for mitochondrial translocation of Parkin and degradation of Parkin substrates [22]. Iguchi et al. found that phosphorylation of Ser65 is required for ubiquitin-thioester formation with Parkin Cys431, suggesting that this phosphorylation is required for activation of Parkin E3 activity [32]. Consistent with this idea, manipulation of the Drosophila Parkin Ser94, which corresponds to the human Parkin Ser65, altered the stability of the known Parkin substrates Mfn and Miro, leading to morphological changes in muscular mitochondria.
We also observed a decrease in the respiratory complex I subunit NDUFS3 upon expression of phospho-mimetic forms of Parkin. A recent Drosophila study demonstrated that the PINK1 and Parkin pathway selectively promotes the turnover of respiratory complex proteins [33]. A reduction in the NDUFS3 level was also observed in PINK1- or Parkin-deficient flies, which indicates that the integrity of respiratory complex I is maintained by a fine balance between PINK1 and Parkin activity. In support of this notion, the ATP production of muscular mitochondria in aged flies expressing SE Parkin was reduced by approximately 40% of a control. Although the motor activity and mitochondrial phenotypes in SE Parkin-expressing flies closely resemble those of PINK1- or Parkin-deficient flies, a noticeable difference was observed when SE Parkin was expressed in the PINK1-deficient genetic background. The drooped-wing phenotype caused by SE Parkin was considerably improved in the absence of PINK1 in young flies. Our previous results indicated that Ser65 in human Parkin is the sole phosphorylation site utilized by PINK1. Similarly, the phos-tag western blot performed in this study also suggests that Drosophila Parkin Ser94 is the only phosphorylation site utilized by PINK1. Given that the E3 activity of SE Parkin is equivalent regardless of PINK1 activity, an unknown Parkin regulator(s) is likely modulated by PINK1. Although we cannot rule out the possibility of endogenous Parkin contribution, we prefer this idea, which is supported by our previous finding that Parkin phosphorylation is not sufficient to trigger mitochondrial translocation of Parkin in mammalian cultured cells [22]. Because activated Parkin is preferentially degraded by the proteasome, marked reduction of SE Parkin expression in the wild-type but not PINK1-deficient genetic background also suggests that SE Parkin is more active in the presence of PINK1 (Figure 2A, 2D) [22]. However, the abnormal wing phenotype caused by SE Parkin eventually emerged in aged PINK1-deficient flies expressing SE Parkin, which suggests that constitutive expression of SE Parkin overwhelmed the function of the PINK1-dependent regulator(s).
Introduction of SA Parkin did not fully restore mitochondrial morphology and the abnormal wing posture in Parkin-deficient flies; however, SA Parkin rescued all mitochondrial phenotypes and behavioral defects in PINK1-deficient flies. This finding indicates the possibility that endogenous Parkin converts non-phosphorylated latent Parkin into an active form of Parkin, as a study by Lazarou et al. demonstrated that Parkin oligomerizes on the mitochondria upon activation [34]. However, co-expression of SA and WT Parkin failed to ensure that ubiquitin was loaded at the catalytic residue of SA Parkin in the cultured cells, suggesting that endogenous Parkin does not activate the E3 activity of SA Parkin in trans (Figure S5, lane 6). Thus, a molecular explanation for this observation has yet to be provided.
Expression of Parkin phospho-mutants in the presence or absence of PINK1 activity in DA neurons revealed that Parkin phosphorylation by PINK1 is required for regulation of mitochondrial morphology and motility in DA neurons. Our group and Wang et al. have proposed a model in which PINK1 and Parkin prevent damaged mitochondria from moving to the nerve terminals by degrading Miro, an adaptor for microtubule-dependent mitochondrial transport [20], [21]. Supporting this notion, ectopic expression of WT Parkin in DA neurons suppressed mitochondrial distribution outside cell bodies, which was recovered in the PINK1-null genetic background. Regardless of the status of PINK1, SE Parkin enhanced the mitochondrial phenotypes observed when WT Parkin was expressed with endogenous PINK1 such that perinuclear accumulation of mitochondria was observed in the cell bodies. These findings indicate that phosphorylation of Parkin by PINK1 boosts its E3 activity, thus regulating mitochondrial motility and morphology through degradation of mitochondrial proteins, such as Miro and Mfn.
We also reveal that Parkin phosphorylation regulates neuronal activity, as our data indicate that spontaneous dopamine release in the nerve terminals and flying activity were compromised in the presence of both SA and SE Parkin expression. The impairment of dopamine release was reported in both Parkin-deficient mice [35]–[37] and PINK1-deficient mice [38]. The regulation of dopamine release may be independent from the mitochondrial function regulated by PINK1 and Parkin, as Parkin is implicated in the regulation of vesicle trafficking [39]. Our results indicate that both SA and SE Parkin impaired dopamine release, suggesting that the appropriate phosphorylation cycle of Parkin regulates spontaneous dopamine release independently from mitochondrial activity. SE Parkin compromised mitochondrial transport in DA neurons, leading to the perinuclear accumulation of mitochondria. In contrast, reduction of VMAT-pHluorin expression by SE Parkin is unlikely due to inhibition of axonal transport because we did not observe VMAT-pHluorin accumulation in the cell bodies. VMAT might be degraded by activated Parkin.
WT Parkin expression in DA neurons in a wild-type genetic background showed a more toxic effect than those in a PINK1-deficient genetic background (Figure 7D, 7E). This difference in neurotoxicity could be explained by the difference in the extent of Parkin phosphorylation by PINK1, leading to a difference in Parkin E3 activity. SA Parkin expression in DA neurons unexpectedly exhibited more neurotoxic activity than did WT Parkin in the PINK1-deficent flies (Figure 7E). This result may suggest the existence of a neuron-specific Parkin kinase(s) other than PINK1, although the effect of the kinase on Parkin phosphorylation appears to be smaller than that of PINK1. Tricornered/NDR kinase, which rescues the mitochondrial degeneration caused by the loss of PINK1 in Drosophila, could be a candidate of Parkin kinase [28]. Another possibility is that SA Parkin acts in a dominant-negative fashion, as demonstrated by a report that transgenic expression of pathogenic Parkin Q311X resulted in an age-dependent degeneration of DA neurons in the substantia nigra of mice, suggesting that mutant Parkin exerts dominant toxic effects in DA neurons [40].
Inhibition of DA release in the adult brain alters sleep behavior and age-dependent locomotor deficits, which might be associated with PD symptoms. While visual perception is largely maintained in adult flies lacking brain dopamine [41], expression of pathogenic LRRK2, a late–onset PD gene, by the TH-GAL4 driver resulted in non-autonomous visual neurodegeneration [42]. In another experimental setting, expression of pathogenic LRRK2 by the TH-GAL4 driver dramatically shortened the lifespan of Drosophila [43]. The above reports and our results may suggest that expression of mutant PD gene products including Parkin SA and SE in DA neurons does not only impair DA transmission but also leads to widespread neurodegeneration that affects lifespan non-cell-autonomously.
In summary, we have shown that Parkin phosphorylation by PINK1 drives Parkin E3 activity in vivo. Although cell culture studies suggest that PINK1 is inactivated by constitutive breakdown under steady-state conditions, in this study, we have used Drosophila models to reveal that endogenous PINK1 precisely controls Parkin activity to maintain the mitochondrial function in muscle tissue and the neuronal function in DA neurons. Our genetic study also suggests the presence of PINK1-regulating factor(s), which may be Parkin regulators. Identification of these unknown factor(s) will be pursued in a further study, and elucidation of the Parkin activation mechanism, including phosphorylation of the Ubl domain, and role of Parkin phosphorylation in neuronal activities will contribute to the identification of a potential therapeutic target in PD pathogenesis.
Fly culture and crosses were performed on standard fly food containing yeast, cornmeal and molasses, and the flies were raised at 25°C. The w1118 (w–) line was used as a wild-type genetic background. Complementary DNAs for Drosophila WT, S94A and S94E Parkin were subcloned into the pUAST vector, and UAS-Parkin WT, S94A and S94E transgenic lines were generated in the w– background. All other fly stocks and GAL4 lines used in this study were obtained from the Bloomington Drosophila Stock Center and have been previously described: UAS-PINK1 RNAi [11]; PINK1B9 [10]; park1, parkΔ21 [8]; daughterless–Gene-Switch [44]; and VMAT-pHluorin [31]. PINK1B9 and park1/parkΔ21 were used as PINK1-deficient and Parkin-deficient alleles, respectively.
Rabbit anti-Drosophila Parkin polyclonal antibody was raised against recombinant MBP-tagged Drosophila Parkin (275–482 aa) produced in the E. coli strain Rosetta 2 (Novagen). The antibodies used in the western blot analysis were as follows: anti-Parkin (1∶5,000 dilution), anti-Mfn (1∶2,000 dilution; a kind gift of Dr A. Whitworth), anti-Miro (1∶2,000 dilution; a kind gift of Dr E. Zinsmaier), anti-Drp1 (1∶2,000 dilution; a kind gift of Dr L. Pallanck), anti-ATP5A (1∶10,000 dilution; Abcam, 15H4C4), anti-NDUFS3 (1∶10,000 dilution; Abcam, 17D95), anti-Actin (1∶10,000 dilution; Millipore, MAb1501) and anti-Hsp60 (1∶1,000 dilution; Cell Signaling, D307). The antibody used in immunocytochemistry was anti-TH (1∶1,000 dilution; ref. [11]). Complementary DNAs for human and Drosophila Parkin and PINK1 were described in previous studies [11], [14], [45]. Parkin phospho-mutants were generated by PCR-based mutagenesis followed by sequence confirmation of the entire gene.
Fly heads and thoraxes were directly homogenized in 20 µl and 40 µl of 3x SDS sample buffer per head and thorax, respectively, using a motor-driven pestle. After centrifugation at 16,000×g for 10 min, the supernatants were subjected to western blotting. For Gene-Switch experiments, newly eclosing flies crossed with the daughterless–Gene-Switch driver were raised with media containing 5% glucose, 1% agarose and various concentrations of RU486 for 3 days, and their thorax samples were subjected to western blot analysis as described [24]. The band intensity was analyzed using ImageJ software.
S2 cells were cultured in Schneider's medium (Invitrogen) supplemented with 10% FCS (Invitrogen) and 1% penicillin-streptomycin. The cells were transfected using HilyMax reagent (Dojindo) following the manufacturer's instructions. After 24–48 h, the cells were collected and lysed in lysis buffer containing 0.2% NP-40, 50 mM Tris (pH 7.4), 150 mM NaCl and 10% glycerol supplemented with protease inhibitor (Roche Diagnostics) and phosphatase inhibitor (Pierce) cocktails. Phos-tag western blotting was performed as previously described [46].
The mitochondrial morphology of the indirect flight muscle and TH-positive neurons was analyzed by whole-mount immunostaining as described previously [46]. The length of the long axis of mitochondria was calculated using ImageJ software. TEM images were obtained at the Laboratory of Ultrastructural Research of Juntendo University.
Brain tissue samples from 3- to 5-day-old adult flies isolated in HL-3 solution (pH 7.5) [47] were mounted in HL-3 solution for FRAP analysis. Live imaging was performed using a Leica SP5 DM6000 confocal microscope equipped with a 40X oil immersion objective. A 488-nm argon laser applied at 100% power was used to photobleach the whole brain for 2 min. Images were taken as Z-stacks (3 µm slices) at 10% laser power every 3 min. The DA release rates of whole brains were calculated by normalizing against the fluorescent intensity just after photobleaching as the fluorescent recovery rate from 0 to 9 min.
For the lifespan studies, approximately 20 adult flies per vial were maintained at 25°C, transferred to fresh fly food and scored for survival every 2 days. To control for isogeny, the driver and PINK1B9 lines were backcrossed to the w- wild-type background for six generations. All UAS-Parkin transgenic flies were generated in the w- genetic background and thus have matched genetic backgrounds. The number of flies exhibiting defective, abnormal wing posture (held-up or drooped) was determined for each genotype [11]. For flight analysis, 25 control and experimental flies were placed in individual vials (9.3 cm height ×3.5 cm2 area), which were then gently tapped to bring the flies down to the vial bottoms. Flight events were counted for one minute at approximately 1 p.m., and the results of ten trials were subjected to averaging. A climbing assay was performed as described previously [48].
The ATP content in the thorax muscle was measured as described previously with some modifications [48]. Briefly, the thorax of an adult fly was dissected and homogenized in 20 µl of homogenization buffer (6 M guanidine-HCl, 100 mM Tris and 4 mM EDTA [pH 7.8]). After freezing, the samples were centrifuged at 16,000×g. The supernatant was diluted 1∶1000 with water for ATP measurement and 1∶10 for measurement of protein concentration. ATP and proteins were measured using the CellTiter-Glo luminescent cell viability assay kit (Promega) and the BCA protein assay kit (Pierce), respectively. Mitochondria isolation, citrate synthase activity assay and complex I activity assay were performed as described previously with some modifications [24]. Briefly, thoraxes were homogenized in mitochondrial isolation medium (250 mM sucrose, 10 mM Tris-HCl, pH 7.5, 0.15 mM MgCl2) on ice using a plastic pestle homogenizer, and centrifuged at 500×g for 5 min at 4°C. The pellet was resuspended in 50 µl mitochondrial isolation medium per fly. Mitochondrial suspension (5 µl each) was used for citrate synthase activity and complex I activity assays. Complex I activity was calculated as values from which values with 2 µl rotenone were subtracted, normalizing to protein concentrations.
A one-way repeated measures analysis of variance (ANOVA) was used to determine significant differences among multiple groups unless otherwise indicated. If a significant result was achieved (p<0.05), the mean values of the control and the specific test group were analyzed using Tukey-Kramer tests.
|
10.1371/journal.pbio.1001571 | Dual Host-Virus Arms Races Shape an Essential Housekeeping Protein | Transferrin Receptor (TfR1) is the cell-surface receptor that regulates iron uptake into cells, a process that is fundamental to life. However, TfR1 also facilitates the cellular entry of multiple mammalian viruses. We use evolutionary and functional analyses of TfR1 in the rodent clade, where two families of viruses bind this receptor, to mechanistically dissect how essential housekeeping genes like TFR1 successfully balance the opposing selective pressures exerted by host and virus. We find that while the sequence of rodent TfR1 is generally conserved, a small set of TfR1 residue positions has evolved rapidly over the speciation of rodents. Remarkably, all of these residues correspond to the two virus binding surfaces of TfR1. We show that naturally occurring mutations at these positions block virus entry while simultaneously preserving iron-uptake functionalities, both in rodent and human TfR1. Thus, by constantly replacing the amino acids encoded at just a few residue positions, TFR1 divorces adaptation to ever-changing viruses from preservation of key cellular functions. These dynamics have driven genetic divergence at the TFR1 locus that now enforces species-specific barriers to virus transmission, limiting both the cross-species and zoonotic transmission of these viruses.
| Genetic differences between mammalian species dictate the patterns of viral infection observed in nature. They also define how viruses must evolve in order to infect new mammalian hosts, giving rise to new and sometimes pandemic diseases. Because viruses must enter cells before they can replicate, new diseases often emerge when existing viruses evolve the ability to bind to the cell-surface receptor of a new species. At the same time, host cell receptors also evolve to counteract virus attacks. This back-and-forth evolution between virus and host can lead to an arms race that shapes the sequences of the proteins involved. In wild rodent populations, the retrovirus MMTV and New World arenaviruses both exploit Transferrin Receptor 1 (TfR1) to enter the cells of their hosts. Here we show that the physical interactions between these viruses and TfR1 have triggered evolutionary arms race dynamics that have directly modified the sequence of TfR1 and at least one of the viruses involved. Computational evolutionary analysis allowed us to identify specific residues in TfR1 that define patterns of viral infection in nature. The approach presented here can theoretically be applied to the study of any virus, through analysis of host genes known to be key to controlling viral infection. As such, this approach can expand our understanding of how viruses emerge from wildlife reservoirs, and how they drive the evolution of host genes.
| Transferrin receptor (TfR1) is the cell-surface receptor for iron-loaded transferrin circulating in the blood [1]. TfR1-transferrin complexes are internalized via clathrin-mediated endocytosis and iron is released in acidic endosomes. Besides transferrin, the other major binding partner of TfR1 is the hereditary hemochromatosis protein (HFE), which negatively regulates iron uptake. In addition to these host-beneficial interactions, three different families of viruses are known to interact with TfR1 to trigger their own cellular entry. TfR1 likely constitutes an attractive target for viruses because it is both ubiquitous and specifically up-regulated in rapidly dividing cells [1]. Because of the tremendous investment that has been made in understanding both TfR1 and the viruses that exploit it, there are rich structural and functional data available. For instance, co-crystal structures have been solved of human TfR1 in complex with both of its cellular iron-transport binding partners [2]–[4] and with the surface glycoprotein of a zoonotic rodent arenavirus, Machupo virus, which uses TfR1 for entry [5]. For this reason, TfR1 provides an ideal opportunity to investigate how cellular housekeeping proteins evolve to combat viruses that are exploiting them while simultaneously preserving critical cellular functions.
The entry of viruses into cells is often mediated by specific physical interactions between virus surface proteins and host-encoded cell surface receptors. In the case of the New World arenaviruses, the surface glycoprotein, GP, contacts TfR1 to trigger cellular entry [6]. These viruses infect various rodent species found in the Americas, and each virus has evolved compatibility with the particular TfR1 ortholog encoded by its host species (Figure 1A) [7]–[9]. Several of these viruses, including Junin virus, Machupo virus, and Guanarito virus, have acquired the ability to bind human TfR1 and are currently emerging into human populations through zoonotic transmission [10],[11]. These viruses cause hemorrhagic fevers in humans with case fatality rates of 15–30%, but fortunately, they do not yet spread from human to human efficiently enough to cause large epidemics. Another rodent virus that uses TfR1 for cellular entry is the retrovirus mouse mammary tumor virus (MMTV). The MMTV surface glycoprotein, Env, contacts TfR1 to trigger cellular entry [12]. MMTV infects Muridae rodents specifically of the genus Mus, including Mus musculus, the house mouse (Figure 1A). In contrast to the arenaviruses, MMTV is not known to infect other rodent species or humans. Incompatibility with human TfR1 appears to be the major cellular barrier to zoonosis because MMTV replicates robustly in human cells when receptor-mediated entry is bypassed by transfection of the viral genome directly into cells [13]–[15]. Finally, in carnivores, parvoviruses also bind TfR1 for cellular entry [16]. Canine parvovirus serves as one of the most important models for disease emergence in the wild, as this virus first came into existence in the 1970s when a virus was passed to dogs from another carnivore species [17]. This event centered around viral evolution for compatibility with the dog TfR1 ortholog [18],[19]. Thus, in all three of the virus families that use TfR1, existing evidence suggests that the ability to enter cells through the TfR1 ortholog of a particular species is a necessary criterion for infection in the wild, and that viral adaptation is often required to utilize the TfR1 of new species.
While infectious disease research has long focused on host antiviral proteins, host proteins that facilitate viral replication are now an exploding area of inquiry [20]. These proteins represent novel targets for the development of antiviral drugs because interruption of the interactions between virions and host proteins like TfR1 are predicted to block viral replication. In nature, evolution has utilized two paradigms for achieving this same goal. In some cases, host genes encoding pathogen entry receptors have accumulated promoter or other mutations that cause reduced or no expression of the receptor protein [21]–[27]. However, TFR1, given the essential nature of its housekeeping functions, would be unlikely to tolerate hypomorphic mutations. For retroviruses, host genomes are known to employ a second mechanism to block virus entry, one that exploits a unique property of the retroviral lifecycle. Unlike other viruses, retroviruses permanently integrate into the host genome during viral replication. If viral genomes become integrated in the host germline, they can be passed to future generations and inherited in a Mendelian fashion [28],[29]. In several instances, retroviral surface proteins (Envs) expressed from these integrated retroviral copies compete with exogenous viruses for receptor use [30]–[35]. Host genomes are presumably selected to keep these retroviral env open reading frames intact because they offer protection against infection by exogenous viruses that use the same receptor [28],[29],[36]. Given the critical role of TfR1 in iron homeostasis, there may be a fitness cost to competitive binding by genome-encoded copies of the retroviral Env. Indeed, there is no evidence for either of these models (hypomorphic mutations or competitive inhibition) in the TFR1 literature. How, then, do critical genes like TFR1 respond to virus-driven selective pressure?
Most of what is known about the evolutionary dynamics between host and virus genomes comes from studies of antiviral genes, particularly those encoding viral sensors. Viral sensors (also referred to as “pattern recognition receptors” or “restriction factors”) are host proteins like RIG-I and TRIM5α that recognize and destroy viruses that are attempting to replicate inside of host cells [37],[38]. Because these sensors can be so effective, viruses often encode proteins that antagonize them or their downstream executors [39],[40]. Host genomes are continually selected to encode sensors that better recognize viruses, and viruses are continually selected to evade or disrupt these sensors [41]–[50]. This ongoing evolutionary struggle is called a molecular “arms race” (reviewed in [51]–[53]). Arms races play out in the protein–protein interactions that exist between host and virus proteins, and they drive endless rounds of “positive selection” for mutations that alter these interactions. This results in the rapid evolution of both proteins (host and virus) engaged in the conflict. Indeed, host-encoded viral sensors are often exceptionally genetically divergent between species and diverse within species [41]–[50],[54]–[58]. As a result, such genes are appreciated as major genetic barriers to host switching by viruses in nature, because unique virus mutations are required to counteract the divergent viral sensors present in each new host species [43],[59]–[61].
Arms races have not traditionally been documented in important housekeeping genes. Here, we document recurrent positive selection in rodent TFR1 and demonstrate that both the protein sequence and the interaction specificities of this receptor are far from static. Using a small evolutionary dataset consisting of TFR1 gene sequences from only seven rodent species, we identify specific codons in TFR1 that have been repeatedly targeted by positive selection for amino acid replacement. We find that these rapidly evolving positions correlate to the surfaces on TfR1 that mediate interaction with the two rodent viruses that bind this receptor. We demonstrate experimentally that mutations at these specific receptor residues are potent at altering interactions with virions while not altering receptor expression or function. We show that this evolutionary scenario has driven genetic divergence at this receptor locus that now enforces species barriers to viral transmission. We address the implications of these findings for human TfR1 and identify a human SNP that conveys some protection against cellular entry of a zoonotic rodent arenavirus. Our study demonstrates that the influence of viral pathogens on mammalian genomes goes well beyond the shaping of antiviral genes, as we can now appreciate that even the sequence of important housekeeping genes can be shaped by unremitting antagonism by viruses. However, in this case, collateral damage to cellular functions must be carefully controlled as the evolutionary battle with viruses plays out.
We investigated the evolution of TFR1 in rodents, where two different virus families use this receptor for cellular entry. The type of selective pressure that has acted on a gene can be inferred from the pattern of mutations that it has accumulated over time [62],[63]. The rate at which mammalian genes accumulate amino acid–altering DNA mutations (dN; nonsynonymous mutations) is typically far slower than the rate at which they accumulate mutations that leave the amino acid unchanged (dS; synonymous mutations) [51]. This is because most amino acid–altering mutations are deleterious. This signature (dN/dS<<1) stands in contrast to the pattern that is observed when genes have experienced multiple rounds of positive selection for protein-altering mutations (dN/dS>1). However, in host-virus arms race situations, patterns of dN/dS>1 would not be expected throughout the entire length of a gene, but rather specifically in the codons correlating to the interaction interface between host and virus proteins (reviewed in [51],[52]). We used the codeml program in PAML [64] to analyze dN/dS ratios in codons in an alignment of TFR1 from seven rodent species, five of which are known host species for the New World arenaviruses or MMTV (Figure 1A). We found variable patterns of codon evolution in TFR1. For instance, in codon model M2a, maximum likelihood estimation indicates that 78% of codons are extremely conserved with dN/dS = 0.09, 19% evolve neutrally with dN/dS = 1, and 2.4% are under positive selection with dN/dS = 4.2. Codon models that allow a subset of codons to evolve under positive selection (dN/dS>1) fit the data significantly better than models where positive selection is not allowed (p<0.001; Table S1). Thus, while much of the protein sequence of TfR1 is extremely conserved, a small percentage of residue positions are rapidly evolving.
The crystal structure of the TfR1 ectodomain has been solved [65]. Six codons that correspond to residues in this structure were assigned to the dN/dS>1 site class with a high posterior probability: K205, L209, N215, S296, T569, and E575 (Table S1). While discontinuous on the linear polypeptide (Figure 1B), the residues corresponding to these codons are located on a single ridge trailing down the outer edge of each monomer of the human TfR1 dimer (red residues in Figure 1C). Remarkably, all of these sites map precisely to the two known virus-binding surfaces on TfR1. Three of these rapidly evolving residue positions (K205, L209, and N215) map to the arenavirus binding surface of TfR1 (gray residues in Figure 1C) [5]. The other three rapidly evolving residues (S296, T569, and E575) fall directly in the surface of TfR1 that binds MMTV (blue residues in Figure 1C) [13]. We hypothesized that rodent TFR1 is subject to not just one but two different host-virus arms races.
Arms races are predicted to drive positive selection in both the host and virus genes involved, so we next analyzed the gene encoding the arenavirus surface protein, GP, for signatures of positive selection. Because the co-crystal structure has been solved of the Machupo virus surface glycoprotein subunit GP1 in complex with TfR1, the specific residues on GP1 that interact with TfR1 are known (blue lines below protein schematic in Figure 2A). We analyzed an alignment of gp1 from 13 human and mouse isolates of Machupo virus (Figure 2B). In this alignment, 11 codons bear the signature of dN/dS>1 (red lines above diagram in Figure 2A and Table S2). Ten of these correspond to surface-exposed residues in the GP1 structure [66]. Strikingly, all 10 are located on the surface of GP1 that faces TfR1, and none fall on the opposite side of GP1 that faces the virion (Figure 2C). Four of the residues under positive selection directly contact TfR1, and the rest are located near residues that do (Figure 2C). Using a permutation test, we find that the 16 TfR1-binding residues of GP1 are significantly enriched for sites of positive selection (p<0.005). Like all virus surface proteins, GP1 will have also experienced selection for immune escape, a complication that makes signatures of dN/dS>1 more difficult to interpret in viral genes than in host genes. However, GP1 residues in direct contact with TfR1 are unlikely to successfully mutate for the purpose of immune escape during an active infection. An arms race between rodent TfR1 and arenavirus GP1 is thus supported by the rapid evolution of each partner in this interaction, specifically in residues that are known to mediate contact with the other.
In the TFR1 dataset analyzed, only one of the rodent species included is known to harbor MMTV in the wild (house mouse; Figure 1A). It was thus unclear why we detected positive selection in the MMTV binding surface of TfR1 with the rodent dataset that was used. We hypothesized that either the evolutionary signature in the MMTV binding surface of TfR1 was driven by something else, or that MMTV-like viruses once circulated more widely through rodent genera. We reasoned that if the latter hypothesis is true, “fossils” of these extinct viruses might be found in the form of endogenous retroviruses (ERVs) integrated into the genomes of their former host species. Indeed, we identified MMTV-like ERVs in the genomes of the brown rat (Rattus norvegicus) and the North American deer mouse (Peromyscus maniculatus) (Figure 3 and Figure S1). The full-length ERV identified in the deer mouse genome is particularly interesting because this rodent is in the same family as the arenavirus host species (Cricetidae; Figure 1A). These ERVs reveal that MMTV-like viruses once circulated more widely amongst rodents, supporting the model that rodent TFR1 may have experienced selection imposed by these viruses. Interestingly, MMTV appears to be a virus in retreat, with a shrinking host range. We cannot exclude the possibility that MMTV-like viruses still infect other rodent species and have simply not been identified, but such viruses have not been reported in the literature or in GenBank [67], and are absent from large metagenomic surveys of rodent feces [68]. These MMTV ERVs are thus reminiscent of the many ERV families found in the human genome, none of which currently circulate in infectious form [28].
Based on these findings, TfR1 may have experienced high levels of sequence divergence on the MMTV-binding surface due to selection for mutations that blocked entry by these MMTV-like viruses. Consistent with this, we find that TfR1 orthologs from three different Cricetidae species are highly recalcitrant to entry by MMTV (Figure 4), even though this rodent family appears to once have harbored a similar virus. In an arms race between TfR1 and MMTV, the MMTV Env should also be evolving in response to the evolution of TfR1. Compared to Machupo virus GP1, far less is known about the amino acids in MMTV Env that bind to TfR1, as there is no co-crystal structure of Env in complex with TfR1. However, a five amino acid receptor binding motif in MMTV Env has been identified [69]. We find that this motif has a distinct protein sequence depending on the particular rodent host species from which each virus was isolated (Figure S2), consistent with viruses having uniquely evolved compatibility with each host TfR1 (before they potentially went extinct). An incomplete understanding of receptor binding determinants in MMTV Env, and the fact that most of these viruses now exist as endogenous copies, make it difficult to draw specific conclusions about the evolution of MMTV Env. Nonetheless, an arms race between TfR1 and MMTV is supported by the rapid evolution of residues on the MMTV-interaction surface of TfR1, the discovery that MMTV-like viruses once infected rodents more broadly providing a model for what drove this selection, and the observation that several Cricetidae TfR1 in their current form do not support MMTV entry, suggesting that they could have been selected for this property.
To test this MMTV resistance hypothesis further, we simulated the evolution of an MMTV-resistant receptor by mutating only the residue positions under positive selection in the MMTV binding surface (Figure 5A). We mutated the TfR1 of house mouse, the MMTV host, so that these three positions now encode the amino acids found in the TfR1 of the vesper mouse, which is not susceptible to MMTV. MDCK (dog) cells were transduced to stably express the mutant or wild-type TfR1 protein. These cells were chosen because dog TfR1 does not support entry by arenaviruses [9] or MMTV [13]. An extracellular FLAG tag was added to each receptor so that cell surface expression could be monitored on live cells by flow cytometry. We then measured the cellular entry of GFP-encoding retroviral vectors expressing the MMTV Env on their surface (MMTV pseudoviruses). Indeed, the three mutations in house mouse TfR1 almost completely abolished the entry of MMTV into cells (Welch t-test, p<0.0001, one-tailed; Figure 5B) without significantly altering receptor cell surface expression (Figure 5C). None of the sites of positive selection that we identified are found near the dimerization domain of TfR1, the region known to be most important for interaction with iron-transport binding partners (Figure 6A,B) [2]–[4],[70]. We confirmed that these mutations indeed do not alter transferrin binding (Figure 6C,D). Thus, amino acid substitutions at these sites in TfR1 can block virus entry without deleterious consequences to surface expression or receptor function, providing a clear hypothesis for why they might have a strong selectable advantage in MMTV-infected rodent populations.
If positively selected residues are key modulators of virus compatibility, we reasoned that mutations at these sites should also render MMTV-resistant TfR1s susceptible to MMTV entry. Because species divergence can lead to subtle structural differences in receptors, creating a gain-of-function phenotype with just three amino acid changes should be substantially more difficult than creating a loss of function phenotype in a receptor where virus-binding is currently intact. Nonetheless, mutating the three positively selected residues in the MMTV binding surface of zygodont TfR1 to match the corresponding residues found in TfR1 of house mouse (the MMTV host) led to a significant increase in MMTV entry (Welch t-test, p = 0.008, one-tailed; Figure 5D) without enhancing cell-surface expression (Figure 5E), transferrin binding (Figure 6C,D), or entry of three arenaviruses (Figure 5F). Thus, we have shown that swapping amino acids encoded at positively selected sites can swap virus-susceptibility phenotypes of TfR1 in both a gain-of-function and loss-of-function manner. Mutations at just three residue positions acutely regulate virus entry while preserving receptor expression and transferrin binding for the host.
Every round of positive selection of the rodent TFR1 gene began with a random mutation that arose in a single rodent individual. If this mutation offered protection against virus entry while not otherwise causing major fitness defects related to iron homeostasis, it would have been favored by natural selection and would have become more common or even fixed in the population where it arose. Because the New World arenaviruses are currently emerging into human populations, they are now beginning to exert selective pressure on the human population as well. For instance, there have been approximately 30,000 cases of Argentine hemorrhagic fever caused by the Junin virus since the 1950s, with a case fatality rate of 20% [11]. The geographic region at risk for this disease is expanding into north-central Argentina, and currently includes an area populated by around 5 million people [11]. Individuals with genotypes that make them less susceptible to infection or severe illness are expected to survive with bias over other individuals. This selection would intensify as the frequency or severity of the disease increases. In such cases, natural selection would be expected to act at any genetic locus where functionally distinct alleles exist within the human population. We wished to investigate whether TFR1 may be one such locus.
TfR1 interacts with arenaviruses and MMTV through distinct interaction surfaces (Figure 1C). TfR1 is 760 amino acids long, but a small stretch of nine residues from 204 to 212 is the major determinant of species-specificity for arenavirus entry (colored yellow in Figure 7A). These residues span two beta strands and the intervening loop (βII-1–βII-2). Two of the sites of positive selection (residues 205 and 209) fall in this stretch of nine residues, and the third (residue 215) falls three amino acids away (colored red in Figure 7A). As we demonstrated for the sites under positive selection in the MMTV binding surface, the introduction of amino acids from different rodent species at positions in this stretch has been previously shown to alter patterns of virus compatibility [7],[8]. Additionally, substitution of rodent-encoded amino acids at these residues can convert human TfR1 into an entry receptor for currently non-zoonotic rodent arenaviruses [5],[8]. By querying SNP databases, we identified a human SNP located in this structural feature, L212V (colored blue in Figure 7A). Because of the localization of this SNP near the residues under positive selection, we hypothesized that the L212V human polymorphism might affect arenavirus entry.
To test this, we again focused on Machupo virus. We constructed stable cell lines that express either human 212L or 212V TfR1. In the context of MDCK cells, dog TfR1 does not allow entry by Machupo virus, so the expression of either human allele allows more virus entry than is observed in MDCK cells alone (Figure 7B). However, the minor TfR1 212V variant supports about half the level of entry as seen with TfR1 212L (Figure 7B). Valine at position 212 may lead to a modest decrease in binding affinity with GP1 due to loss of a hydrophobic contact, based on the observation that two residues of Machupo GP1 (Phe226 and Pro223) are in van der Waals contact with TfR1 Leu212 [5]. We next stably expressed the human 212V and 212L TFR1 alleles in human cell lines that are themselves homozygous for 212L: HEK293 (kidney) and HEL299 (lung). Lung cells are especially relevant since arenaviruses are transmitted to humans through respiratory inhalation. In both cases, expression of the minor 212V allele was again protective against virus entry compared to the wild-type allele (Figure 7C,D). Thus, we have identified a SNP (L212V) that conveys some protection against arenavirus entry, at least in vitro.
The L212V SNP has only been reported in Asian populations (Chinese and Japanese), while TfR1-utilizing arenaviruses have only been found in the Americas. We sequenced TFR1 from 18 indigenous Central and South American individuals, but identified no instances of this polymorphism. Like all SNPs, this SNP arose randomly and may have no fitness advantage or disadvantage in the Asian populations where it is found, since TfR1-utilizing arenaviruses are not found in that part of the world. Nonetheless, this SNP could begin to experience selection if the rodent populations that carry these viruses were introduced into Asia, if these arenaviruses ever evolved to spread efficiently from human to human, or in the event of an intentional release of these viruses [71]. The data shown in Figure 7C,D indicate that protective TFR1 alleles can act in a semidominant fashion with regards to virus entry, because the human cells used in these experiments also express wild-type TfR1. We speculate that this occurs either because mutant and wild-type TfR1 proteins are forming heterodimers with one another, or because expression of a second allele that is functional for iron-uptake results in lower levels of wild-type TfR1 (TFR1 expression levels are tightly regulated for the purpose of maintaining iron homeostasis [10]). Either model would also be relevant in heterozygous individuals, suggesting that selection could act on SNPs conveying protection against viral entry even when they are rare and found predominantly in heterozygotes.
In this study we show that the protein sequence and interaction specificities of rodent TfR1 have been dynamic over time, shaped by selective pressures imposed by viruses. These dynamics have played out through mutations accumulated at just a small number of residue sites, where mutations decrease virus entry without measurably affecting receptor expression or iron-transport functions. TFR1 represents the first case, to our knowledge, where the evolution of a single host gene is driven by two host-virus arms races at once. In the case of the MMTV binding surface, this has played out through three residue positions coordinated in three-dimensional space. In the arenavirus binding surface, the target of selection has been a small surface-exposed structural feature, in which we were able to detect positive selection of three of the residues. Outside of rodents, TfR1 is used by a third family of viruses, the parvoviruses, and carnivore TFR1 is also under positive selection [72]. TFR1 evolution has thus been shaped by viruses in two separate species groups (rodents and carnivores) and by every viral pathogen known to use this receptor. These findings now explain how TFR1 became divergent enough to create species-specific interactions with all three of these virus families. If even a few residue positions can evolve to block virus entry without collateral damage to cellular function, host-virus arms race dynamics can unfold even in genes encoding highly conserved and essential housekeeping proteins.
This evolution of TFR1 can be put into contrast with other types of pathogen-driven positive selection of host genes. The human CCR5 gene encodes a co-receptor for HIV cellular entry. Some humans encode a variant allele of CCR5, CCR5Δ32, where a 32 base pair deletion gives rise to a defective receptor that is not expressed on the cell surface [73]. Individuals homozygous for this allele are almost completely resistant to HIV infection, and even heterozygous genotypes afford some protection due to reduced expression of wild-type CCR5. Like the model proposed herein for TFR1 L212V, CCR5Δ32 pre-dates HIV and may or may not have had any functional significance before the HIV pandemic. Nonetheless, it has become highly relevant in a world with HIV/AIDS. Like HIV, most simian immunodeficiency virus (SIV) strains also use CCR5 as a co-receptor. In a fascinating case of convergent evolution, some sooty mangabeys and red-capped mangabeys also encode null or defective alleles of CCR5 [21],[23]. Similarly, the DARC gene encodes a chemokine receptor that is used as an entry receptor by some malaria-causing Plasmodium species. A cis-regulatory polymorphism that silences DARC expression in erythrocytes has arisen independently in human populations from different parts of the world and is highly protective against Plasmodium vivax and Plasmodium knowlesi infection [24],[25]. Similar mutations have arisen in the cis-regulatory region of DARC in African baboons, and these are associated with resistance to a malaria-like parasite common in baboon populations [26]. In all of these cases, it has been speculated that selective pressure exerted by pathogens has driven these hypomorphic receptor alleles to high frequency in the affected human and nonhuman primate populations.
These CCR5 and DARC examples represent a more common mode of pathogen-driven positive selection (not recurrent) than the one demonstrated for TFR1, and there are several important differences. When receptor genes experience hypomorphic mutations, the predominant evolutionary strategy available to viruses will be to use a new receptor altogether. Indeed, the SIV strains that infect sooty and red-capped mangabeys (SIVsmm and SIVrcm) have both evolved to use alternate co-receptors [21],[23]. A few CCR5Δ32 homozygous humans have also been reported to be infected with HIV, again through mutations that allow the virus to use an alternate co-receptor (CXCR4 in this case). Hypomorphic mutations in receptors are not expected to be “serially replaced” due to arms race dynamics. Rather, viral evolution to use a new receptor ends the arms race with the original receptor gene and starts a new one with the new receptor gene. The CCR5 and DARC examples also involve evolutionary time scales millions of years shorter than what has been demonstrated in the current study; because these hypomorphic alleles are circulating in populations of individuals and are not shared between species, they have arisen relatively recently. Also, because these mutations simply reduce cellular expression of the encoded receptors, they presumably have some negative fitness effect on the host. The TfR1 example that we provide here is unique because solutions to viral entry have been found that appear to lack collateral damage to transferrin binding, and presumably to other host functions as well. Because of this, these mutations become common or fixed in populations where they occur, and are serially replaced as viruses continue to evolve and as rodents continue to speciate.
There is reason to believe that host-virus arms races are also shaping the protein sequence of other virus entry receptors in the manner described here. There are several other examples where significant sequence and functional divergence exist both on the side of a virus and its host entry receptor. For instance, certain strains of murine leukemia virus (MLV) use the rodent XPR1 receptor for cellular entry [74]. There are several functionally distinct variants of the XPR1 gene encoded by rodents of the genus Mus, each with its own pattern of virus susceptibilities. The viruses that use this receptor are also highly variable in the receptor-binding portion of their surface protein, Env. High levels of sequence divergence and disparate interaction specificities have also been observed between the entry receptor TVB encoded by birds and the avian leukosis virus (ALV) strains that use this receptor [75]. In neither of these cases is the housekeeping function or structure of the receptor known, so the pleiotropic consequences of pathogen-driven selection remain to be explored. However, both of these viruses can evolve to use new allelic forms of their receptor encoded by new hosts, suggesting that the receptors are important determinants of host range. High levels of sequence divergence, along with polymorphic and species-specific interactions between receptors and viruses, should be the hallmark for this type of evolution. These patterns have also been observed in other pairs of receptors and viruses [72],[76]–[80], suggesting that arms races might shape many receptors and potentially other types of housekeeping proteins exploited by viruses as well [81],[82].
Traditionally, TfR1 has been viewed as a housekeeping protein with an immensely important and conserved role in the cell. This study provides a much richer understanding of the multiple dynamic roles that this receptor is balancing in nature.
Rodent TFR1 and Machupo gp1 sequences were analyzed for positive selection. Database accession numbers for sequences used are listed in Tables S1 and S2. Sequences were aligned in Clustal [83], with minor adjustments made by hand (these two alignments contain few or no indels, respectively). jModeltest v2.1.1 [84] was used to select the best-fit model of nucleotide substitution, which was HKY+G in both cases. Phylogenetic trees for each sequence set were built by the maximum likelihood method implemented in MEGA5 [85]. The TFR1 gene tree matches the species tree of these rodents [86]. Because the Machupo gp1 sequences represent viral isolates from the same population, GARD [87] was run on the gp1 alignment to confirm the lack of phylogenetic breakpoints indicative of recombination. For both datasets, maximum likelihood analysis of dN/dS was then performed with codeml in the PAML 4.1 [64] software package. To detect selection, multiple alignments were fit to the NSsites models M1a (neutral model, codon values of dN/dS are fit into two site classes, one with value between 0 and 1, and one fixed at dN/dS = 1), M2a (positive selection model, similar to M1a but with an extra codon class of dN/dS>1 allowed), M7 (neutral model, codon values of dN/dS fit to a beta distribution, dN/dS>1 disallowed), M8a (neutral model, similar to M7 except with a fixed codon class at dN/dS = 1), and M8 (positive selection model, similar to M7 but with an extra class of dN/dS>1 allowed). Model fitting was performed with multiple seed values for dN/dS (ω) and assuming either the f61 or f3x4 model of codon frequencies [88]. Likelihood ratio tests were performed to assess whether permitting some codons to evolve under positive selection gives a significantly better fit to the data than models where positive selection is not allowed. The results obtained were shown to be robust to changes in the codon frequency model used, and the seed value for dN/dS (Tables S1 and S2). Posterior probabilities of codons under positive selection in M8 were then inferred using the Naive Empirical Bayes (NEB) algorithm. Coordinates for molecular structures were obtained from the RSCB protein database (http://www.pdb.org/) and rendered using PyMOL (http://www.pymol.org).
Full-length MMTV sequences were obtained on GenBank (AF228552, D16249, AF033807, AF228551). These sequences were used to BLAT [89] the current assemblies of the M. musculus (mm9) [90] and R. norvegicus (rn4) [91] genomes on the UCSC genome browser [92], recovering the indicated ERVs in these genomes. The nr/nt database for rodents (taxid:9989) at NCBI was searched for similar sequences in other species using the discontiguous megablast search algorithm with full-length MMTV as a query, and using the tBLASTx algorithm with MMTV pol as a query. Both of these approaches identified the Peromyscus maniculatus ERV buried in the sequence of GenBank record EU204642 (a BAC clone containing the deer mouse beta-globin gene cluster). A relatively young age of this ERV can be inferred from the fact that one open reading frame (pol) is still uninterrupted, and from the observation that the 5′ and 3′ LTRs differ at only 1 out of 917 positions. The giraffe, bison, and musk ox sequences are from [93]. Exogenous and endogenous beta-retrovirus genome sequences were aligned with MUSCLE [94] as implemented in MEGA5 [85]. jModeltest v2.1.1 [84] was used to select GTR+I+G as the best-fit model of nucleotide substitution. Phylogenetic trees were built by the maximum likelihood method implemented in MEGA5. Positions in which one or more sequences contained a gap were excluded during tree building. One thousand bootstrap replicates were performed and results are presented as percentage of replicates that supported each node.
The L212V SNP in human TFR1 (rs41301381) was identified in data deposited by the 1000 Genomes Project (http://browser.1000genomes.org). As of Release 12, L212V had been found as a heterozygous SNP in 11 individuals, with no homozygous carriers identified. Three of these individuals were Han Chinese from the South (CHS population), six were Han Chinese from Beijing (CHB population), and two were Japanese individuals (JPT population). In total, 11 out of 286 Asian individuals surveyed were heterozygous at this position, yielding a genotypic frequency of 0.038 in Asia. This SNP has not been included in the HapMap Genotyping Project (as of Release 28).
Human embryonic kidney 293T cells (ATCC CRL-11268), HEK293 cells (ATCC CRL-1573), human embryonic lung HEL299 cells (ATCC CCL-137), and canine kidney MDCK.2 cells (ATCC CRL-2936) were all maintained in Dulbecco modified Eagle's medium (Cellgro) supplemented with 10% fetal bovine serum (Gibco), 100 units ml−1 penicillin, 100 µg ml−1 streptomycin, and 2 mM L-glutamine (Cellgro). Human, Mus musculus, Calomys musculinus, Calomys callosus, and Zygodontomys brevicauda TFR1 with an encoded C-terminal FLAG tag were moved from pcDNA3.1 (+) vectors (described previously [7]) into the Gateway entry vector pCR8 using the pCR8/GW/TOPO TA Cloning Kit (Invitrogen). The following primers were used to amplify TfR1 for TA cloning: 5′-TTAATACGACTCACTATAGGG-3′ and 5′-TAGAAGGCACAGTCGAGGC-3′. Gateway LR recombination (Invitrogen) was performed to transfer TFR1 genes from pCR8 into the entry site in a Gateway-converted LPCX retroviral vector. Site-directed mutagenesis of the human, M. musculus, and Z. brevicauda TFR1 orthologs was performed using QuikChange Site-Directed Mutagenesis kit (Stratagene). Plasmids encoding Machupo, Junin, and Guanarito GP have been described previously [6]. An MMTV Env-encoding plasmid (pQ61) was kindly provided by Dr. Susan Ross (via Dr. Jackie Dudley).
The above described LPCX:TFR1 retroviral vectors were packaged in 293T cells by co-transfecting them along with the NB-MLV packaging plasmid pCS2-mGP [95] and pC-VSV-G using Fugene (Roche). Supernatants were collected and used to infect MDCK.2 (dog) cells. After 24 h, media containing 3.5 µg ml−1 puromycin was added to select for transduced cells (1.0 µg ml−1 puromycin was added when creating the HEK293 and HEL299 stable cell lines). These receptors have a C-terminal FLAG tag that is extracellular when the receptor is at the cell surface [8]. Expression of TfR1 proteins was detected in live cells by flow cytometry using an anti-FLAG antibody conjugated with Allophycocyanin (Abcam, catalog ab72569). Stable cell lines expressing human 212L and 212V TFR1 alleles were made in MDCK, HEK293, and HEL299 cells as described above.
Arenavirus GP or MMTV Env pseudotyped MLV recombinant retroviruses were packaged in 293T cells. Fugene (Roche) was used to co-transfect the GFP-encoding transfer vector pQCXIX (BD Biosciences) along with plasmids encoding MLV Gag-Pol and one of the viral surface glycoproteins Machupo GP, Junin GP, Guanarito GP, or MMTV Env. After 48 h, supernatants containing viruses were harvested, filtered, and frozen at −80°C. For entry assays, cell lines stably expressing various TfR1 orthologs or human alleles were plated at a concentration of 1.0×105 cells per well in a 24-well plate and, after 24 h, infected with pseudotyped virus along with 5 µg ml−1 polybrene. The plates were spinoculated with centrifugation at 350g for 1.25 h at 30°C. After 2 h of incubation at 37°C, cells were washed once with PBS and the media was replaced. Two days postinfection, cells were analyzed by flow cytometry. Cells were first gated for live cells and then, using an anti-FLAG antibody conjugated with Allophycocyanin (APC; Abcam, catalog ab72569), further gated such that all samples were narrowed to the same log decade of receptor expression (capturing the majority of cells but excluding outliers). Where TfR1 expression levels are reported, this is the mean fluorescent intensity within this gated population (10,000 cells). These same 10,000 cells were scored for expression of GFP (viral entry). Analysis of flow cytometry data was performed using FlowJo 8.8.6 (TreeStar Inc, Ashland, OR).
MDCK.2 stable cell lines expressing FLAG-tagged TfR1 orthologs were trypsinized and aliquoted in triplicate at a concentration of 2.5×105 cells/tube. The cells were washed with DPBS with 1% ovalbumin (Sigma). The cells were then resuspended in 200 µL of DPBS with 1% ovalbumin containing 1∶500 dilution of FITC-conjugated Mouse transferrin (2.0 mg/mL stock concentration; Jackson ImmunoResearch, 015-090-050) and incubated at 37°C for 60 min. Anti-DDDDK (FLAG) tag antibody conjugated with Allophycocyanin (0.1 mg/mL stock concentration; Abcam, catalog ab72569) was added to the cells at a 1∶100 dilution and incubated on ice for 20 min. The cells were then washed twice, resuspended in DPBS with 1% ovalbumin, and then analyzed by flow cytometry. Cells were first gated for live cells and then further gated such that all samples were narrowed to the same log decade of receptor expression (capturing the majority of cells but excluding outliers). Where TfR1 expression levels are reported, this is the mean fluorescent intensity within this gated population (10,000 cells). These same 10,000 cells were simultaneously analyzed for transferrin binding with FITC. Analysis of flow cytometry data was performed using FlowJo 8.8.6 (TreeStar Inc., Ashland, OR).
|
10.1371/journal.pbio.1000139 | RIN4 Functions with Plasma Membrane H+-ATPases to Regulate Stomatal Apertures during Pathogen Attack | Pathogen perception by the plant innate immune system is of central importance to plant survival and productivity. The Arabidopsis protein RIN4 is a negative regulator of plant immunity. In order to identify additional proteins involved in RIN4-mediated immune signal transduction, we purified components of the RIN4 protein complex. We identified six novel proteins that had not previously been implicated in RIN4 signaling, including the plasma membrane (PM) H+-ATPases AHA1 and/or AHA2. RIN4 interacts with AHA1 and AHA2 both in vitro and in vivo. RIN4 overexpression and knockout lines exhibit differential PM H+-ATPase activity. PM H+-ATPase activation induces stomatal opening, enabling bacteria to gain entry into the plant leaf; inactivation induces stomatal closure thus restricting bacterial invasion. The rin4 knockout line exhibited reduced PM H+-ATPase activity and, importantly, its stomata could not be re-opened by virulent Pseudomonas syringae. We also demonstrate that RIN4 is expressed in guard cells, highlighting the importance of this cell type in innate immunity. These results indicate that the Arabidopsis protein RIN4 functions with the PM H+-ATPase to regulate stomatal apertures, inhibiting the entry of bacterial pathogens into the plant leaf during infection.
| Plants are continuously exposed to microorganisms. In order to resist infection, plants rely on their innate immune system to inhibit both pathogen entry and multiplication. We investigated the function of the Arabidopsis protein RIN4, which acts as a negative regulator of plant innate immunity. We biochemically identified six novel RIN4-associated proteins and characterized the association between RIN4 and the plasma membrane H+-ATPase pump. Our results indicate that RIN4 functions in concert with this pump to regulate leaf stomata during the innate immune response, when stomata close to block the entry of bacterial pathogens into the leaf interior.
| Plants are continuously exposed to a variety of microorganisms. In order to successfully avoid infection, they have evolved a series of defense mechanisms that work in concert to limit pathogen invasion and multiplication [1]. Unlike vertebrates, plants lack an adaptive immune system and rely on their innate immune system to recognize and restrict pathogenic microbes. Conceptually, there are two primary branches of plant innate immunity. One branch employs extracellular receptors to recognize conserved microbial features termed pathogen-associated molecular patterns (PAMPs), resulting in PAMP-triggered immunity (PTI). The second branch uses intracellular plant resistance (R) proteins to recognize pathogen effectors delivered inside host cells during infection, resulting in effector-triggered immunity (ETI). Despite the importance of plant innate immunity, how pathogen perception activates immune responses and signaling overlap between PTI and ETI remain elusive.
PAMPs are conserved microbial features, such as bacterial flagellin or fungal chitin, which fulfill a function crucial to the lifestyle of the organism. PAMPs are perceived by pattern-recognition receptors resulting in PTI. The activation of PTI leads to the induction of mitogen-activated protein kinase (MAPK) signaling, transcriptional reprogramming, production of reactive oxygen species, and callose deposition, which serves as a physical barrier at infection sites (reviewed in [2]).
In order to colonize plants, virulent microorganisms need to overcome PTI. Plant pathogenic bacteria use the type III secretion system to deliver 20–30 effector proteins into the plant cell during pathogenesis. Collectively, these effectors are required for virulence and individual effectors have been shown to inhibit PTI through a variety of mechanisms [3]. The most well-studied bacterial effectors come from P. syringae pv. tomato (Pst), the causal agent of bacterial speck on Arabidopsis and tomato. In susceptible plant genotypes effectors enhance pathogen virulence and can inhibit PTI and ETI; in resistant plant genotypes effectors are recognized, culminating in an inhibition of pathogen growth [4],[5]. Despite the wide range of pathogens recognized, the majority of R genes can be grouped into one large family encoding proteins with a nucleotide-binding site (NB) and C-terminal leucine rich repeat (LRR) domains [6]. Several plant R proteins can detect effectors indirectly by monitoring for effector-induced perturbations of key host proteins.
To date, RIN4 (At3g25070) is the only known protein that can regulate both branches of the plant immune system. RIN4 overexpression lines exhibit decreased callose deposition after PAMP treatment as well as enhanced growth of virulent and type III secretion-deficient Pst, indicating a reduction in PTI [7]. rin4 knockout lines exhibit increased callose deposition after PAMP treatment and decreased Pst growth, consistent with enhanced PTI signaling [7]. These data indicate that RIN4 is a negative regulator of PTI. In addition, two R proteins, RPM1 (At3g07040) and RPS2 (At4g26090), monitor RIN4. RPM1, RPS2, and RIN4 are all localized to the plasma membrane [8]–[10]. In the absence of pathogen perception, RIN4 acts as a negative regulator of RPM1 and RPS2. When the P. syringae effectors AvrRpm1 or AvrB are delivered to the plant cell, RIN4 is hyper-phosphorylated, which in turn leads to the activation of RPM1-mediated resistance [8]. Another P. syringae effector, AvrRpt2, is a protease that directly targets RIN4, leading to the activation of RPS2-mediated resistance [11]–[14]. Investigation of the Arabidopsis–P. syringae interaction has identified RIN4 is a point of convergence for the regulation of both PTI and ETI. However, a mechanistic understanding of how RIN4 negatively regulates PTI remains elusive.
Many pathogenic bacteria can proliferate as epiphytes on the plant leaf surface, but in order to infect a plant they must colonize host tissues. Bacterial pathogens gain entry inside plant leaves through wounds or natural openings like stomata. Stomatal pores, located on the aerial epidermis, permit gas exchange between plants and the atmosphere. A pair of guard cells surrounds stomatal pores. Guard cells respond to diverse stimuli in order to regulate stomatal apertures including: blue light, temperature, humidity, CO2, plant hormones, and pathogen inoculation [15]–[17]. Stomatal pores operate as osmotic machines that open when the PM H+-ATPase of guard cells is allowed to be active. The activity of this proton pump generates a large transmembrane electrochemical gradient that drives the uptake of charged solutes and, as a consequence, water, which in turn causes the cells to swell and the pore between them to open. Stomatal closure is initiated upon depolarization of the guard cell plasma membrane by inhibiting the PM H+-ATPase.
Historically, stomata were thought to be passive ports of entry, but recent evidence reveals that stomatal closure is induced by both PTI and ETI in an attempt to restrict bacterial invasion [15],[18],[19]. Upon perception of PAMPs, stomata will close within 1 h. However, virulent bacteria are able to re-open stomata after 3 h, facilitating their entry into the plant leaf. For example, virulent Pst secretes the polyketide toxin coronatine, which stimulates the plant to re-open their stomata [15],[20]. Several other pathogenic microorganisms also act to regulate stomatal apertures during infection [19],[21]–[23]. One particularly well-characterized example is the toxin fusicoccin, produced by the fungal pathogen Fusicoccum amygdali [24]. Fusicoccin is a strong activator of the plasma membrane H+-ATPase and rapidly induces stomatal opening, presumably in order to facilitate fungal penetration [25]–[27]. Taken together, these data highlight the importance of stomatal pores and guard cell signaling during pathogen infection.
In this study, we report the identification and characterization of the Arabidopsis RIN4 protein complex. We were able to purify several associated proteins by immunoaffinity chromatography and identify them by mass spectrometry. We identified the PM H+-ATPases AHA1 (At2g18960) or AHA2 (At4g30190), whose interaction we characterized in greater detail. The C-terminal regulatory domain of AHA1 and AHA2 interact with RIN4 by yeast two-hybrid and we can detect a specific interaction between AHA1/AHA2 and RIN4 in planta using bimolecular fluorescence complementation (BiFC). RIN4 overexpression enhanced PM H+-ATPase activity, while the rin4 knockout line exhibited decreased PM H+-ATPase activity. Importantly, we demonstrate that the rin4 knockout cannot re-open its stomata in response to virulent Pst. We also show that RIN4 is expressed in guard cells along with other PTI and ETI signaling components. Our findings are consistent with a model in which RIN4 associates with the C-terminal autoinhibitory domain of the PM H+-ATPase to regulate leaf stomata in response to PAMPs.
In order to gain a more comprehensive understanding of the proteins involved in plant immune signaling, we investigated the components of the RIN4 protein complex in Arabidopsis thaliana. We used affinity-purified antibody recognizing RIN4 to purify associated proteins by immunoaffinity chromatography (Figure S1). The rps2-101c mutant complemented with the RPS2 transgene containing a C-terminal fusion to the hemagglutinin (HA) epitope was used for RIN4 purifications. This line is biologically relevant because RPS2:HA is expressed from its native promoter, can complement the rps2-101c mutation, and confers resistance to Pst expressing AvrRpt2 [11]. RPS2 associates with RIN4 in planta, and we used this association to troubleshoot purification conditions. Because the rin4 knockout is lethal in the presence of RPS2, we used the rps2/rin4 double mutant line to control for nonspecific protein binding [13]. Multiple purification protocols were tested in order to identify conditions that would enable us to detect the presence of both RIN4 and RPS2 by mass spectrometry. We found that wash conditions containing more than 150 mM NaCl eliminated most nonspecific protein binding, but also eliminated our ability to copurify RPS2 in the positive controls. Protein complex purifications were also conducted after plasma membrane fractionation, but this eliminated our ability to copurify RPS2 (unpublished data). Therefore, we used whole leaf protein extracts and mild wash conditions to purify RIN4 associated proteins across three biological replicates. Proteins from each sample were analyzed directly using high performance liquid chromatography coupled to tandem mass spectrometry (MS; Figure S1). Proteins were identified using the MASCOT algorithm to search the Arabidopsis genome. All experiments captured native, biologically relevant levels of RIN4 and associated proteins.
We reproducibly identified RIN4 and RPS2 as well as six novel RIN4-associated proteins across three biological replications (Tables 1, S1, and S2). In order to be classified as a RIN4-associated protein, the protein had to be identified by a minimum of two unique peptides and be present in all three replications of the positive control, but never identified in the negative control rps2/rin4. Although we were able to identify RPS2 and RIN4 by mass spectrometry, we did not identify two additional proteins that are known to interact with RIN4: NDR1 (At3g20600) and the R protein RPM1 [8],[28]. Both proteins have been demonstrated to interact by yeast two-hybrid and co-immunoprecipitation. Our inability to detect RPM1 could be because only a small percentage of RPM1 interacts with RIN4 in the plant, indicating that these two proteins may transiently interact during ETI [8]. Alternatively, our mass spectrometry analysis may have only identified the most abundant RIN4 associated proteins. In contrast to RIN4, which is easily detected by western blot, RPM1 and NDR1 are expressed at very low levels, making them difficult to identify by mass spectrometry.
A MATH domain protein, two Jacalin domain proteins, ERD4, a remorin, and the PM H+-ATPases AHA1 and/or AHA2 were identified by mass spectrometry (Tables 1 and S1). The MATH domain is broadly represented in eukaryotes [29]. Proteins containing MATH domains, primarily the well-characterized TNF Receptor Associated Factor family, are involved in human disease resistance signaling through their regulation of inflammation and apoptosis responses [30]. MATH domains are thought to act as protein adapters, transferring signals to intracellular signaling pathways. Proteins containing MATH domains are prevalent throughout the plant kingdom, but have not been characterized or implicated in plant disease resistance. Jacalins are lectins, which have been shown to be induced in response to the hormone methyl jasmonate [31]. ERD4 (Early Responsive to Dehydration 4) was originally identified because it is rapidly induced during drought stress [32]. Microarray analysis has revealed that ERD4 is also induced in response to multiple biotic and abiotic stresses, although its function remains elusive (unpublished data). Remorins are plasma membrane associated proteins of unknown function with C-terminal coiled-coiled domains. Multiple remorins possess an N-terminal domain with similarity to viral movement proteins [33]. All of these proteins are predicted to be membrane-localized, which is where RIN4 resides [8].
We also identified the PM H+-ATPase (AHA), the proton pump responsible for energization of the plasma membrane. We were unable to distinguish between the highly homologous AHA1 and AHA2 proteins by mass spectrometry in two out of three biological replications. We were able to identify AHA1 specific peptides in the first MS run (Table S2). There are 11 AHA genes in Arabidopsis, which pump H+ from the cytosol to the apoplast in an ATP-dependent manner. AHA1, AHA2, and AHA5 are the major transcripts found in guard cells [34]. AHA1 and AHA2 are predicted to have molecular masses of 104.2 and 104.4 kDa, respectively, and share 94% amino acid identity. In light of recent data implicating AHA1 in stomatal regulation and the role of stomatal closure in the innate immune response, we decided to analyze the association between RIN4 and AHA1/AHA2 in greater detail [15],[35].
In order to validate the RIN4 AHA1/AHA2 association detected by mass spectrometry, we subjected them to BiFC and yeast two-hybrid analyses. AHA1 and AHA2, which are negatively regulated by their C termini, possess multiple transmembrane domains (reviewed in [36]). Therefore, we employed the hydrophilic C-terminal regulatory domain of AHA1 and AHA2 in our yeast two-hybrid analyses. As shown in Figure 1A, we detected an interaction between RIN4 and the C termini of both AHA1837–950 and AHA2837–949, when compared with the negative control T-antigen/Lamin-C using the Matchmaker system. We were unable to detect any interaction between RPS2, AHA1837–950, or AHA2837–949 by yeast two-hybrid (unpublished data). We verified that RIN4, AHA1837–950, and AHA2837–949 are expressed in yeast and do not autonomously activate His auxotrophy (Figures 1A and S2). We also tested beta-galactosidase activity, but could only detect a faint blue color (unpublished data). These results indicate that RIN4 can weakly interact with the C terminus of AHA1 and AHA2 by yeast two-hybrid.
To provide additional evidence for the AHA and RIN4 interaction, we investigated the association between AHA1, AHA2, and RIN4 in planta using a BiFC approach to directly visualize protein interactions in living cells. A specific interaction between either AHA1 or AHA2 and RIN4 was detected in Nicotiana benthamiana leaves (Figure 1B, a, b). The yellow fluorescent protein (YFP) fluorescence was clearly localized to the plasma membrane, where RIN4, AHA1, and AHA2 have been previously shown to be located. The background fluorescence of chloroplasts in the green channel is due to the excitation at 488 nm. Meanwhile, we were unable to detect any YFP fluorescence between AHA1 or AHA2 and RPS2 (Figure 1B, d, e). As a negative control we co-expressed each protein with the auxin influx carrier AUX1 (At2g38120), an integral plasma membrane protein. None of the proteins were able to induce YFP fluorescence in the presence of the negative control, indicating a specific interaction between AHA1/AHA2 and RIN4 in planta. In order to ensure that the proteins used as negative controls indeed were expressed, expression of AUX1 was detected by western blotting employing the His tag included in the construct (unpublished data) and expression of RPS2 was tested by observation of cell death 48 h after infiltration (unpublished data).
RIN4 can interact with the C-terminal regulatory domains of AHA1 and AHA2. Therefore, we investigated the hypothesis that RIN4 can regulate H+-ATPase activity. Because it is not possible to measure the biochemical activity of single PM H+-ATPase isoforms in planta, we analyzed PM H+-ATPase activity as a whole, even though RIN4 may only affect a subset of ATPases. Plasma membrane vesicles were purified from Col 0, dexamethasone (Dex) inducible RIN4 overexpression [7], rpm1/rps2, and rpm1/rps2/rin4 leaf tissue by aqueous two-phase partitioning. We have used the rpm1/rps2/rin4 triple mutant for experiments to avoid the weak activation of RPM1 that occurs in the absence of RIN4 [37]. PM H+-ATPase activity was subsequently measured on inside-out plasma membrane vesicles as described by Palmgren and colleagues [38]. In this assay, the PM H+-ATPase hydrolyzes ATP and pumps H+ into vesicles, which creates a pH gradient across the membrane. The pumping activity was measured by quenching of the ΔpH probe acridine orange at an absorbance of 495 nm. H+ transport measured from plasma membrane vesicles purified from wild-type Col 0 leaves demonstrated that these vesicles were both transport competent and highly enriched for plasma membrane (Figure S3). In rpm1/rps2/rin4 leaves, H+-ATPase activity was 30% lower than Col 0 (p<0.001, Figure 2A and 2C). In RIN4 overexpression lines, H+-ATPase activity was 65% higher than Col 0 (Figure 2B and 2D). We also noticed that the rpm1/rps2 double mutant exhibited slightly higher H+-ATPase activity than Col 0 (7%–13%) across independent plasma membrane isolations (p<0.05, Figure 2A and 2C). Because both RPS2 and RPM1 interact with RIN4, this line may possess more RIN4 protein that can interact with the H+-ATPase, thus increasing its activity. RIN4 overexpression was induced by spraying the Dex:RIN4 line with 20 µM Dex and harvesting tissue 48 h later (Figure 2E). We also found that Dex treatment itself slightly inhibited the H+-ATPase enzymatic activity in Col 0. This is not surprising, because previous studies have revealed that Dex treatment alone can lead to significant changes in gene expression [39]. Nevertheless, when comparing to Col 0 and Dex:RIN4 lines after treating with Dex, it is clear that RIN4 overexpression leads to enhanced PM H+-ATPase activity. These results are consistent with the hypothesis that RIN4 can act to regulate H+-ATPase activity at the plasma membrane. On the basis of these results, RIN4 acts as a positive regulator of AHA1/AHA2, as RIN4 overexpression lines exhibit enhanced AHA activity and the rin4 knockout exhibits decreased AHA activity.
To test the in vitro effect of RIN4 on H+ pumping, recombinant RIN4 protein was purified from E. coli and added directly to H+ transport assays. H+ transport activity in vesicles isolated from the rpm1/rps2/rin4 knockout was increased in the presence of 3 µg of RIN4 (Figure 3). No effect on H+ transport was observed when recombinant RIN4 protein was added to vesicles isolated from wild-type Col 0 plants (Figure 3).
In order to determine if altering the activity of AHA1 or AHA2 could lead to changes in PTI or ETI, we first analyzed aha1 (salk_118350) and aha2 (salk_022010) knockout lines. We were unable to detect any obvious morphological or altered disease phenotypes in either knockout line (unpublished data). We were unable to generate an aha1/aha2 double mutant by crossing salk_118350 and salk_022010, a result that has been reported previously [40]. These results suggest that knocking out both AHA1 and AHA2 is a lethal combination, indicating that that AHA1 and AHA2 may be functionally redundant in Arabidopsis. Therefore, we analyzed ost2-1D and ost2-2D, which possess point mutations of P68S and L169F/G867S in AHA1, respectively, and act as dominant activation mutations [35]. The ost2-1D mutant background is in the Landsberg erecta (Ler) ecotype and the ost2-2D is in the Col 0 ecotype.
The ost2-1D and ost2-2D mutants were originally identified based on their open stomata phenotype [35]. Because stomata can serve as ports of entry for microbial pathogens, we hypothesized that these mutants may facilitate enhanced bacterial entry inside leaves. We were unable to detect a difference between Col 0, Ler, and ost2-1D or ost2-2D after syringe infiltration with virulent Pst DC3000 or avirulent Pst DC3000 expressing the effector AvrRpt2, which induces ETI (Figure 4A and 4B). Col 0 and Ler exhibited clear bacterial speck symptoms by 4–5 d after spray inoculation. However, the leaves of ost2-2D lines were completely collapsed by 4 d after spray inoculation. Therefore, all growth curves were performed at 3 d post-inoculation, when disease symptoms were clearly visible on ost2-1D and ost2-2D (Figure 4D). When we spray-inoculated with Pst DC3000 or Pst DC3000 (AvrRpt2), the bacteria were able to grow 5- to 10-fold more in the ost2-1D and ost2-2D mutant lines compared to Ler and Col 0 and displayed enhanced disease symptoms (Figure 4A, 4B, and 4D). These results show that AHA1 activation can facilitate Pst entry into the plant leaf interior.
Our genetic analysis suggests that AHA1 and AHA2 are functionally redundant. Therefore, we hypothesized that AHA2 overexpression lines would also enable enhanced bacterial entry into the leaf interior. AHA2 regulation has been well-studied in vitro, and the C terminus acts as a negative regulator of the PM H+-ATPase [36],[41],[42]. Removing the C terminus induces strong auto-activation in vitro and in planta [41],[43]. We generated an AHA2 overexpression line in Col 0 by transforming a truncated version of AHA2 (amino acids 1–837) without its C-terminal inhibitory domain under the control of the cauliflower mosaic virus 35S promoter. Because of the small leaf size of the 35S:AHA21–837 line, we were unable to syringe inoculate or harvest large quantities of leaf tissue necessary for PM H+-ATPase enzymatic analysis. The resulting transgenic plants were dwarf with pronounced leaf chlorosis, decreased germination rates, and possessed enhanced AHA2 expression (Figure S4). Pst DC3000 was able to grow 20-fold more in this line compared to Col 0 after spray inoculation (Figure 4C). However, the 35S:AHA21–837 line did not have a constitutively open stomata phenotype like ost2-1D and ost2-2D mutants (unpublished data). The pleotropic phenotypes generated by overexpressing AHA21–837 in Col 0 are not surprising because strong constitutive activation of plasma membrane H+-ATPase(s) can result in a nonspecific expression in different cell types, profound changes in plasma membrane potential, and will affect multiple biological processes [43]. For these reasons, we did not investigate the 35S:AHA21–837 line further and concentrated our analyses on the AHA1 activation mutants.
The ost2-1D and ost2-2D mutants were previously reported as lesion-mimic mutants and displayed salicylic acid-induced necrosis on leaflets [35]. Under standard growth conditions for pathogen inoculation, we did not observe this phenotype on any of the lines exhibiting enhanced PM H+-ATPase activity (Figure 4D). However, we were able to visualize leaflet necrosis on both lines when they were grown under conditions to promote flowering (140 µmol/sec/m2, 16-h days, 23°C). The phenotypes of lesion-mimic mutants can be variable and sensitive to variations in growth conditions [44]. Lesion-mimic mutants are often associated with mutations in ion channels [44],[45]. As the AHA family is an important regulator of multiple cellular processes, spatial and temporal regulation of PM H+-ATPases inside mesophyll cells may also be an important component of plant immune signaling.
In order to test the hypothesis that enhanced bacterial growth on ost2 mutant leaves is due to their increased ability to gain entry into the leaf interior via open stomata, we inoculated wild-type Arabidopsis and ost2 mutant lines with the nonmotile Pst flagellin mutant flaA [46]. The flaA mutant grew to similar levels as wild-type Pst when syringe infiltrated in Col 0 leaves (Figure 5A). We were unable to detect enhanced growth of the flaA mutant after spray inoculation onto ost2-1D and ost2-2D, indicating that these mutant plants promote bacterial colonization of the leaf by allowing bacteria to gain entry by swimming through their stomatal apertures (Figure 5B). Interestingly, we noticed that growth of the flaA mutant was decreased in ost2 mutants after spray inoculation, but not syringe infiltration (Figure 5B). This may be due to an inability of the flaA mutant to swim away from unfavorable microenvironments (such as low pH) near stomatal openings with enhanced PM H+-ATPase activity.
Lines exhibiting increased AHA1 activity are more susceptible to bacterial inoculation due to their open stomata phenotype (Figures 4 and 5). Previously, Melotto and colleagues showed that upon perception of PAMPs, Col 0 stomata will close within 1 h [15]. Virulent Pst can re-open stomata after 3 h through the production of coronatine, facilitating pathogen entry. Because RIN4 can interact with the C-terminal regulatory domain of AHA1 and AHA2 (Figure 1), we investigated the stomatal response in the rin4 knockout line after pathogen inoculation. Leaf epidermal peels from Col 0, rpm1/rps2, and rpm1/rps2/rin4 were floated on 1×108 colony-forming unit (CFU)/ml Pst DC3000 and their stomatal apertures were measured in response to pathogen inoculation. Stomatal apertures from all genotypes closed after 1 h (Figure 6A). Importantly, we observed that Pst DC3000 could not re-open the stomata in rpm1/rps2/rin4 after 3h (Figure 6B). The stomata of rpm1/rps2 lines were open after 3 h, indicating that this phenotype is solely due to the lack of RIN4 (Figure 6B). We also tested ndr1-1 mutant plants for a defect in stomatal response to PAMPs, but ndr1-1 lines were still able to re-open their stomata 3 h after exposure to Pst DC3000 (unpublished data), indicating that NDR1 is not required for the RIN4-mediated stomatal phenotype. These observations are consistent with RIN4 being a negative regulator of plant innate immunity. These results also support the hypothesis that RIN4 and AHA1/AHA2 work together to regulate stomatal apertures in response to PTI.
Previously, the rpm1/rps2/rin4 triple mutant was shown to be more resistant than rpm1/rps2 after spray inoculation with Pst DC3000 [7]. In addition, rin4 knockout lines exhibit enhanced callose deposition in response to PTI, whereas RIN4 overexpression lines display the opposite phenotype [7]. Therefore, RIN4 may play a role in PTI signaling in both guard cells and mesophyll cells. In order to test this hypothesis, we inoculated rpm1/rps2 and rpm1/rps2/rin4 plants grown under the same conditions by both spray inoculation and syringe infiltration. Spray inoculation always resulted in a significant decrease of 4- to 9-fold in bacterial growth on the rpm1/rps2/rin4 mutant when compared to rpm1/rps2 (Figure 6C). We were also able to detect a slight decrease (2- to 4-fold) in bacterial growth on the rpm1/rps2/rin4 mutant after syringe infiltration. These results indicate that RIN4 contributes significantly to PTI signaling in guard cells and has a subtle phenotype with respect to PTI in mesophyll cells. Because rin4 knockout lines do not re-open their stomata after inoculation with Pst, this may be the reason why lines lacking RIN4 exhibit increased resistance to virulent bacteria after spray inoculation.
Our observation that virulent Pst cannot re-open stomata in rin4 knockout lines led us to investigate what cell types express RIN4. We investigated RIN4's expression pattern in intact leaves and guard cells. Guard cell protoplasts were isolated from Col 0, visually inspected for purity, and analyzed for the presence of RIN4 (Figure 7A). We used the expression of phosphoenolpyruvate carboxylase 2 (ATPPC2, At2g42600), which has low-level expression in guard cells and high-level expression in mesophyll cells, as a control to verify guard cell protoplast purity [47]. Each batch of purified guard cell protoplasts was divided in two for extraction of RNA and protein. Our reverse transcriptase (RT)-PCR analysis showed that RIN4 was expressed in both Col 0 guard cells as well as intact leaves (Figure 7A). Guard cells make up less than 2% of the leaf epidermal cells, which highlights the expression of RIN4 within guard cells. Next, we performed immunoblot analysis on leaf and guard cell protoplast protein extracts (30 µg) with the anti-RIN4 antibody. RIN4 protein was detected in Col 0 guard cells as well as in the intact leaf (Figure 7B). Given the abundance of mesophyll cells in the leaf sample, this result indicates that RIN4 is strongly expressed in guard cells.
AHA1 expression in guard cells was previously demonstrated [35]. On the basis of our interaction studies we therefore tested if AHA2 is also expressed in guard cells. Transgenic plants expressing an AHA2 promoter:GUS construct clearly demonstrated AHA2 expression in guard cells (Figure S4C) supporting the hypothesis that both AHA1 and AHA2 interact with RIN4 and that this interaction is physiologically relevant.
Given the importance of guard cells in regulating bacterial invasion, we investigated if additional immune signaling components were present in guard cells. Like Melotto and colleagues [15], we were able to detect the flagellin PAMP receptor FLS2 (unpublished data). We detected expression of the EF-Tu PAMP receptor EFR and the chitin PAMP receptor CERK1 in Col 0 guard cell protoplasts by RT-PCR (Figure 7C). We were also able to detect the expression of EDS1, PAD4, and NDR1, which are involved in the manifestation of ETI (Figure 7C). By mining publicly available microarray data from Yang and colleagues [48], we analyzed the expression of the following genes in both guard cell and mesophyll cell protoplasts: FLS2, EFR, CERK1, EDS1, PAD4, NDR1, RPS2, RPM1, and RIN4. With the exception of CERK1, all genes were expressed at a detectable level in both guard cells and mesophyll cells (unpublished data).
The stomata of ost2 mutants are ABA insensitive, but do respond to other stimuli such as CO2 and blue light, indicating that individual PM H+-ATPases may exhibit defined biological roles [35]. Therefore, we investigated the ability of ost2 mutant lines to respond to PTI-mediated stomatal closure. We floated epidermal peels of Ler, Col 0, and ost2 mutant lines on 1×108 CFU/ml Pst DC3000 and measured their stomatal apertures in response to pathogen inoculation. Pst could not induce stomatal closure in ost2-1D or ost2-2D, while 80% of the stomata from Col 0 and Ler were closed after 1 h (Figure 8A, 8B). Epidermal peels from ost2 mutants were also incubated with the flg22 peptide of flagellin and lipopolysaccharide (LPS), which are recognized as bacterial PAMPs. We clearly observed that incubation with 10 nM/ml flg22 or 100 µM LPS can induce stomatal closure in either Ler or Col 0 plants, but not in ost2-1D or ost2-2D (Figure 8C), suggesting that AHA1 inactivation contributes to stomatal closure during PTI signaling.
PTI induces an oxidative burst within minutes after pathogen perception, and treatment with reactive oxygen species, such as H2O2 and nitric oxide (NO) results in stomatal closure [49]. We were interested in determining if stomata from plants with enhanced AHA1 activity would respond to the presence of reactive oxygen species. In Figure S5, we treated plants with 0.2 mM H2O2 and 100 µM sodium nitroprusside (SNP, an NO donor). Neither H2O2 nor SNP could induce closure in ost2-1D and ost2-2D, but could rapidly induce stomatal closure in wild-type Arabidopsis. These results demonstrate that the stomata of ost2 mutants, which exhibit enhanced AHA1 activity, do not close in response to PTI, therefore enabling virulent bacteria to gain entry into the plant apoplast. Melotto and colleagues also demonstrated that PAMP-induced stomatal closure required the OST1 protein kinase, a key component of the ABA signaling pathway [15].
Recognition of pathogens by the host innate immune system is a critical component controlling survival and fitness of both animals and plants. We investigated the function of RIN4, an Arabidopsis protein that acts as a negative regulator of both PTI and ETI [7],[8],[11],[13]. Here, we have identified six novel RIN4 associated proteins. We have investigated the association between RIN4 and PM H+-ATPases AHA1 and AHA2 in detail. These data are consistent with the model of RIN4 acting in concert with the PM H+-ATPases AHA1 and AHA2 to regulate stomatal apertures in response to pathogen attack in resistant genotypes (Figure 9).
Stomata are surrounded by a pair of two guard cells, whose turgor controls opening and closure of the aperture. Changes in the turgor of guard cells are strongly influenced by the activity of PM H+-ATPase. Activation of PM H+-ATPase can lead to hyperpolarization of the plasma membrane and subsequent induction of inward K+ channels resulting in an increase in turgor due to concomitant entry of water and stomatal opening. In contrast, inhibiting the PM H+-ATPase and anion channel activation initiate plasma membrane depolarization, resulting in the activation of outward rectifying K+ channels [50],[51]. These ion effluxes result in a loss of guard cell turgor and stomatal closure. A number of secondary messengers are important for initiating membrane depolarization, including reactive oxygen species and Ca2+.
We have demonstrated that the RIN4 protein acts in concert with PM H+-ATPases to regulate stomatal apertures during PTI. Importantly, the rin4 knockout line does not re-open its stomata in response to virulent Pst (Figure 6). This result solidifies the importance of RIN4 in regulating stomatal apertures in response to pathogen attack. Previously, RIN4 was found to be a negative regulator of both PTI and ETI [7],[8],[11],[13]. Our results were consistent with these findings and suggest that RIN4's association with AHA1 and AHA2 is an important component of RIN4 function. Autoactive AHA1 mutants display increased susceptibility to virulent Pst, because of the bacteria's enhanced ability to gain access to the plant interior via open stomata (Figures 4 and 5). RIN4 overexpression lines exhibit enhanced disease susceptibility and increased PM H+-ATPase activity. Conversely, rin4 knockout lines exhibit decreased disease susceptibility and lower PM H+-ATPase activity (Figures 2 and 6C). These results can now explain how RIN4 acts to regulate plant innate immunity at the level of pathogen invasion.
Despite the importance of RIN4 in plant innate immunity, the pattern of RIN4 expression remained unknown. Using a combination of RT-PCR, western blotting, and microarray analyses we were able to demonstrate that RIN4 is expressed in guard cells (Figure 7). These results highlight the importance of RIN4 in PTI-induced stomatal closure. We were also able to detect the expression of multiple PAMP receptors, R genes, and innate immune signaling components in guard cells at the RNA level, emphasizing the importance of this cell type in the innate immune response (Figure 7).
Inhibition of the PM H+-ATPase is one of the first steps required to induce stomatal closure. These data are consistent with a model in which RIN4 acts in concert with AHA1 and/or AHA2 to regulate stomatal apertures in response to pathogen attack during PTI (Figure 9). Perception of the flagellin flg22 peptide during PTI was found to inhibit both inward and outward rectifying K+ channels [52]. Therefore, flagellin perception can not only induce stomatal closure, but can inhibit stomatal opening [15],[52]. Because stomata serve as points of entry for multiple bacterial, fungal, and oomycete pathogens, it is not surprising that several different classes of pathogens have evolved to manipulate stomatal apertures during pathogenesis. For example, the polyketide toxin coronatine, produced by several strains of P. syringae, can induce stomatal opening after PTI-mediated closure [15]. Coronatine can reverse the inhibition of inward rectifying K+ channels, leading to stomatal opening [52]. Xanthomonas campestris employs a small diffusible signal molecule, which can also induce stomatal opening on compatible hosts [19]. The most well-characterized example of stomatal manipulation by a pathogen is the toxin fusicoccin, produced by the fungal pathogen F. amygdali, the causal agent of almond and peach canker [24]. Fusicoccin is a potent activator of the PM H+-ATPase and strongly induces stomatal opening by binding to and stabilizing an activated H+-ATPase/14-3-3 complex [25],[27],[53]. These studies highlight the importance of stomatal regulation during plant innate immunity, as components of the signaling pathways controlling stomatal apertures can be regulated by the plant immune system as well as by virulent pathogens.
What is the mechanism RIN4 uses to regulate PM H+-ATPase activity? PM H+-ATPase regulation has been well studied over the last 20 years (reviewed in [54]). Both crystallographic data and homology modeling of the PM H+-ATPase indicate that it possesses a similar structure to other P-type ATPases [55],[56]. The PM H+-ATPase also possesses an extended C terminus [57], which is lacking in other P-type ATPases [57] and is involved in negative regulation of pump activity [58]. Activation of the PM H+-ATPase can be achieved by phosphorylation of the penultimate threonine residue. Phosphorylation of this residue leads to subsequent binding of regulatory 14-3-3 proteins, which displace the autoinhibitory C-terminal domain. This apparently induces the formation of a dodecamer consisting of six H+-ATPase and six 14-3-3 molecules in the PMA2 H+-ATPase isoform from N. plumbaginifolia [54],[56]. Additional phosphorylated residues have recently been identified that can contribute to both positive and negative regulation of the PM H+-ATPase, highlighting the complexity of this pump's regulation [59]–[61].
Data presented in this manuscript are consistent with RIN4 being a positive regulator of the PM H+-ATPases AHA1 and AHA2. Previous studies have demonstrated that RIN4 is phosphorylated in planta [8],[62]. It will be interesting to test if the phosphorylation status of RIN4 plays a role in regulating PM H+-ATPase activity. Future research investigating if RIN4 is transcriptionally or posttranslationally modulated during the guard cell response to PAMPs and Pst DC3000 may help elucidate the mechanism employed by RIN4 to regulate the PM H+-ATPase. In addition, RIN4 homologs can be detected in many plants where substantial DNA sequences are available. In the future, it will be important to determine the role of RIN4 as well as RIN4-associated proteins across different species. For example, stomatal closure in response to PTI occurs in multiple plants [15],[18]. Does the association of RIN4 with PM H+-ATPases act to regulate stomatal apertures in other species?
It will also be important to elucidate how innate immune complexes change in response to pathogen attack and if complex constituents are the same between different cell types. It is plausible that components of the innate immune complexes exist in distinct pools within each cell, with each pool controlling different aspects of PTI and ETI. There is evidence for RIN4 existing in different cellular pools within plant leaves based on data obtained from co-immunoprecipitation experiments [8],[11],[13]. In this study, we were able to elucidate members of the RIN4 complex in the absence of pathogen infection. An in-depth investigation how the RIN4 complex assembles and changes during PTI, ETI, and after pathogen-induced modification in different cell types (e.g., guard cells and mesophyll cells) and plant genotypes will greatly facilitate our understanding of innate immune signaling.
Arabidopsis plants, Columbia (Col 0), Landsberg erecta (Ler), and the mutants derived from them as indicated in the figures were sown in soil and stratified at 4°C for 2 d. In the text, the rps2, rpm1, and rin4 mutants refer to rps2-101c, rpm1-3, and the rin4 T-DNA knockout [8],[9],[63]. Dex:RIN4 lines were previously described, and all figures refer to line 31 [7]. Plants were grown in controlled environment chamber at 24°C with a 10-h light/14-h dark photoperiod under a light intensity of 85 µE/m2/s. For all the experiments, 4–5 wk old plants were used. 35S:AHA2(1–837) transgenic lines were generated by following the standard floral dip transformation procedure [64]. The AHA2 (1–837) fragment was cloned into the BamH I/Xho I site of binary vector pMD-1 and transgenic plants were screened on 50 µg/ml kanamycin. Two independent T3 lines were used for bacterial inoculation.
Pst DC3000, Pst DC3000 (AvrRpt2), and the flagellin deficient mutant Pst DC000 flaA− were grown on NYG plates for 30 h, then cultured at 28°C in NYG media for 48 h [46]. Pst DC3000 (AvrRpt2) expressed AvrRpt2 from the broad-host range vector pDSK519 [65]. Antibiotics were used for plate selection at the following concentrations: 25 µg/ml kanamycin, 100 µg/ml rifampicin, and 35 µg/ml chloramphenicol. For spray inoculation, Arabidopsis leaves were sprayed until runoff with a Preval sprayer containing 1×109 CFU/ml bacteria in 10 mM MgCl2 with 0.025% silwett L-77. Inoculated plants were left uncovered for 30 min and then covered with a plastic dome for 2 d. For syringe infiltration, bacteria were resuspended in 10 mM MgCl2 and inoculated at a concentration of 0.5×105 CFU/ml with a needleless syringe. Leaves were surface sterilized for 30 s in 70% ethanol, and bacterial populations were determined by growth curve analysis as described by Kim and colleagues [7]. All experiments were repeated at least three times, with a minimum of three biological replicates per time point.
Stomatal aperture measurements were conducted according to a published procedure [15]. Plants were induced to open stomata under white light for 2 h. Epidermal peels were floated on a 1×108 CFU/ml of Pst in water or purified PAMPs. For PAMP treatments, epidermal peels were floated on 5 µM flg22 peptide (synthesized by GenScript) in MES buffer (10 mM KCl, 0.2 mM CaCl2, 10 mM MES-KOH [pH 6.15]), 100 ng/µl LPS (Sigma) in MES buffer or MES buffer alone as a negative control. Stomatal apertures were analyzed by microscopy with a digital camera and measured with SPOT4.1 software (Diagnostic Instruments) at 0-h, 1-h, and 3-h timepoints. All experiments were repeated at least three times, with a minimum of three biological replicates per time point.
Arabidopsis plants were grown as described above for 5 wk in soil at a pH of 7.5. To determine the effect of overexpressing RIN4, Dex:RIN4 and Col 0 leaves were sprayed with water and 0.025% silwett or 20 µM Dex in 0.025% silwett. Leaf tissue was harvested after 48 h. For all experiments, plasma membranes were immediately purified after harvesting leaf tissue. Arabidopsis leaves (30 g) were homogenized with a blender in 200 ml ice-cold buffer containing 50 mM MOPS (pH 7.0), 0.33M sucrose, 5 mM EDTA, 2 mM DTT, 1.5 mM ascorbate, 0.2% (w/v) insoluble polyvinylpolypyrrolidone, 1 mM phenylmethylsulfonyl fluoride, 1 µg/ml leupeptin, and 1 µg/ml pepstain A. Plasma membranes were purified from the microsomal fraction (10,000 g to 50,000 g pellet) by partitioning at 4°C in an aqueous polymer two-phase system as described previously [66]. The final plasma membrane pellet was suspended in re-suspension buffer (5 mM potassium phosphate buffer [pH 7.8], 0.33 M sucrose, 10% (v/v) glycerol, 50 mM KCl, 0.1 mM EDTA, 2 mM DTT, 1 µg/ml leupeptin, and 1 µg/ml pepstain A). H+-pumping activity was detected by a decrease of acridine orange absorbance at 495 nm [38]. The assay buffer contained 20 mM MES-KOH (pH 7.0), 140 mM KCl, 3 mM ATPNa2, 30 µM acridine orange, 0.05% Brij 58, and 50 µg of plasma membrane protein in a total volume of 1 ml. Membranes were pre-incubated at 25°C for 5 min in assay buffer. The assay was initiated by the addition of 3 mM MgSO4. To determine if purified RIN4 protein could alter H+-pumping activity in vitro, 3 µg of purified recombinant RIN4 protein was added to the assay medium and pre-incubated at 25°C for 10 min before the addition of MgSO4. Recombinant RIN4 protein was expressed in E. coli and purified by Ni+ affinity chromatography as described previously [14].
The Bradford assay was used to calculate total plasma membrane protein content [67]. Each experiment was repeated two times with independent plasma membrane isolations.
The yeast strain AH109, containing the HIS3 and lacZ reporter genes, was used for yeast two-hybrid analyses (Matchmaker, Clontech). The coding sequence of RIN4, AHA1(837–950), and AHA2(837–949) fragments were obtained by PCR amplification and sequenced. The RIN4 PCR product was cleaved and cloned into the BamH I/Pst I site of the pGBKT7 vector (binding domain). AHA1(837–950) and AHA2(837–949) PCR products were cloned into the EcoR I/Xho I sites of pGADT7 vector (activation domain). pGBKT7-RIN4, pGADT7-AHA1(837–950), pGADT7-AHA2(837–949), the positive control pGAL4 and the negative control pGBKT7 vector were all transformed into the yeast strain AH109 following the manufacturer's protocol. Protein expression was detected in transformed strains by immunoblotting. Transformants were dilution plated onto yeast potato dextrose agar (YPDA) and synthetic dextrose lacking leucine/tryptophan/histidine (SD-3). Yeast growth was examined as previously described [28].
SDS-PAGE and subsequent immunoblotting were performed according to standard procedures [69]. RIN4 immunoblots were performed with affinity purified rabbit polyclonal anti-RIN4 at a concentration of 1∶1,000. AHA immunoblots were performed with rabbit polyclonal anti-AHA antisera at a concentration of 1∶5,000. The AHA antibody was raised against a C-terminal peptide of AHA2 (amino acids 852–949) [41]. Secondary goat anti-rabbit IgG-HRP conjugate (Biorad) was used at a concentration of 1∶3,000 for detection via enhanced chemiluminescence (Pierce).
Protein complexes from nproRPS2:HA in rps2-101c and the rps2-101c/rin4 negative control were purified three separate times for identification by mass spectrometry. For protein complex purifications, all steps were carried out on ice or at 4°C. 5 g of leaf tissue were ground in liquid N2 and resuspended in 15 ml IP buffer (50 mM HEPES, 50 mM NaCl, 10 mM EDTA, 0.2% Triton X-100, pH 7.5). Debris was removed from the lysate by centrifugation at 10,000g, 10 min. The supernatant was filtered through a 0.45-µm low-protein binding filter (Millipore) and incubated with 0.5 ml of affinity-purified RIN4 antisera coupled to Protein A beads (GE Healthcare). RIN4 antiserum was affinity purified according to standard protocols and 2 mg of antibody was coupled per ml of Protein A with dimethylpimelimidate [69]. The mixture was incubated end-over-end in batch format for 3 h then poured into a 20-ml glass column. Immunocomplexes were washed twice with 20 ml of wash buffer A (50 mM HEPES, 50 mM NaCl, 10 mM EDTA, 0.1% Triton X-100, pH 7.5), then twice with wash buffer B (50 mM HEPES, 150 mM NaCl, 10 mM EDTA, 0.1% Triton X-100, pH 7.5). Immunocomplexes were then washed with 5 ml of phosphate buffer (10 mM Na2PO4, 50 mM NaCl [pH6.8]) and eluted in 3×1 ml of low pH buffer (50 mM Glycine-Cl [pH2.5], 50 mM NaCl, 0.1% Triton X-100). The eluted proteins were neutralized, concentrated to a final volume of 30 µl with StrataClean resin (Stratagene), and loaded onto a single lane on a 10% SDS-PAGE gel. Proteins were run 5 mm into the separating gel and stained with colloidal coomassie blue. The resulting gel blobs were excised from the SDS-PAGE gel using a sterile blade.
Total RNA was extracted by a QIAGEN RNeasy Plant Mini kit and subjected to Dnase I digestion (Invitrogen). The first strand cDNA was synthesized by using 5 µg of total RNA with a cDNA synthesis kit (Promega) in a 20-µl reaction, and the reaction without reverse transcriptase served as a non-RT control. The expression level of the following genes RIN4 (AT3G25070), EDS1 (AT3G48090), PAD4 (AT3G52430), NDR1 (AT3G20600), EFR (AT5G20480), and CERK1 (AT3G21630) were normalized to the expression of Actin2 (AT3G18780). RT-PCR was run for 28 cycles. The primers for all genes are listed in Table S3.
Guard cell protoplasts were isolated enzymatically from the lower leaf epidermis according to a previously described method [73]. 100–150 rosette leaves were used. Purified guard cells were visually inspected for purity by light microscopy. Guard cells were immediately used for RNA and protein extraction. Cellulose R-10 and Macerozyme R-10 were purchased from Yakult Honsha Corporation. Nylon meshes were purchased from Spectrum Laboratories, Inc.
The AHA2:GUS construct contained a 2,000-bp AHA2 promoter fragment cloned into pCAMBIA 1303. AHA2 localization in the roots of the plant lines are previously described [60].
|
10.1371/journal.pgen.1004200 | Lsd1 Restricts the Number of Germline Stem Cells by Regulating Multiple Targets in Escort Cells | Specialized microenvironments called niches regulate tissue homeostasis by controlling the balance between stem cell self-renewal and the differentiation of stem cell daughters. However the mechanisms that govern the formation, size and signaling of in vivo niches remain poorly understood. Loss of the highly conserved histone demethylase Lsd1 in Drosophila escort cells results in increased BMP signaling outside the cap cell niche and an expanded germline stem cell (GSC) phenotype. Here we present evidence that loss of Lsd1 also results in gradual changes in escort cell morphology and their eventual death. To better characterize the function of Lsd1 in different cell populations within the ovary, we performed Chromatin immunoprecipitation coupled with massive parallel sequencing (ChIP-seq). This analysis shows that Lsd1 associates with a surprisingly limited number of sites in escort cells and fewer, and often, different sites in cap cells. These findings indicate that Lsd1 exhibits highly selective binding that depends greatly on specific cellular contexts. Lsd1 does not directly target the dpp locus in escort cells. Instead, Lsd1 regulates engrailed expression and disruption of engrailed and its putative downstream target hedgehog suppress the Lsd1 mutant phenotype. Interestingly, over-expression of engrailed, but not hedgehog, results in an expansion of GSC cells, marked by the expansion of BMP signaling. Knockdown of other potential direct Lsd1 target genes, not obviously linked to BMP signaling, also partially suppresses the Lsd1 mutant phenotype. These results suggest that Lsd1 restricts the number of GSC-like cells by regulating a diverse group of genes and provide further evidence that escort cell function must be carefully controlled during development and adulthood to ensure proper germline differentiation.
| The mechanisms that govern the formation, size and signaling output of in vivo niches remain poorly understood. Studies of Drosophila germline stem cells (GSCs) have suggested that chromatin programming greatly influences the behavior of these cells and their progeny. Previous work has shown that loss of the highly conserved histone demethylase Lsd1 results in ectopic niche signaling and an expanded GSC phenotype. To determine direct regulatory targets of Lsd1, we employed chromatin immunoprecipitation coupled with massive parallel sequencing (ChIP-seq) using specific cell populations inside and outside of the GSC niche. These experiments revealed that Lsd1 exhibits highly enriched binding to over one hundred genomic sites within a specific cell population. Furthermore, mis-regulation of some of these direct targets contributes to the expanded stem cell phenotype observed in Lsd1 mutants. These results provide insights into how Lsd1 directly restricts the size of the GSC microenvironment and establish a platform for understanding and exploring chromatin programming inside and outside an in vivo stem cell niche.
| Stem cells undergo self-renewing divisions in which at least one daughter retains its stem cell identity, while the second daughter may or may not differentiate, depending on intrinsic and extrinsic cues. A balance between stem cell self-renewal and differentiation must be maintained for proper organ formation during development and tissue homeostasis in adulthood. Stem cells often reside in microenvironments called niches, and specific mechanisms tightly regulate the size and signaling output of these structures [1]. However, in vivo niches have often proven difficult to identify in mammalian tissues. As a result, much of the current understanding of niches stems from the study of invertebrate models such as the germline stem cells (GSCs) of the Drosophila ovary.
Drosophila female GSCs reside in a well-characterized niche at the tip of a structure called a germarium (Figure 1A). Within germaria, GSCs lie immediately next to a somatic cell niche comprised of cap cells and terminal filament cells [2]. Escort cells reside adjacent to the cap cells and line the anterior portion of the germarium. These cells act to shepherd the germ cells during the earliest stages of their differentiation [3], [4], after which developing germline cysts are enveloped by follicle cells derived from a second stem cell population within the germarium. Cap cells produce Decapentaplegic (Dpp), which in turn activates a canonical Bone Morphogenic Protein (BMP) signal transduction pathway in GSCs [5], [6]. BMP pathway activation results in the transcriptional repression of bag of marbles (bam) [7]–[9], a factor both necessary and sufficient for germ cell differentiation [10], [11]. Ectopic Dpp signaling outside the tip of the germarium results in an expanded GSC phenotype [5], [9].
Other pathways and neighboring cells likely regulate niche specific BMP signaling. For example, a recent study provides evidence that hedgehog (hh) produced by the cap cells stimulates the anterior escort cells to produce niche specific signals [12] Moreover, several additional intrinsic and extrinsic mechanisms help restrict Dpp ligand production and diffusion within the niche (reviewed in [13], [14]). One such mechanism involves the histone demethylase Lysine Specific Demethylase 1 (Lsd1). Lsd1 uses a flavin-dependent monoamine oxidative based mechanism to remove mono- and di-methyl groups from histone H3 on lysine 4 (H3K4me1 and H3K4me2) [15]. In mammals, Lsd1 has been shown to silence a number of distinct gene sets in different cellular contexts, including Notch targets, TGFβ-1 and various loci involved in the maintenance of embryonic stem cells [16]–[20]. Additional studies suggest that Lsd1 may also promote gene expression under certain circumstances [21]. Disruption of Drosophila Lsd1 results in a male and female sterile phenotype, marked by the expansion of GSC-like cells in the germarium [22], [23]. These cells exhibit ectopic BMP responsiveness and fail to initiate a normal differentiation program once they leave the cap cell niche [24].
To characterize the molecular mechanism by which Lsd1 restricts signaling outside the Drosophila female GSC niche, we used ChIP-seq to define direct binding sites of Lsd1 specifically in either escort cells or cap cells. These experiments revealed that Lsd1 binds to over one hundred sites in escort cells and provide further insights into how Lsd1 contributes to the chromatin programming of cells inside and outside of an in vivo niche.
Escort cells send out extensions that closely contact germline cysts undergoing the early steps of differentiation [3], [4]. Escort cell death or genetically disrupting escort cell extensions can lead to an inappropriate expansion of GSC-like cells in the germarium [3]. Previous results showed that Lsd1 functions within escort cells to prevent expanded BMP signaling outside of the GSC niche [24]. This phenotype was accompanied by widespread cell death in both somatic cells and germ cells. Therefore we considered the possibility that the expansion of BMP signaling exhibited by Lsd1 mutants may depend on changes to escort cell morphology and number. To test this, we knocked down the expression of Lsd1 specifically in the escort cells and early follicle cells by crossing UAS-Lsd1RNAi into a c587-gal4; UAS-mCD8::GFP background and stained the resulting ovaries for GFP and the fusome marker Hts. Fusomes are highly vesiculated organelles that appear round in GSCs and cystoblasts, and become branched as these germ cells differentiate into multi-cellular cysts [25], [26]. Three days after eclosion, control samples appeared normal. These germaria typically contained two to four single cells (GSCs and cystoblasts) with round fusomes and escort cells that extended cytoplasmic processes between the developing cysts (Figure 1B). In contrast, the Lsd1 RNAi samples showed an expansion of GSC-like cells with round fusomes. Escort cell extensions were clearly present in some germaria, but were missing in others (Figure 1C,D). These observations suggested that while the knockdown of Lsd1 caused changes in escort cell morphology, the presence of extra single cells with round fusomes did not absolutely depend on a complete loss of escort cell extensions. However changes in escort cell morphology likely contributed to the phenotype over time. In addition, expression of Lsd1RNAi also led to an increase of death within escort cells, consistent with the widespread cell death previously noted in Lsd1 null mutant germaria (Figure 1E,F) [24].
Next we performed clonal analysis using the mosaic analysis with a repressible cell marker (MARCM) system to further analyze the Lsd1 mutant phenotype. Clonal germaria were stained for the positive clone marker GFP and for the fusome marker Hts. We categorized the relative position of escort cell clones along the anterior-posterior axis of the germarium. Cells were considered anterior escort cells if they were immediately next to the cap cells, posterior escort cells if they were immediately adjacent to the follicle stem cells and middle escort cells if they were located in any position in between. We induced control and Lsd1 mutant clones in parallel. Control escort cell clones were never associated with an obvious robust germline tumor phenotype, although we noted one exception in which a single germarium with control escort cell clones contained six single germline cells with round fusomes (1 out of 143 counted). By contrast, 17% (27/155) of the germaria that contained Lsd1 mutant escort cell clones displayed an expanded germline stem cell-like cell phenotype (Figure 1G–J). The relative position of Lsd1 mutant clones appeared to correlate with the appearance of a germline phenotype. The vast majority of germaria (96%; n = 27) that contained greater than 5 germline stem cell-like cells carried at least one Lsd1 mutant middle escort cell clone. We observed one example in which a germarium with a mild expansion of GSC-like cells contained an Lsd1 mutant anterior clone and posterior clone but no middle escort cell clones (Figure 1H). Of note, most germaria that carried middle escort cell clones did not exhibit a GSC expansion phenotype (98/125 germaria). While their appearance was rare, germaria with only anteriorly or posteriorly (Figure 1K) positioned escort cell clones did not display a robust GSC-like cell expansion phenotype. Similarly, loss of Lsd1 in the terminal filament did not result in an obvious phenotype (Figure 1L).
Germaria that contained Lsd1 mutant escort cell clones and exhibited an increased number of GSCs occasionally had an elongated and abnormal morphology (Figure 1J). Moreover, 22.1% of the germaria (n = 199) from Lsd1 mutant females lacked marked escort cell clones, compared to 13.9% of control germaria (n = 166), and the average number of Lsd1 mutant clones per germarium (4.72 escort cell clones/germarium) was lower compared to controls (8.77 escort cell clones/germarium), suggesting that either Lsd1 mutant escort cell progenitors exhibited reduced proliferation during development or died during the course of the experiment. If increased death within the escort cell population accounted for all the observed phenotypes, one might predict that complete loss of all Lsd1 mutant escort cell clones within a particular germarium would result in an increased number of GSC-like cells. However, we did not observe an expanded GSC phenotype in germaria that lacked clones from Lsd1 mutant females. Together all the phenotypic data suggest both that escort cells require Lsd1 function to limit GSC number and that loss of Lsd1 compromises the growth and survival of escort cells, consistent with previous observations [24] and those noted above, which in turn further exacerbates the observed germ cell phenotypes.
To directly define the molecular mechanisms by which Lsd1 influences escort cell function, we elected to identify direct targets of Lsd1 regulation in these cells. Determining whether Lsd1 targeted potential candidate genes represented a significant challenge. For example, the size and complexity of the dpp promoter precluded our ability to assay whether Lsd1 directly targeted this gene using a PCR based chromatin immunoprecipitation (ChIP) approach. To systematically define Lsd1 binding sites, we conducted ChIP experiments coupled with massive parallel sequencing (ChIP-seq). We used a number of different Hemagglutinin (HA) tagged transgenes, including a N-terminally tagged UASt-HA::Lsd1 transgene that exhibits high expression in the somatic cells and a N-terminally tagged UASp-HA::Lsd1 transgene that displays relatively lower levels of expression in the somatic cells (Figure 2; S1). The UASt-HA::Lsd1 and UASp-HA::Lsd1 transgenes both fully rescued the Lsd1ΔN GSC tumor phenotype when driven by the c587-gal4 driver. Because Lsd1 is expressed ubiquitously throughout the ovary [24], we sought to determine whether this protein bound to distinct sites in different cell populations within the germarium. We expressed the UASt-HA::Lsd1 and UASp-HA::Lsd1 trangenes in cap cells and terminal filament cells using hh-gal4 (Figure 2A; S1) and in the escort cells and early follicle cells using the c587-gal4 driver (Figure 2B; S1). HA-directed ChIP assays were performed on dissected ovaries and the immunoprecipitated chromatin was compared to input chromatin as a control. Within escort cells and early follicle cells, products of the UASp-HA::Lsd1 and UASt-HA::Lsd1 trangenes bound to 207 and 191 sites respectively (Based on the FindPeaks algorithm using a p-value threshold of 1.00e-3 to maximize the number of potential peaks; Table S1, S2), with 100 common sites sharing some degree of overlap (Figure 2C). Within cap cells and terminal filament cells, the UASp-HA::Lsd1 and UASt-HA::Lsd1 transgenic products associated with 98 and 167 genomic loci respectively (Table S3, S4), with 37 overlapping loci in common between the two datasets (Figure 2C). Comparing all four datasets revealed 66 common peaks between terminal filament/cap cells and escort cells/early follicle cells (Figure 2C,D). 232 peaks appeared specific for escort cells and early follicle cells and 162 specific for cap cells and terminal filament cells (Figure 2C,D). MACs analysis [27] showed similar but broader peak calls (Table S5, S6, S7, S8). Lsd1 enrichment peaks were spread throughout the Drosophila genome (Figure S2) and showed a preference for the promoter and 5′UTR regions of genes (Figure S3). We were unable to isolate a sufficient number of cells to map H3K4 methylation across the escort cell genome. However, comparing our data with available data from the modENCODE project revealed that Lsd1 binding peaks correlate with valleys of H3K4 methylation in embryos (Figure S4), consistent with the established biochemical activity of Lsd1.
Strikingly, we did not observe any enrichment for Lsd1 binding near the dpp locus in escort cells (Figure 2E), indicating that the repression of BMP signal transduction by Lsd1 must be through an indirect mechanism. We examined the annotation of genes near escort cell and early follicle cell peaks, cap cell and terminal filament peaks, shared peaks and from the UASt-HA::Lsd1 data sets (Table S9, S10, S11, S12). This analysis indicated that genes near escort cell specific Lsd1 binding peaks encode products with a diverse array of functions needed for development, basic cellular processes and reproduction (Figure S5A) [28], [29]. Further analysis of this gene set did not reveal significant enrichment for components of specific pathways. MEME analysis showed an enrichment of ACTGGAA elements within Lsd1 binding sites (Figure S5B). The significance of this enrichment remains unclear. These results suggested that the mis-expression of a functionally diverse set of genes likely contributes to the Lsd1 mutant phenotypes.
The engrailed gene stood out as one potentially relevant target among the list of candidate genes. engrailed encodes a homeobox transcription factor that acts as a segment polarity gene during embryogenesis [30]–[32]. Previous results showed that engrailed regulates early ovarian development and that Engrailed protein expression is restricted to the terminal filament and cap cells in adult germaria [33]. Engrailed functions within these cells to help maintain GSCs [12]. Our ChIP-seq data revealed that Lsd1 exhibits enriched binding to a 2 kb region of the engrailed promoter in the escort cells (Figure 3A). We performed RNA RT-qPCR to look at the transcript levels of engrailed in Lsd1 mutants. We compared bam mutants to bam Lsd1 double mutants because these samples are comparable in size and have the same basic cellular makeup (Figure S6). This analysis revealed that engrailed transcript levels increased 6 fold in the absence of Lsd1 (Figure 3B).
Next, we tested whether Engrailed protein expression expanded in the absence of Lsd1. In wild type germaria, cap cells and terminal filament cells express readily detectable levels of Engrailed, whereas the escort cells do not (Figure 3C) [12], [33]. In Lsd1ΔN mutants, however, we observed ectopic Engrailed protein expression in a limited number of escort cells, in addition to the terminal filament and cap cells, in 85.7% (n = 91) of the germaria examined (Figure 3D). These Engrailed expressing escort cells tended to be positioned immediately adjacent to the normal niche, although occasionally we observed Engrailed expressing escort cells several cell positions away from the cap cells (Figure 3E). We cannot rule out the possibility that other cells also mis-expressed Engrailed, but at a level below our detection threshold. These data together suggest that Lsd1 serves to repress engrailed expression within a subpopulation of escort cells.
To test the functional relevance of ectopic Engrailed expression in escort cells, we assayed whether disruption of engrailed function, either through RNAi knockdown or loss-of-function mutations, modified the Lsd1 mutant phenotype. Knockdown of engrailed in an Lsd1 RNAi background (Figure 4A) suppressed the expanded GSC-like cell Lsd1 mutant phenotype (Figure 4B,E,F). Furthermore, three independent engrailed mutations also suppressed the Lsd1 RNAi-induced phenotype, so that the number of single cells with round fusomes decreased and cyst development within the germarium proceeded normally (Figure 4C–F). In all cases, engrailed suppression of the Lsd1 RNAi phenotype resulted in the formation of morphologically normal ovarioles with maturing egg chambers.
In Drosophila, Engrailed drives the expression of hedgehog (hh), which in turn leads to the expression of dpp in adjacent cells [34]–[37]. Previous analysis showed an expansion of hh expression in Lsd1 mutant germaria [24]. To determine whether the mis-expression of hh in escort cells contributed to the Lsd1 mutant phenotype, we crossed both hh-specific UAS RNAi and hh mutant lines into a c587-gal4>UAS-Lsd1RNAi background. This analysis revealed that loss of hh function, similar to engrailed, suppressed the GSC-like expansion phenotype, resulting in the formation of germaria that exhibited normal germ cell differentiation (Figure 4G–J).
To assess whether mis-expression of engrailed and hh are sufficient to expand the number of stem cell-like cells in the germarium, we expressed transgenes corresponding to both genes within escort cells and early follicle cells using the c587-gal4 driver. Similar to the phenotype exhibited by Lsd1 mutants, ectopic expression of engrailed resulted in a stem cell-like cell expansion within 49.5% of germaria examined (n = 111). Many of these germline cells remained as single cells with round fusomes (Figure 5A). However, mis-expression of engrailed did not completely block cyst development and many ovarioles from c587-gal4>UAS-engrailed females contained maturing egg chambers.
In contrast to engrailed, over-expression of hh using two different transgenes did not obviously perturb early germ cell differentiation (Figure 5G,H). However, the mis-expression of these transgenes did result in follicle cell defects, consistent with previously published results [38]. These results indicated that the hh transgene is active in these cells but that hh over-expression in the escort cells and early follicle cells is not sufficient to induce an expansion of GSC-like cells in germaria.
Loss of Lsd1 results in expanded BMP signaling within the germline [24]. Based on the expansion of Engrailed expression in Lsd1 mutants and the similarities between the Lsd1 mutant and engrailed over-expression phenotypes, we reasoned that mis-expression of engrailed may also induce ectopic BMP signaling in the ovary. To test this, we used a Dad-lacZ enhancer trap as a positive transcriptional reporter of dpp signal transduction [9], [39]. In control ovarioles, stem cells express high levels of Dad-LacZ, whereas the expression of this reporter sharply decreases in differentiating cysts (Figure 5B). Upon engrailed mis-expression in the escort cells, the number of Dad-LacZ positive germline cells increases, likely reflecting expanded Dpp signaling (Figure 5C). Next, we knocked down the expression of dpp in the presence of the engrailed transgene and found that disruption of dpp suppressed the engrailed over-expression phenotype (Figure 5D–F). Together these results suggest that mis-expression of engrailed in Lsd1 mutants drives ectopic BMP signaling, resulting in the expanded GSC-like cell tumor phenotype.
To test whether ectopic engrailed expression can specifically affect adult escort cells, we performed a temporally controlled over-expression experiment. c587-gal4>UAS-engrailed larvae were kept at low temperature to prevent robust expression of the engrailed transgene. Ovaries from adult females maintained at a low temperature did not display ectopic Engrailed expression or an expanded undifferentiated cell phenotype (Figure 6A,A′). However, shifting c587-gal4>UAS-engrailed females to a higher temperature after eclosion resulted in ectopic engrailed expression in escort cells and early follicle cells, and a concomitant expansion of germline stem cell-like cells (Figure 6B–C). Thus, engrailed expression specifically in adults appears sufficient to induce ectopic BMP signaling in the anterior region of the germarium.
Lsd1 associates with the promoters of many genes besides engrailed, some of which could potentially play a role in regulating escort cell function. To begin to characterize whether Lsd1 modulates the expression of other potential target genes, we stained c587-gal4 control and c587-gal4>UAS-Lsd1RNAi ovaries using available antibodies. Cap cells and escort cells exhibit a shared peak of Lsd1 binding near the Rho1 gene (Table S11, 12). Previous results showed that loss of Rho1 results in escort cell defects and an expansion of GSC-like cells [3]. Knocking down Lsd1 levels did not appear to result in any dramatic change in Rho1 expression within the escort cells (compare Figure 7A and 7D). Likewise, Lsd1 also exhibits enriched binding near Apc1 (Tables S9, S12), a component of the Wnt signaling pathway. However antibody staining showed that Apc1 protein levels did not change to an appreciable degree upon knock-down of Lsd1 (Figure 7B,E). In contrast, the product of a third potential Lsd1 target gene, Broad, exhibited increased expression in c587-gal4>UAS-Lsd1RNAi samples relative to controls (Figure 7C,F). However, loss of broad did not appear to suppress the Lsd1 mutant phenotype (data not shown).
The raw gene functions in the developing gonad to regulate the morphology of somatic cells as they interact with primordial germ cells [40], [41]. Lsd1 exhibits enriched binding just 3′ to the raw transcription termination site (Figure 7G). Antibodies were not available to assay whether loss of Lsd1 caused a change in Raw expression levels but raw mutant and RNAi lines partially suppressed the Lsd1 phenotype (Figure 7H–L). The raw134.47 allele weakly modified the GSC-like cell expansion phenotype exhibited by c587-gal4>UAS-Lsd1RNAi germaria, while both rawRNAi and a single copy of raw155.27 more strongly suppressed the c587-gal4>UAS-Lsd1RNAi phenotype, giving rise to a reduced number of germaria that carried more than 5 single cells with round fusomes. These genetic interactions suggest that mis-regulation of raw expression or activity also contributes to the Lsd1 mutant phenotype.
Encouraged by the finding that disruption of raw suppressed the Lsd1 mutant phenotype, we crossed a number of additional knockdown lines, corresponding to other putative Lsd1 target genes, into the c587-gal4>UAS-Lsd1RNAi background. We counted the total number of single germ cells with round fusomes within individual germaria from the resulting ovaries. This analysis showed that knockdown of 7 out of the 34 genes tested suppressed the c587-gal4>UAS-Lsd1RNAi phenotype to the point where fewer than 50% of the assayed germaria contained greater than 5 single cells with round fusomes (Figure 8A). This group included FK506-binding protein 1 (FK506-bp1), Glutamine:fructose-6-phosphate aminotransferase 1(Gfat1), CG11779, ken and barbie (ken), Anaphase Promoting Complex subunit 7 (APC7), barren (barr) and Hepatocyte nuclear factor 4 (Hnf4). These genes have varied functions and play roles in cell cycle regulation (APC7 and barr), juvenile hormone signaling (FK506-bp1), development of the genitalia (ken) and lipid metabolism (Hnf4). Lack of a clear functional link between these suppressors suggests that escort cells are particularly sensitive to perturbations in their gene expression programs. Together these data show that disruption of Lsd1 results in a complex phenotype, marked by increased BMP signaling in the germline and disruption of normal escort cell function, which likely involves the mis-expression of several direct and potentially indirect target genes (Figure 8B).
Using ChIP-seq, we have systematically identified Lsd1 binding sites throughout the genomes of the cap cells/terminal filament cells and the escort cells/early follicle cells of the Drosophila ovary. The establishment of ChIP-based techniques using these specific cell populations and this comprehensive Lsd1 genome-wide dataset will facilitate further studies on the transcriptional hierarchies that regulate escort cell and cap cell function. Although Lsd1 mutants exhibit increased dpp expression [23], Lsd1 does not appear to directly target the dpp locus for silencing. Rather, Lsd1 represses the expression of the transcription factor engrailed. Mutations in the engrailed gene suppress the ovarian Lsd1 mutant phenotype and mis-expression of engrailed in escort cells results in increased BMP signaling and an expansion of undifferentiated germline cells within germaria. Thus engrailed may regulate dpp expression in germaria through either hh-dependent or independent mechanisms. Lsd1 also regulates additional genes adjacent to its other binding sites. Genetic analysis shows that the mis-regulation of these other target genes likely contributes to the Lsd1 mutant phenotype to varying degrees.
We have found Lsd1 associates with a limited number of loci within two different cell populations. Lsd1 exhibits fairly broad peaks of binding, ranging in size from 166–262 bp based on the FindPeaks algorithm. MACs analysis calls even wider peaks (Supplementary Tables S5, S6, S7, S8). The significance of the width of these peaks remains unclear but suggests that Lsd1 either does not associate with single sequence specific elements at these sites or exhibits a certain degree of spreading upon recruitment to a particular locus. In Drosophila escort cells, Lsd1 binds to over 100 sites spread throughout the genome. Lsd1 binds to fewer sites in cap cells. While some Lsd1 binding sites overlap in cap cells and escort cells, the relatively large number of different sites suggests that Lsd1 recruitment depends on multiple and perhaps distinct cell-specific co-factors. MEME analysis [42], [43] reveals that many of the identified Lsd1 binding sites contain ACTGGAA elements. GGAA sequences are often present in the core binding sites of ETS transcription factors. The Drosophila genome encodes a number of ETS family members, none of which have been characterized in the somatic cells of the germarium. Determining the functional relevance of these specific GGAA sites within gene promoters and identifying the transcription factors that bind to them will require additional efforts.
For technical reasons and to enable comparisons of Lsd1 binding between escort cells/early follicle cells and cap cells/terminal filament cells, we elected to express the Lsd1 HA-tagged transgenes in an otherwise wild-type background. We acknowledge the possibility that endogenous Lsd1 may outcompete the HA-tagged transgenes for binding at specific sites in the escort cells and early follicle cells. Therefore sites identified in this study may be an underrepresentation of the total number of sites bound by endogenous Lsd1. Repeating the ChIP-seq analysis using material from rescued Lsd1ΔN females that express the HA-tagged Lsd1 transgene in escort cells and early follicle cells represents important work for the future.
We found that Lsd1 mutant samples exhibit a 6-fold increase in engrailed transcript levels relative to controls. Curiously, ectopic Engrailed protein expression was only observed in a small number of escort cells. Perhaps additional post-transcription mechanisms regulate the translation of Engrailed, and potentially other proteins, inside and outside of the cap cell niche. Such mechanisms would allow these cells to fine-tune their signaling output more than what could be achieved through transcriptional based mechanisms alone. Results presented here also suggest that escort cells are not uniform in nature and perhaps have specific functions or capabilities depending on their lineage and where they reside within the germarium. MARCM analysis shows that the loss of Lsd1 in some but not all escort cells results in a marked expansion of GSC-like cells within the germarium. Previous studies have also suggested that specific escort cells have distinct roles in supporting GSCs [12].
Technical considerations aside, the severity of the Lsd1 null phenotype compared to the engrailed over-expression phenotype, both in terms of the penetrance and extent of the GSC expansion phenotype and the accompanying germline and somatic cell death, suggests that engrailed is not the only biologically relevant target of Lsd1 regulation in the escort cells. Based on expression analysis (Figure 7), Lsd1 regulates the expression of some but not all genes adjacent to its binding sites. Our genetic analysis suggests that mis-regulation of the putative target raw (Figure 7) and several additional genes (Figure 8) also contribute to the Lsd1 mutant phenotype. Characterizing the transcriptional profile of escort cells from wild-type and Lsd1 mutant samples, which will have to await for improvements in current cell isolation and RNA profiling techniques, will help to further resolve which genes are direct and indirect targets of Lsd1 regulation. Such approaches may also reveal additional genes that participate in niche formation and function. Of note, the observation that functionally diverse genes can suppress the Lsd1 mutant phenotype suggests that escort cells are acutely sensitive to changes in their gene expression programs. While our data support a model that loss of Lsd1 initially results in mis-regulation of engrailed and other genes that, in turn, drive GSC expansion, it is clear that many of the escort cells that experience reduced Lsd1 function retract their cellular extensions and undergo cell death. This loss of escort cells further exacerbates the GSC expansion phenotype. Given the phenotypic complexity described here and elsewhere [3], care should be taken when analyzing gene function within the escort cell population.
How niches maintain stem cells and adjust their signaling output to ensure tissue homeostasis remains a fundamental question in stem cell biology. Elegant work has shown that terminal filament cells, cap cells and escort cells help to support the self-renewal of two to three germline stem cells at the tip of Drosophila germaria [5], [44]. The predominant signal emanating from the anterior tip of the germarium is Dpp, which acts locally to induce a canonical signal transduction cascade in GSCs, which in turn represses their differentiation [5], [9]. Several expression and genetic studies strongly suggest that terminal filament and cap cells, and perhaps the most anterior escort cells, are the primary source Dpp ligand [5], [9]. More recent work has suggested that Engrailed expression in cap cells non-autonomously promotes dpp expression in escort cells through a hedgehog dependent mechanism [12]. Loss of Lsd1 results in ectopic expansion of hh expression in escort cells [24] and data shown here (Figure 4) reveals that disruption of hh partially suppresses the Lsd1 mutant phenotype. However, consistent with previous results [38], over-expression of hh in escort cells does not result in an Lsd1-like mutant tumor phenotype (Figure 5G,H), demonstrating that hh is not sufficient to induce ectopic BMP signaling in the germarium. Given these observations, ectopic engrailed expression in escort cells likely targets additional genes besides hh to induce ectopic BMP signaling and promote the expansion of undifferentiated germ cells.
The finding that loss of Lsd1 or mis-expression of engrailed in adult escort cells leads to expanded Dpp signal transduction within germ cells throughout the anterior portion of the germarium indicates that subpopulations of escort cells are capable, and perhaps even poised, to express dpp under certain conditions. Such plasticity might allow the niche to expand and contract in response to various stimuli and environmental cues. Indeed, previous studies have shown that Jak/Stat and insulin signaling can influence the number of GSCs in the ovary [45]–[49]. Moreover, ongoing dynamic regulation of signaling may be a regular feature of niches under resting homeostatic conditions. The observation that long-term knock-down of dpp in escort cells results in a reduced number of GSCs at the tip of the germarium, but not their complete elimination, is consistent with the notion that escort cells contribute to the maintenance of GSCs in some manner [12]. Further work, with single-cell spatial and small-scale temporal resolution, will be needed to help clarify what cells express niche signals and when.
Inappropriate and extensive expansion of niches would be predicted to upset tissue homeostasis and perhaps even result in pathological conditions. Therefore robust but flexible mechanisms that depend on chromatin factors such as Lsd1 may be in place to precisely control the expansion and contraction of in vivo stem cell niches. The continued study of Drosophila cap cells and escort cells will provide further insights into how chromatin programming regulates niche plasticity.
Drosophila stocks were maintained at room temperature on standard cornmeal-agar medium unless specified otherwise. The following fly strains were used in this study: w1118 was used as a control; Lsd1ΔN was provided by N. Dyson (Massachusetts General Hospital Cancer Center, Charlestown, MA); hh-gal4 and UAS-hh lines [50] were provided by J. Jiang (University of Texas Southwestern, Dallas, TX); c587-gal4 and Dad-LacZ were provided by A. Spradling (Carnegie Institution for Science, Baltimore, MD); the UAS-engrailed::GFP transgenic line was provided by Florence Maschat (Institute of Human Genetics, France); the raw134.47 and raw155.27 alleles were provided by Jennifer Mierisch (Loyola University of Chicago); en7, en4, enspt, UAS-mCD8::GFP, UAS-dppRNAi-1 (BL#-31172), UAS-dppRNAi-2 (BL#-31530), UAS-dppRNAi-3 (BL#-31531) and UAS-rawRNAi (BL#-31393), UAS-enRNAi-1 (BL#-33715), UAS-enRNAi-2 (BL#-26752), UAS-hhRNAi (BL#-31042), hhAC (BL#-1749), hh2 (BL#-3376), broadnpr-3 (BL#-29971), broad5 (BL#-29972), par-1HMS00405 (BL#- 32410), NaPi-THMS00966 (BL#- 34003), CG17186JF02425 (BL#- 27079), CG12054HMJ03134 (BL#- 50910), Atg1HMS02750 (BL#- 44034), mudHMS01458 (BL#- 35044), CG12128HMS00960 (BL#- 33997), Nek2HM05088 (BL#- 28600), GstS1HM05063 (BL#- 28885), RhoGEF2HMS01118 (BL#- 34643), lwrHMS01648 (BL#- 37506), UGPHMJ03120 (BL#- 50902), Nmdar2HMS02176 (BL#- 40928), FdhHMS01268 (BL#- 34937), SpredHMS00637 (BL#- 32852), CG10949JF02129 (BL#- 26231), CG13192HMS01384 (BL#- 34390), LaspJF02075 (BL#- 26305), ApcHMS00188 (BL#- 34869), CycAGLV21059 (BL#- 35694), ari-1JF03352 (BL#- 29416), CG16989HMJ21141 (BL#- 51017), PezHMS00862 (BL#- 33919), AcCoASHMS02314 (BL#- 41917), Pdk1HMS01250 (BL#- 34936), LrchHMJ03119 (BL#- 50901), DabHMS02482 (BL#- 42646), FK506-bp1HMS00339 (BL#- 32348), Gfat1HMS02585 (BL#- 42892), CG11779GL01496 (BL#- 43155), kenHMS01219 (BL#- 34739), APC7GL01114 (BL#- 38932), barrHMS00049 (BL#- 34068) and Hnf4JF02539 (BL#- 29375) lines were obtained from the Bloomington Stock Center. UAS-Lsd1RNAi was obtained from the National Institute of Genetics, Japan.
Lsd1 mutant MARCM clones were generated by crossing Lsd1ΔN FRT 2A to yw122 tub-gal4 UAS-GFP;;tub-gal80 FRT 2A/TM6B (gift from Ben Ohlstein). The resulting larvae were heat-shocked twice a day at 37°C on days 5–7 after the cross was set. The resulting adult females were dissected and stained 7 days after they eclosed.
The HA tagged transgenes of Lsd1 were created using Gateway Cloning (Invitrogen). The open reading frame (ORF) of Lsd1 was cloned into modified pTHW and pPHW destination vectors (http://emb.carnegiescience.edu/labs/murphy/Gateway%20vectors.html) that contained φC31 attB sites [51]–[53]. These constructs were injected into flies and transformed using φC31 integrase into the 51D landing site on the 2nd chromosome.
Adult ovaries were dissected in Grace's medium and fixed in 4% (vol/vol) formaldehyde for 10 min. The ovaries were washed with PBT (1X PBS, 0.5% BSA, and 0.3% Triton-X 100) and stained with primary antibody overnight at 4°C. The ovaries were washed and incubated in secondary antibody at room temperature for 5 hrs. Ovaries were then washed again and mounted in Vectashield containing DAPI (Vector Laboratories). The following primary antibodies were used: mouse anti-Hts (1B1) (1∶20), mouse anti-Engrailed (4D9) (1∶2), mouse anti-Broad-core (25E9.D7) (1∶10), mouse anti-Rho1 (P1D9) (1∶50) and rat anti-VASA (1∶20) (Developmental Studies Hybridoma Bank), mouse anti-β-galactosidase (1∶200) (Promega), rabbit anti-APC1 [54] (1∶1000)(gift of E. Wieschaus), Rabbit anti-α-Spectrin [55](1∶1000)(gift from Ron Dubreuil), rabbit anti-GFP (1∶1000) (Molecular Probes), rat anti-HA 3F10 (Roche) and rabbit anti-cleaved Caspase-3 (1∶250) (Cell Signaling Technology). Fluorescence-conjugated secondary antibodies (Jackson Laboratories) were used at a dilution of 1∶200.
RNA was isolated from bamΔ86 and Lsd1ΔN bamΔ86 mutant ovaries using TRIzol (Invitrogen). The RNA was treated with DNase and subjected to RT-qPCR reaction using the Superscript III First-Strand Synthesis SuperMix (Invitrogen). The primers used to amplify engrailed mRNA are as follows:
engrailed forward: 5′ - GCCCGCCTGGGTGTACTG
engrailed reverse: 5′ - CGCTTCTCGTCGTTGGTCTTG
We used 1000 pairs of ovaries for each ChIP-seq reaction. Every 200 pairs of ovaries were dissected, fixed, washed and frozen immediately at −80°C. The entire protocol was done at 4°C unless otherwise indicated. The ovaries were dissected and fixed in 1 ml of 1% formaldehyde at room temperature for 10 mins. The crosslinking was stopped by the addition of 100 ml 1.25M glycine solution. The ovaries were washed three times with 1X cold PBS buffer and then sonicated in 500 µl ChIP Sonication Buffer (1%Triton X-100, 0.1% Deoxycholate, 50 mM Tris 8.1, 150 mM NaCl, 5 mM EDTA) on ice to achieve a final DNA length of 100 to 600 base pairs. The sample was centrifuged at maximum speed at 4°C to remove cell debris. The supernatant was transferred to a new tube and the sonicated sample was then blocked by adding Protein G agarose beads and incubating at 4°C for one hour. The beads were removed. 1% of the sample was kept aside as INPUT and to the remaining sample 3 ug rabbit-HA antibody (Abcam) was added and incubated overnight at 4°C.
The next day protein agarose G beads were added and incubated for 3 hours at 4°C. The beads was then washed well with ChIP Sonication Buffer (two times), High Salt Wash Buffer (1% Triton X-100, 0.1% Deoxycholate, 50 mM Tris 8.1, 500 mM NaCl, 5 mM EDTA) (three times), LiCl Immune Complex Wash Buffer (two times) and TE buffer. The protein bound to the beads was eluted using 500 µl Elution Buffer (1% SDS, 0.1M NaHCO3). The elution buffer was added to the INPUT samples and they were treated the same as the IP samples from this point. 20 µl of 5M NaCl was added to 500 µl of elution buffer and incubated at 65°C overnight.
The third day, the sample was treated with RNase A, Proteinase K and the DNA isolated using Qiagen PCR Purification kit. Subsequent library construction and sequencing of the input and immunoprecipitated DNA were conducted by the UT Southwestern McDermott Sequencing Center.
The primary sequencing data was mapped to the fly reference genome dm3 using BioScope (1.2.1). During the alignment, three filter steps were applied to remove low quality, ambiguous and redundant reads. HA-Lsd1 binding regions were identified as genomic regions with a significant read enrichment and binding peak profile in the ChIP reads over the input reads using the FindPeaks module in the Homer software tool [56] with 10% false discovery rate (FDR). ChIP enrichment at important genome features such as specific chromosomes, promoters, downstream, exonic, intronic and distal intergenic regions was statistically analyzed with the Cis-regulatory Element Annotation System (CEAS) [57]. De novo motif discovery analysis for HA-Lsd1 binding regions was performed with the Multiple EM for motif elicitation (MEME) software tool [42], [43]. High quality motifs were aligned against transcription factor motifs retrieved from JASPAR [58] and TRANSFAC [59] using the TOMTOM software tool [42] to identify known transcription factor motifs that match the MEME predicted motifs. Potential protein-coding target genes associated with the identified HA-Lsd1 binding regions were identified based on the distance of their transcription start sites (TSSs) according to their RefSeq annotation in the dm3 assembly to binding peak summits. Genes with TSSs within 5 kb or nearest to an HA-Lsd peak summit were called as target genes. ChIP-Seq data has been deposited with NCBI GEO (http://www.ncbi.nlm.nih.gov/geo) under the accession code GSE54376.
|
10.1371/journal.pgen.1001306 | Mapping of the Disease Locus and Identification of ADAMTS10 As a Candidate Gene in a Canine Model of Primary Open Angle Glaucoma | Primary open angle glaucoma (POAG) is a leading cause of blindness worldwide, with elevated intraocular pressure as an important risk factor. Increased resistance to outflow of aqueous humor through the trabecular meshwork causes elevated intraocular pressure, but the specific mechanisms are unknown. In this study, we used genome-wide SNP arrays to map the disease gene in a colony of Beagle dogs with inherited POAG to within a single 4 Mb locus on canine chromosome 20. The Beagle POAG locus is syntenic to a previously mapped human quantitative trait locus for intraocular pressure on human chromosome 19. Sequence capture and next-generation sequencing of the entire canine POAG locus revealed a total of 2,692 SNPs segregating with disease. Of the disease-segregating SNPs, 54 were within exons, 8 of which result in amino acid substitutions. The strongest candidate variant causes a glycine to arginine substitution in a highly conserved region of the metalloproteinase ADAMTS10. Western blotting revealed ADAMTS10 protein is preferentially expressed in the trabecular meshwork, supporting an effect of the variant specific to aqueous humor outflow. The Gly661Arg variant in ADAMTS10 found in the POAG Beagles suggests that altered processing of extracellular matrix and/or defects in microfibril structure or function may be involved in raising intraocular pressure, offering specific biochemical targets for future research and treatment strategies.
| Primary open angle glaucoma (POAG) is a leading cause of vision loss and blindness affecting tens of millions of people. Ocular hypertension is a strong risk factor for the disease and the only effective target of treatment. Ocular hypertension results from increased resistance to outflow of aqueous humor through the trabecular meshwork, a specialized filtration tissue consisting of alternating layers of cells and connective tissue, but the specific reasons for the increased resistance are not known. The animal model for human POAG used in this study was a colony of Beagle dogs that carry an inherited form of the disease in which ocular hypertension is the primary manifestation. We have found a variant in ADAMTS10 that belongs to a family of genes that contribute to formation of extracellular matrix and may itself be involved in formation of elastic microfiber structures. We found that the ADAMTS10 protein is expressed at particularly high levels in the trabecular meshwork. The candidate variant in ADAMTS10 found in the POAG–affected Beagles suggests that altered processing of connective tissue and/or elastic microfiber defects may be involved in raising eye pressure, offering specific biochemical targets for future research and treatment strategies.
| Elevated intraocular pressure is a strong risk factor for glaucoma development and progression [1]. In POAG, increased resistance to outflow of aqueous humor through the trabecular meshwork is the cause of elevated intraocular pressure [2]. Currently, the only proven treatments for POAG patients involve reduction of intraocular pressure by inhibiting aqueous humor production, or bypassing the diseased trabecular meshwork. The mechanisms of increased resistance to aqueous humor outflow are not well-understood [2], but may involve changes in extracellular matrix composition of the trabecular meshwork [3].
Linkage studies have identified a number of POAG loci [4]. So far only three genes have been shown to be associated with POAG [4], but they account for only a small fraction of POAG cases, and none have shed much light on the disease process. Although genome-wide association studies could be a powerful tool to establish more POAG loci, this requires recruitment of a large number of patients. Moreover, causal association between sequence variants and disease can be difficult to establish in human studies. In this study, we have used a canine model to identify a candidate POAG gene, which has the advantage of availability of tissues from normal and affected dogs as well as future gene rescue experiments to investigate the pathogenic mechanisms of the gene variant.
A colony of Beagle dogs established in 1972 [5], which is a well-characterized and naturally occurring animal model of POAG, was used for this study. For POAG-affected dogs in this colony, increases in intraocular pressure begin at 8 to 16 months of age, due to increased resistance to outflow of aqueous humor [6], despite normal appearing open iridocorneal angles. As with POAG in humans, optic nerve cupping, loss of optic nerve axons [7] and vision loss occur in affected Beagles following slowly progressing and sustained elevations of intraocular pressure, if left untreated. Multigenerational breeding experiments have shown that POAG in the Beagle colony is inherited as an autosomal recessive trait [8].
Domestication of the dog from wolves and recent breed creations have resulted in extensive linkage disequilibrium and large haplotype blocks, which makes mapping Mendelian traits possible with far fewer markers and fewer individuals as needed for human studies [9], [10]. The dog genome has been sequenced and microarrays for whole-genome high-density SNP genotyping have been established and used to map traits in dogs [10]–[12]. The aim of this study was to map the disease locus and then to identify candidate disease genes by high-throughput sequencing of the entire disease locus.
To map the POAG locus, we genotyped 19 affected and 10 carrier dogs from the POAG Beagle colony using version 2 of the Affymetrix Canine Genome SNP array. Since the colony has been maintained primarily by affected to affected breeding, with periodic introduction of unrelated normal Beagles (Figure 1), we hypothesized that the disease allele would be contained within an extensive haplotype block homozygous for affected and heterozygous for carrier dogs. Therefore, we identified SNPs that fulfilled the zygosity criterion, defined as being both homozygous for all affected dogs and heterozygous for all carriers. Regions of homozygosity for all affected dogs were common for all chromosomes, as expected for the highly inbred pedigree. However, only Chromosome 20 contained SNPs heterozygous for all carriers (Figure 2A), consisting of 41 consecutive SNPs covering 4.7 Mb. Of those 41 SNPs, 27 consecutive SNPs were also homozygous for all affected dogs, satisfying the zygosity criterion (Figure 2B and Figure S1). Haplotype analysis of the region revealed informative recombination events within the pedigree that defined a 4 Mb locus in which all carriers were heterozygous and all affected dogs homozygous for the affected haplotype (Figure 2C).
In addition to applying the zygosity criterion, two-point and multipoint parametric linkage analyses of the pedigree genotype data were performed. Initial power calculations predicted that with the available pedigree, a single locus could be identified with a LOD score of 2.67. With genome-wide two-point analysis, regions with LOD score >2 were found on chromosomes 5, 15 and 20 (Figure 3A). Follow-up multipoint linkage analysis reduced the LOD score of the chromosome 5 region to below 0.5, excluding this as a candidate locus (Figure 3B). For chromosome 15, multipoint analysis did not reduce the LOD score (Figure 3C). However, haplotype analysis revealed a pattern of inheritance discordant with phenotype (Figure 4), excluding chromosome 15. The distal end of chromosome 20 had a two-point LOD score of 2.42 and a multipoint LOD score of 2.70 (Figure 3D), consistent with initial power calculations. This region identified by linkage analysis coincided with the 4 Mb locus identified by the zygosity criterion. Therefore, the results of genome-wide linkage analysis independently verified that the disease locus in the POAG Beagles maps to the same 4 Mb region of chromosome 20 identified using the zygosity criterion.
Comparison of the 4 Mb POAG locus in dog with the human genome revealed shared synteny within a segment of human chromosome 19, previously identified as a quantitative trait locus for intraocular pressure in humans [13] (Figure 5A). The order and number of genes within the POAG locus on the canine chromosome are highly conserved in the human syntenic region (Figure 5B). Since increased intraocular pressure is the initial manifestation of disease in the POAG Beagles, synteny with the human intraocular pressure locus offers compelling biological support that the 4 Mb region contains the disease-causing genetic variant.
To identify the disease gene, the entire 4 Mb POAG locus in an affected and a carrier dog, as well as a normal dog from the colony (dogs 3, 9 and 11, Figure 1) was isolated by microarray-based sequence capture and then sequenced with the Illumina Genome Analyzer. Alignment of the sequences to the reference canine genome revealed 2,692 sequence variants segregating with disease (homozygous for the affected dog and heterozygous for the carrier dog, with the additional criterion that the normal dog is not homozygous for the same allele as the affected dog). Of the segregating variants, 54 were located within coding regions of canine genes identified by the human protein alignment track of the UCSC genome browser (http://genome.ucsc.edu). Of the 54 variants within coding regions, 8 resulted in non-synonymous amino acid substitutions in 7 genes. Among those 8 variants, based on BLOSUM62 score for amino acid substitution and mammalian conservation score from the vertebrate multiz alignment and conservation track of the UCSC human genome browser, the best candidate variant was at position 56,097,365 of chromosome 20 (canFam build 2) from a G in the reference sequence, to an A in the affected dog. This variant was confirmed by conventional Sanger sequencing of affected, carrier and normal dogs from the POAG colony (Figure 6A). To determine the frequency of the disease allele (56097365 A) in the normal Beagle population, 48 Beagles not affected by glaucoma and not related to the colony were sequenced. Only one of the unaffected dogs was found heterozygous for the disease allele, the rest were homozygous for the normal allele, suggesting a disease allele frequency of ∼1% in Beagles.
The 56097365 G->A variant is within exon 17 of ADAMTS10, a member of the disintegrin and metalloproteinase with thrombospondin motifs family of secreted proteases involved in formation of the extracellular matrix [14]–[16]. The variant results in a Gly->Arg substitution at position 661 within the protein sequence (NCBI accession XP_854320). The glycine at position 661 is completely conserved in 38 vertebrate species (7 representative species shown, Figure 6B). The Gly661Arg substitution was predicted to have a deleterious effect on protein function by the prediction programs SIFT [17] and SNPs3D [18] and occurs within the cysteine-rich domain (Figure 6C), which may be involved in regulation of protease activity [19].
Western blot analysis of protein extracts from tissues dissected from normal dog eyes showed high expression of ADAMTS10 protein in the trabecular meshwork, relative to other eye tissues examined (Figure 7). ADAMTS10 was also expressed in the cornea, and to a much less extent in the iris, ciliary body and optic nerve (Figure 7).
Structural modeling was performed using crystal structure of ADAMTS13 [20] to predict the structures of normal and Gly661Arg ADAMTS10 proteins. In the predicted fold of ADAMTS10, Gly661 is located within a tight turn (Figure 8A), suggesting a glycine may be required at this position for proper folding. Gly661 is predicted to be buried in the structure within the interface between the CA and T1 domains (Figure 8B). Substitution of arginine for glycine at position 661 would be sterically unfavorable, with the longer charged side chain of arginine extending into the T1 domain (Figure 8C), suggesting that the Gly661Arg change would likely disrupt normal ADAMTS10 structure.
To investigate possible effects of the Gly661Arg substitution on ADAMTS10 protein stability, the protein half-lives for normal and mutated ADAMTS10 were determined. Since ADAMTS10 produced by trabecular meshwork cells would be secreted into aqueous humor, half-lives were determined in the presence of aqueous humor. In vitro transcribed normal and mutated ADAMTS10 protein labeled with biotinylated lysine was incubated in aqueous humor for various time periods and the amount of ADAMTS10 remaining at each time point was determined by Western blotting with fluorescently labeled streptavidin. The Gly661Arg mutant appeared to decay more rapidly than did normal ADAMTS10 (Figure 9). The Log2 of the band intensities were plotted vs. time to determine the protein half life, which is equal to the negative inverse of the slope of the best fit line. In each of four independent experiments, mutated ADAMTS10 decayed more rapidly than did normal ADAMTS10 (261+/−29.5 vs. 601+/−219.7 min., mean +/− SD, half-lives for mutated and normal, respectively, significantly different, p<0.05). The slopes of the lines fit to data from all four experiments, combined by normalizing band intensities to the initial time point, were significantly different (p<0.001) and correspond to half lives of 255.8 min. for mutated and 636.9 min. for normal ADAMTS10 (Figure 9C). These results suggest that mutated ADAMTS10 decays more rapidly, with a protein half-life ∼40% that of normal.
By application of the zygosity criterion, linkage and haplotype analyses, we were able to map the Beagle POAG locus to a single 4 Mb region on chromosome 20. This canine POAG locus is syntenic with a region on human chromosome 19 within a quantitative trait locus for regulation of intraocular pressure identified by a genome-wide scan of 486 families [13]. Since ocular hypertension occurs early in the disease process in the Beagles, synteny with the intraocular pressure locus gives biological support to the genetic identification of the POAG locus and suggests that the disease gene directly participates in intraocular pressure regulation. Furthermore, synteny with the human locus suggests that the disease gene found in the Beagles may also be disrupted in human glaucoma patients.
Using affected and unaffected dogs of other breeds to fine map the disease locus has proven to be an effective approach in other canine genetic studies [10]. However, because clinical identification of POAG cases in dogs is rare, this approach would be challenging and further would require the assumption that the POAG locus in Beagles is shared with affected dogs of other breeds. Alternatively, refining the locus by further breeding within the colony to allow for informative recombinations would be time consuming and costly since definitive diagnosis cannot be made until two years of age.
To overcome these limitations, we obtained high quality sequence information for the entire 4 Mb locus by sequence capture and next-generation sequencing. Using this approach, 2,692 single nucleotide variants that segregated with disease were identified, 54 of which were within exons, 8 of which were nonsynonomous. Since POAG in the POAG Beagle colony is autosomal recessive with 100% penetrance, we focused on nonsynonomous changes because these are likely to have strong functional effects. However, synonomous changes in coding regions or variants outside coding regions could have pathogenic effects and cannot be ruled out. In addition, our sequence capture and sequence analysis rely on the quality of the reference canine genome and therefore our approach could miss variants due to errors in the reference genome assembly or annotation.
Among the 8 nonsynonomous variants segregating with disease, the strongest candidate identified was a single base pair change in the affected dogs that results in a non-conservative amino acid substitution in a region of ADAMTS10 that is highly conserved in vertebrate species. In POAG-affected dogs, an arginine is substituted for a glycine at amino acid position 661 which is an invariant amino acid in ADAMTS10 in 38 species, from lamprey to human. Consistent with a highly penetrant rare disease allele, the frequency of the variant in ADAMTS10 estimated from genotyping 48 unrelated normal Beagles was 1%.
ADAMTS10 is a member of a family of secreted metalloproteinases [14], [16]. All ADAMTS family members share a common structural organization including a metalloproteinase domain followed by a disintegrin-like module, a thrombospondin repeat unit, a cysteine-rich domain and a spacer region (see Figure 6C). Diversity within the ADAMTS family largely arises from structural differences in ancillary domains of the carboxy-terminal half of the proteins. The Gly661Arg variant found in this study is within the cysteine-rich domain of ADAMTS10 and is predicted to disrupt protein function by the amino acid substitution prediction programs SIFT [17] and SNPs3D [18]. Consistent with this, our homology modeling of the ADAMTS10 structure suggests that the Gly661 residue is located within a tight turn and is buried within the interface between the cysteine rich and thrombospondin repeat domains. The long polar side chain of arginine substituted at this position is predicted to disrupt the normal protein fold. Consistent with disruption of normal protein folding, we found that the Gly661Arg form of ADAMTS10 is less stable, with a protein half-life ∼40% that of normal ADAMTS10. Although we cannot be certain if the reticulocyte lysate-based in vitro transcription and translation system produced normally folded protein, this system has been used to produce functional secreted proteins such as metalloproteinases [21], neutrophil elastase [22] and myocilin [23]. Any effects of the in vitro system on folding would be experienced by both the normal and mutated proteins in our assays. Our data show that the mutated form of ADAMTS10 has a shortened half-life, consistent with our homology modeling which suggested that the Gly661Arg substitution would disrupt interactions at the interface of two domains.
Clinical evidence for the importance of the cysteine rich domain in ADAMTS function comes from patients with thrombotic thrombocytopenic purpuria (OMIM #274150) who have autoantibodies recognizing the cysteine-rich domain of ADAMTS13, causing reduced proteolytic activity of ADAMTS13 in vivo and in vitro [24], [25]. Structural studies and deletion analysis have established that the cysteine-rich domain plays a vital role in regulation of protease activity or substrate recognition for ADAMTS family proteins [20], [25]. In addition, alignment of the cysteine-rich domains of all 19 human ADAMTS family members and 5 related ADAMTSL proteins by Akiyama et al. [20], revealed that Gly661 of ADAMTS10 is an invariant amino acid. Such stringent evolutionary conservation of this glycine residue, across 38 vertebrate species and within 24 protein superfamily members, supports the hypothesis that the arginine substitution would have a detrimental effect on ADAMTS10 function.
Unlike the three POAG genes identified thus far in humans (MYOC, WDR36 and OPTN) [4], the ADAMTS10 variant identified in this study has obvious functional implications, supporting ADAMTS10 as a strong candidate gene. A role for metalloproteinases in ocular hypertension has long been suggested by numerous in vitro studies [26]. Changes in the amount or composition of extracellular matrix within the trabecular meshwork have been hypothesized to contribute to ocular hypertension by increasing resistance to outflow of aqueous humor through the trabecular meshwork [3]. Although the specific substrate for ADAMTS10 is unknown, other ADAMTS family members are known to participate in collagen processing and proteoglycan degradation. ADAMTS10 is likely to function in some capacity in regulation of extracellular matrix and therefore disruption of its function could lead to POAG by increasing resistance to aqueous humor outflow through the trabecular meshwork. Several ADAMTS family members have been investigated as candidates for regulating outflow resistance, and it has been shown that perfusion of anterior segment organ cultures with ADAMTS4 increases outflow facility [27]. The faster decay of the Gly661Arg ADAMTS10 would likely reduce the amount of ADAMTS10 available, which could possibly result in increased resistance to aqueous humor outflow. Future studies with anterior segment organ cultures perfused with normal and mutated ADAMTS10 could test this hypothesis.
The Beagles of the POAG colony are phenotypically normal with no systemic abnormalities other than POAG in the affected dogs. Our Western blotting results showed that ADAMTS10 is expressed at high levels within the trabecular meshwork as compared to other eye tissues, which would be consistent with an effect of the Gly661Arg ADAMTS10 variant specific to aqueous humor outflow.
Mutations in ADAMTS10 have been identified in human patients with autosomal recessive Weill-Marchesani syndrome (WMS) [28], [29], a rare disease with systemic features including short stature and stubby hands and feet (OMIM #277600). A mutation in type I fibrillin has also been found in autosomal dominant WMS [30], which is clinically indistinguishable from the autosomal recessive form [31], suggesting a functional link between ADAMTS10 and type I fibrillin. WMS belongs to a group of rare connective tissue disorders, including Marfan syndrome (OMIM #154700), for which causative mutations in type I fibrillin, a major constituent of microfibrils [32], have been found. While glaucoma is common in WMS patients [31], the mechanism is not well-studied, due to the extremely small patient population. The prevalence of glaucoma in Marfan syndrome patients is higher than in the general population [33]. Clinically, glaucoma in Marfan syndrome most often presents as POAG, with elevated intraocular pressure and open iridocorneal angles [33]. As type I fibrillin is involved in microfibril formation and function, presentation of POAG in patients with Marfan syndrome caused by type I fibrillin mutations suggests that microfibril defects may be involved in POAG pathogenesis. This notion is supported by another common ocular manifestation in WMS and Marfan syndrome, ectopia lentis (dislocated or malpositioned lens). Consistent with a defect in microfibril structure or function, ectopia lentis is caused by defects in the zonule fibers that hold the lens in place and are composed of fibrillin-containing microfibrils [34]. Recently, mutations were found in other members of the ADAMTS superfamily, ADAMTS17 in autosomal recessive WMS [29] and ADAMTSL4 in isolated ectopia lentis [35], supporting a role for ADAMTS family members in microfibril structure and function.
Ultrastructural studies of human trabecular meshwork have shown changes with age that are more pronounced in POAG patients, including a thickening of sheaths that surround elastin fibers and are composed of extracellular matrix, including fibrillin and fine fibrils, as well as an accumulation of sheath-derived plagues in the aqueous humor outflow pathway [36]. We have previously described similar changes in the trabecular meshwork of POAG-affected Beagles [37], which could be explained by microfibril defects caused by the Gly661Arg variant in the ADAMTS10 gene. Additionally, microfibrils play an important regulatory role in the homeostasis of extracellular matrix by controlling the activation and localization of TGFβ [32], which is elevated in the aqueous humor of glaucomatous eyes [38], [39]. Involvement of microfibril defects in glaucoma is further suggested by recent findings in primary congenital glaucoma of a null mutation in LTBP2, which shares homology with fibrillins and is a structural and functional component of microfibrils [40]. Identification of the Gly661Arg variant of ADAMTS10 in the POAG Beagles in this study provides genetic evidence that microfibril abnormalities may be involved in increased resistance to outflow of aqueous humor through the trabecular meshwork in POAG.
The precise mechanisms of increased resistance to outflow of aqueous humor have remained a long-standing puzzle in glaucoma research. Current treatments for POAG patients involve reduction of intraocular pressure by inhibiting production of aqueous humor, or bypassing the diseased outflow pathway, but do not address the root of the problem. The robust expression of ADAMTS10 in the trabecular meshwork suggests that any defect in ADAMTS10 function caused by the Gly661Arg variant could have particularly pronounced effects on the functioning of the trabecular meshwork, specifically affecting aqueous humor outflow. Identification of ADAMTS10 as a candidate gene in the POAG Beagles suggests that altered processing of extracellular matrix and/or defects in microfibril structure or function may be involved in raising intraocular pressure, offering specific biochemical targets for future research and treatment strategies.
Blood samples from dogs were obtained by licensed veterinarians or veterinary technicians by standard venipuncture, in accordance with the Institutional Animal Care and Use Committees of Vanderbilt University and the University of Florida. A total of 48 canine DNA samples, including 30 dogs from the POAG Beagle colony, 7 unrelated normal Beagles and 11 unrelated mixed-breed dogs, were genotyped using version 2 of the Affymetrix genome-wide SNP genotyping array (http://www.broadinstitute.org/mammals/dog/caninearrayfaq.html). For combined genotype data of all dogs, 40,600 informative SNPs had call rates >90%, and were heterozygous for <50% of dogs. For duplicate samples, 99.6% of SNPs received identical calls. The disease status of dogs was determined by clinical eye exams by veterinary ophthalmologists. The minimal age of the dogs at final diagnosis was 2.2 years.
Initial power calculations were performed using the SIMLINK V 4.12 program (http://csg.sph.umich.edu/boehnke/simlink.php). Two-point and multipoint linkage analyses of the genome-wide SNP data were performed using SuperLink Online [41] assuming an autosomal recessive model with complete penetrance. Dogs 1 through 31, except dog 6, were included in the analysis (Figure 1). SNPs uninformative for Beagles were removed from analysis. Mendelian error checking was performed and inconsistent SNPs removed for all individuals. Minor allele frequencies were calculated using SNP data from 8 unaffected Beagles whose unrelatedness was confirmed using the Graphical Representation of Relationship (GRR) software package [42].
Enrichment for genomic sequence within the 4 Mb locus was carried out using capture microarrays designed and manufactured by Roche NimbleGen, using build 2 of the canine genome. The capture arrays consisted of 385,000 capture probes >60 bp in length, designed to capture all non-repetitive sequence from base position 55,800,000 to 59,850,000 on canine chromosome 20. Sequence capture was carried out on 3 dogs from the POAG Beagle colony (dogs 3, 9, and 11, Figure 1), essentially as described in Albert et al. [43] and Okou et al. [44], with modifications to optimize for the Illumina Genome Analyzer II sequencing platform.
Hybridization of the captured DNA fragments to the flow cell and amplification to form clusters was performed using the Illumina cluster station, following the standard Illumina protocol. The captured DNA fragments were used at a final concentration of 5 pM during hybridization/cluster generation to achieve cluster density of ∼160,000 clusters/tile. Paired end, 38 base pair read sequencing was carried out with the Illumina Genome Analyzer II. Fluorescent images were converted into base pair calls using the Illumina Pipeline software. Paired end alignments to the canine genome build 2 were carried out using Bowtie [45]. For the 3 samples, 53.7% of the reads aligned to the 4.05 Mb target region representing 0.17% of the genome, yielding a 316-fold enrichment of the target sequence. The percentage of the capture region covered >8-fold ranged from 91.3% to 92.1%. The average coverage of genes, represented by the human protein alignment track in the UCSC genome browser (http://genome.ucsc.edu/) ranged from 93.1% to 94.4%. For the 29 SNPs in the capture region represented on the SNP genotyping array, complete concordance in genotype calls was found between the Illumina sequencing and SNP array data. Bases different from the reference canine sequence (variant SNPs) were identified using SAMtools [46] (http://samtools.sourceforge.net). SNPs segregating with disease were defined as being a homozygous variant in the affected dog and heterozygous for the carrier dog, with the additional criterion that the normal dog could not be homozygous for the variant found in the affected dog. Segregation of the 56097365 G->A variant with disease status was confirmed by Sanger sequencing of affected, carrier and normal dogs from the POAG Beagle colony. To determine the disease allele frequency in Beagles, 48 normal Beagles examined by licensed veterinarians and found to be not affected by glaucoma were also sequenced. The normal Beagles were unrelated to the POAG colony and did not share common grandparents.
Postmortem eyes from dogs were obtained by veterinary technicians in accordance with the Institutional Animal Care and Use Committee of Vanderbilt University. Eyes were removed from normal dogs free of eye disease within 30 min after sacrifice. Cornea, trabecular meshwork, iris, ciliary body and optic nerve were isolated by dissection under a stereo microscope. Protein was extracted by homogenization in 150 mM LiCl, 50 mM Tris/pH 7.5, 1 mM dithiothrietol, protease inhibitors and 1% lithium dodecyl sulfate. Protein concentration was determined using a fluorescence-based protein assay (Nano-Orange Protein Assay, Invitrogen). Lysates of HEK293 cells transiently transfected with either empty vector or vector containing an epitope-tagged, full length human ADAMTS10 construct (Origene) were used as controls. For SDS-PAGE under reducing conditions, 10 µg of total protein from eye tissues or 5 µg from cell lysates were loaded into wells of 10% pre-cast polyacrylamide gels (Criterion, Bio-Rad). After SDS-PAGE, proteins were transferred to PVDF membrane (Bio-Rad). Standard Western blotting was performed using 1 µg/ml goat anti-human ADAMTS10 antibody (Santa Cruz) or 3.3 µg/ml mouse anti-human glyceraldehyde-3-phophate dehydrogenase (GAPDH) antibody (clone 6C5, Millipore). Blots were imaged and molecular weights determined using an Odyssey infrared imaging system (Li-Cor Biosciences). A single immunoreactive band for ADAMTS10 ran at an apparent mw of 130 kDa, the same as previously reported for the intact ADAMTS10 zymogen [15].
The homology model of ADAMTS10 was calculated using the program I-TASSER [47] and is based on the structure of ADAMTS13 (PDB entry 3GHM; [20]). Superposition of the calculated model with ADAMTS13 in the program O [48] resulted in a RMS deviation of 1.0 Å for 347 Cα atoms. Figure 8 was made using MOLSCRIPT [49] and RASTER3D [50]. The domain nomenclature and color coding follow those of Akiyama et al. [20].
Canine aqueous humor was obtained from laboratory-quality dogs using protocols approved by the Institutional Animal Care and Use Committee of Vanderbilt University and placed immediately in sealed sterile tubes and stored at −80°C. An expression vector with a T7 promoter upstream of a cDNA insert encoding full-length human ADAMTS10 corresponding to NCBI accession number NM_030957 with a c-terminal Myc-DDK tag was obtained from Origene. The ADAMTS10 insert was verified by Sanger DNA sequencing on both strands. A PCR-based mutagenesis kit (Quick Change II, Stratagene) was used to introduce the G to A mutation found in the POAG-affected Beagles, resulting in a glycine to arginine substitution at amino acid 661 into the expression construct. Mutagenesis was confirmed by Sanger sequencing of the entire construct. A rabbit reticulocyte-based in vitro coupled transcription/translation kit (TNT Quick, Promega) was used to express normal and mutated ADAMTS10 protein from the expression vector constructs following the manufacturer's protocol. Modified lysine-specific tRNA was included in the reaction (Transcend tRNA, Promega) to produce ADAMTS10 protein with biotinylated lysines. To measure protein half-life, samples were made with 4 µl of in vitro reaction mixed with 26 µl aqueous humor in sterile O-ring-sealed tubes and placed in a 37°C water bath. At various times, samples were removed from the water bath and placed in −80°C. The aqueous humor used in the experiments was pooled from 3 individual dogs and included 50 µg/ml cycloheximide (Sigma) to prevent protein synthesis during incubation. Samples were separated by SDS-PAGE using 7.5% pre-cast polyacrylamide gels (Criterion, Bio-Rad). After SDS-PAGE, proteins were transferred to PVDF membrane (Bio-Rad). The membrane was blocked 1 h in PBS/1% casein and then probed with streptavidin conjugated to IRDye 680 (Li-Cor Biosciences). Membranes were imaged and molecular weights and background subtracted band intensities determined using an Odyssey infrared imaging system (Li-Cor Biosciences). A single band at the expected molecular weight of ∼130 kDa was detected, similar to that reported previously [15] and found in eye tissue in this study. Protein decay was assumed to follow the equation:where A(t) is the amount of protein at time t, A(t = 0) is the amount of protein at time t = 0 and h is the half life. The decay equation can be rearranged to:
By plotting the Log2 of the band intensity versus time of incubation, the half-life of the protein was determined as the negative inverse of the slope of the linear fit to the data. Four independent experiments were performed.
|
10.1371/journal.ppat.1002947 | Binding of Glycoprotein Srr1 of Streptococcus agalactiae to Fibrinogen Promotes Attachment to Brain Endothelium and the Development of Meningitis | The serine-rich repeat glycoprotein Srr1 of Streptococcus agalactiae (GBS) is thought to be an important adhesin for the pathogenesis of meningitis. Although expression of Srr1 is associated with increased binding to human brain microvascular endothelial cells (hBMEC), the molecular basis for this interaction is not well defined. We now demonstrate that Srr1 contributes to GBS attachment to hBMEC via the direct interaction of its binding region (BR) with human fibrinogen. When assessed by Far Western blotting, Srr1 was the only protein in GBS extracts that bound fibrinogen. Studies using recombinant Srr1-BR and purified fibrinogen in vitro confirmed a direct protein-protein interaction. Srr1-BR binding was localized to amino acids 283–410 of the fibrinogen Aα chain. Structural predictions indicated that the conformation of Srr1-BR is likely to resemble that of SdrG and other related staphylococcal proteins that bind to fibrinogen through a “dock, lock, and latch” mechanism (DLL). Deletion of the predicted latch domain of Srr1-BR abolished the interaction of the BR with fibrinogen. In addition, a mutant GBS strain lacking the latch domain exhibited reduced binding to hBMEC, and was significantly attenuated in an in vivo model of meningitis. These results indicate that Srr1 can bind fibrinogen directly likely through a DLL mechanism, which has not been described for other streptococcal adhesins. This interaction was important for the pathogenesis of GBS central nervous system invasion and subsequent disease progression.
| Streptococcus agalactiae (Group B streptococcus, GBS) is a leading cause of meningitis in newborns and infants. This life-threatening infection of the brain and surrounding tissues continues to result in a high incidence of morbidity and mortality, despite antibiotic therapy. A key factor in disease production is the ability of this organism to invade the central nervous system, via the bloodstream. We now report that a GBS surface protein called Srr1 binds fibrinogen, a major protein in human blood. This interaction enhances the attachment of GBS to brain vascular endothelial cells, and contributes to the development of meningitis. A mutation in Srr1 that specifically disrupted binding to fibrinogen significantly reduced GBS attachment to brain endothelium, and markedly reduced virulence in an in vivo model of GBS disease. These studies have identified a new mechanism by which Srr1 contributes to GBS invasion of the central nervous system and may provide a basis for novel therapies targeting Srr1 binding.
| The serine-rich repeat (SRR) glycoproteins are a large and diverse family of adhesins found in Gram-positive bacteria [1], [2]. Each SRR protein is encoded within a large locus that also contains genes encoding proteins responsible for glycosylating the SRR protein, as well as an accessory Sec system that is dedicated to the export of the adhesin. The SRR proteins have a highly conserved domain organization, including a long and specialized signal sequence, two extensive serine-rich repeat regions that undergo glycosylation, and a typical LPXTG cell wall anchoring motif [3], [4]. The N-termini also contain a binding region that varies considerably, both in terms of structure and adherence properties (Figure 1).
Among the best-characterized is GspB of Streptococcus gordonii, which binds human platelets through its interaction with sialyl-T antigen on the platelet receptor GPIbα [2], [5]. This appears to be an important event in the pathogenesis of infective endocarditis, since disruption of Siglec-mediated binding results in reduced virulence, as measured by an animal model of endocardial infection [3], [4]. A number of other SRR proteins have been shown to contribute to virulence, including SraP of Staphylococcus aureus, PsrP of Streptococcus pneumoniae, and the two SRR proteins (Srr1 and Srr2) of GBS [6]–[11]. However, the molecular basis for binding by these other adhesins is less defined. Their binding regions have no homology to that of GspB, indicating that they are not Siglec-like adhesins. Although SraP mediates binding to platelets, the receptor for this SRR protein has not been identified [6]. PsrP binds cytokeratin 10 in vitro, which appears to be important for binding to pulmonary epithelial cells and subsequent pneumonia [12].
Expression of Srr1 or Srr2 by GBS has been shown to contribute to virulence in models of meningitis [7], [8]. Srr1 mediates binding to several types of human epithelial cell lines, as well as human brain microvascular endothelial cells (hBMEC) [7], [13]. Binding of these cells appears to be important for both colonization and invasion. In vitro studies have indicated that one ligand for Srr1 is human keratin 4, which may facilitate attachment to cervical, vaginal, and pharyngeal cells [13], [14]. We now report, however, that Srr1 also binds human fibrinogen directly through its interaction with the Aα chain of the heteromultimeric protein. This interaction mediates the binding of GBS both to fibrinogen and to hBMEC, and appears to be important for virulence in the setting of meningitis.
We first measured the adherence of GBS strain COH31 (a serotype III clinical isolate) to a variety of host plasma and matrix proteins. As shown in Fig. 2A, GBS adhered to immobilized human fibrinogen at levels (mean: 16±2.8% of inoculum) that were significantly higher than those seen to with the negative control, casein (<1%). Low levels of binding (<2%) were observed with thrombin, fibronectin, laminin, plasminogen, collagen IV, and fetuin. Binding was significantly inhibited by pretreatment of immobilized fibrinogen with anti-fibrinogen IgG, indicating that the interaction between GBS and fibrinogen was specific (Figure 2B). We also examined eight additional GBS isolates, representing a range of capsular types, all of which were found to bind immobilized fibrinogen. As was seen with the COH31 strain, binding of all GBS strains tested was significantly reduced during treatment with IgG specific for fibrinogen. These data indicate that GBS can adhere specifically to immobilized fibrinogen and adherence to fibrinogen is a general property of GBS.
To better characterize the GBS surface components responsible for fibrinogen interaction, we examined the binding of soluble human fibrinogen to GBS cell wall proteins by Far Western blotting. Although the GBS cell wall extracts contained numerous proteins (Figure 2C, left panel), fibrinogen binding was restricted to a group of high MW bands (300–400 kDa) (middle panel). Probing the membranes with WGA revealed binding of the lectin to one or more proteins of similar size, indicating that they were glycosylated (right panel). Since the serine-rich repeat protein Srr1 of GBS is a high MW glycoprotein, we next assessed the impact of deleting srr1 on WGA and fibrinogen binding. When cell wall extracts of COH31Δsrr1 (PS954) were probed with WGA or fibrinogen, no binding was observed, confirming that the glycoprotein bound by fibrinogen was Srr1.
To examine the impact of Srr1 expression on bacterial binding to fibrinogen, we tested the ability of GBS strains COH31 and NCTC 10/84, and Δsrr1 variants to bind to immobilized fibrinogen. As shown in Figure 3A, deletion of srr1 markedly reduced GBS binding to fibrinogen. Similar results were observed with additional GBS strains H36B and 515 (data not shown). To confirm the role of Srr1 expression in fibrinogen binding by GBS, we next assessed whether binding by COH31 and NCTC 10/84 to fibrinogen was inhibited by rabbit anti-Srr1 IgG (Figure 3B and Figure S1). In control studies, co-incubation of either strain with rabbit IgG had no effect on fibrinogen binding. In contrast, co-incubation of GBS with anti-Srr1 IgG significantly reduced binding to fibrinogen. The level of inhibition was concentration-dependent, with 100 µg/ml of anti-Srr1 IgG being sufficient to reduce WT GBS binding to levels comparable to those seen with GBSΔsrr1. Complementation of the srr1 mutation in trans restored fibrinogen binding by NCTC 10/84 Δsrr1 (Figure S2), thereby demonstrating that the loss of binding observed with srr1 disruption was not due to polar or pleiotropic effects. These results indicate that GBS binding to immobilized fibrinogen is mediated by the surface expressed Srr1 protein.
The attachment of GBS to human brain microvascular endothelial cells (hBMEC) is thought to be important for the invasion of the central nervous system by this organism [15]–[17]. Previous studies indicate that binding of GBS to brain endothelium is mediated by Srr1 [7]. To assess whether fibrinogen contributed to this interaction, we assessed the role of fibrinogen in Srr1-mediated binding of GBS to hBMEC. Fibrinogen was detectable on the surface of washed hBMEC, as measured by immunofluorescence microscopy (Figure 4A). Exposure of the cells to exogenous human fibrinogen (20 µg/ml), markedly increased the amount of the protein on the cell surface, indicating that hBMEC are capable of binding fibrinogen. Strain NCTC10/84 and an isogenic Δsrr1 variant (PS2645) were incubated with hBMEC in tissue culture wells. After 30 min, WT GBS efficiently adhered to these cells, whereas the Δsrr1 mutant was significantly reduced in binding (p<0.01) (Figure 4B). Preincubation of bacteria with purified human fibrinogen (20 µg/ml) enhanced the binding of the WT strain to hBMEC, but had no effect on binding of the Δsrr1 mutant strain.
The ligand binding site of the SRR proteins characterized to date has been localized to the region bridging the two serine-rich repeat domains (Figure 1) [1]–[3], [6], [9]. To confirm that the putative binding region of Srr1 (Srr1-BR) interacts with fibrinogen, we assessed the binding of the purified FLAG tagged binding region (FLAGSrr1-BR) with fibrinogen. In control studies, no significant binding by FLAGSrr1-BR to immobilized casein blocking regent was detected. In contrast, FLAGSrr1-BR showed significant binding to fibrinogen, which increased in direct proportion to the amount of protein applied (Figure 5A). No fibrinogen binding activity was detected by either the N-terminal of Srr1-BR (AA303–479) or C-terminus (AA480–641) alone, indicating that entire region is required. To determine the apparent KD for the binding of FLAGSrr1-BR to fibrinogen, we analyzed data from six independent ELISA-based binding assays, as described previously. The calculated mean KD was 7.51×10−8, which is within the range reported for staphylococcal fibrinogen binding proteins [18]. To validate these findings, we also examined the inhibition of this interaction with either anti-fibrinogen IgG or unlabeled Srr1-BR (Figure 5C and D). When immobilized fibrinogen was pretreated with anti-fibrinogen IgG, the binding of FLAGSrr1-BR to the protein was subsequently reduced (Figure 5C). In addition, when FLAGSrr1-BR was co-incubated with unlabeled (non-tagged) Srr1-BR, subsequent binding was effectively blocked (Figure 5D). These findings indicate that the fibrinogen binding domain of the Srr1 is indeed located in the binding region (AA 303–641).
We next sought to characterize the region within fibrinogen responsible for Srr1-BR binding. Fibrinogen is a complex protein consisting of two subunits, each containing three polypeptide chains (Aα, Bβ and γ). When separated by SDS-PAGE under reducing conditions, fibrinogen appeared as three bands corresponding to the Aα, Bβ, and γ chains (Aα = 63.5 kDa, Bβ = 56 kDa, γ = 47 kDa) having the expected masses (Figure 6B). When transferred to nitrocellulose and probed with purified FLAGSrr1-BR, the Aα chain was readily detected, with low levels of binding seen to the Bβ and γ chains (Figure 6B). We also assessed the binding of Srr1-BR to recombinant forms of each chain, expressed as MalE fusion proteins. In this case, FLAGSrr1-BR was found to bind the MalE:Aα chain, while no binding was seen to the MalE:Bβ and MalE:γ chains (Figure S3).
We next sought to identify the domains within the Aα chain bound by Srr1-BR, by examining the binding of Srr1-BR to a series of recombinant Aα chain truncates (Figure 6A and 6C). Far Western blot analysis showed that binding of FLAGSrr1-BR was localized to subdomains containing residues 283–410, which correspond to the tandem repeat region of the Aα chain (Figure 6C). To confirm that this region was the Srr1-BR binding site, we assessed by ELISA the interaction of FLAGSrr1-BR with the immobilized fibrinogen Aα subdomains (Fig. 6D). As was observed with the Far Western analysis, we found no significant binding of FLAGSrr1-BR to immobilized MalE:Aα198–282 or MalE:Aα(198–282+411–610). However, FLAGSrr1-BR exhibited levels of binding to MalE:Aα283–410 that were comparable to recombinant full length Aα chain (MalE:Aα1–610), indicating that the Srr1-BR binding site is indeed the 13 AA tandem repeat region within the Aα chain of fibrinogen.
Next we examined whether fibrinogen binding by GBS was mediated by the interaction of Srr1-BR with Aα283–410. GBS strains COH31 and NCTC 10/84, and their respective Δsrr1 mutants (PS954 and PS2645) were incubated with either immobilized MalE:Aα283–410 or MalE:Aα198–282 (Fig. 7A and B). The Δsrr1 mutant strains exhibited low levels of binding to both Aα chain truncates. In contrast, WT GBS strains had high levels of binding to MalE:Aα283–410, as compared with MalE:Aα198–282. In addition, we found that GBS binding to immobilized fibrinogen was subsequently reduced during co-incubation with MalE:Aα283–410 (Figure S4), suggesting that Srr1-BR binds fibrinogen specifically within AA 283–410 of the Aα chain, and that this interaction is important for GBS fibrinogen binding.
To gain a better understanding of the structural determinants present within the binding region of Srr1, bioinformatic analysis was performed on the predicted binding region sequence (AA 303–641). Interestingly, PSI-BLAST analysis identified this region to be related to the fibrinogen binding domain of the staphylococcal adhesins SdrG and ClfA (sharing 22% and 23% identity respectively). Structure prediction analysis using PHYRE -[19], Swiss-Model [20], and HHPRED [21] algorithms also identified the binding region of Srr1 as having structural similarity to the fibrinogen-binding region of SdrG (HHPred; 100% probability, e = 4.5e−51) and ClfA (HHPred; 100% probability, e value = 9.2e−51) (Fig. S5).
The binding regions of ClfA and SdrG are composed of two domains (N2 and N3) (Figure 1), each of which adopts an IgG-like fold [22]–[24]. This domain architecture enables fibrinogen binding through a “dock, lock, and latch” mechanism (DLL) [24], in which fibrinogen engages a binding cleft between the N2 and N3 domains. As the ligand “dock”, the flexible C-terminal extension of the N3 domain (the “latch”) changes conformation, so that it “locks” the ligand in place, and forms a β strand complex with the N2 domain [24]. Bacterial adhesins that are structurally related to Clf-Sdr family are able to bind fibrinogen using this mechanism, which appears to represent a general mode of ligand-adhesin binding [24]–[28]. Collectively, our bioinformatic analysis suggests that the binding region of Srr1 structurally resembles the binding region of the Clf-Sdr family proteins (SdrG, ClfA, ClfB) and may have a similar binding mechanism.
Using structure prediction searches (HHPRED) [21], we did not identify a latch-like sequence in C-terminal end of the Srr1-BR. However, a highly homologous TYTFTDYVD-like “latching cleft” sequence between the D1 and E1 strands was identified at AA 412∼420 (TYTWTRYAS) (Figure S5 and Table S3). To investigate whether the C-terminal end of Srr1-BR contained a functional latch-like domain, we generated a variant of Srr1-BR, in which the C-terminal 13 AA had been deleted (FLAGSrr1-BRΔlatch). As shown Figure 8A, this mutation abolished the binding of the Srr1-BR. Moreover, untagged Srr1-BRΔlatch (100 µg/ml) failed to inhibit the binding of FLAGSrr1-BR binding to immobilized fibrinogen (data not shown). The Srr1-BR protein readily bound to hBMEC and this interaction was increased by preincubating hBMEC with fibrinogen (20 µg/ml) (Figure 8B). In contrast, the Srr1-BRΔlatch protein exhibited lower levels of binding to hBMEC compared with the Srr1-BR protein, which were not enhanced by fibrinogen. To exclude the possibility that this deletion had produced changes in the secondary structure of the protein that might account for the reduction in fibrinogen-binding activity, we analyzed Srr1-BR and Srr1-BRΔlatch proteins by circular dichroism (Fig. S6). The two proteins had a similar CD profile, with a maximum at less than 200 and a minimum at 216–218, resembling previously determined CD spectra for ClfA [29]. These results indicate that the Srr1-BR mediates Srr1 binding to fibrinogen, and that the C-terminal end of Srr1-BR contains a latch-like domain.
We next generated an isogenic variant of strain GBS NCTC 10/84 in which the latch-like domain of the Srr1-BR had been deleted. Of note, deletion of this region did not affect surface expression of Srr1 (Figure 8C). We then examined the impact of this mutation on GBS binding to fibrinogen and brain endothelium. As shown in Fig. 8D and E, deletion of the latch region significantly reduced GBS binding to fibrinogen and hBMEC, as compared with the parent strain. These results strongly suggest that GBS binding to fibrinogen is mediated by Srr1-BR via the “dock, lock, and latch” mechanism.
To investigate the role of Srr1–mediated binding to fibrinogen in the pathogenesis of experimental meningitis, we compared the relative virulence of NCTC 10/84 with its isogenic latch-deficient variant. CD-1 mice were infected intravenously with either the WT or the Δlatch mutant strain. Twenty-four hours after challenge, the levels of GBS detected in the blood of each group were essentially identical (Figure 9A). Despite their initial similarities in establishing a high-grade bacteremia in the mouse, infection with the WT GBS strain resulted in significantly higher mortality (p = 0.017, Log Rank test). By 54 h, 50% of mice infected with NCTC10/84 had died. In contrast, all animals infected with GBSΔlatch were alive at 78 h (Figure 9B). At the time of death (or upon euthanasia at 78 h), blood and brain were harvested from each mouse for quantitative bacterial culture. Mice infected with the WT strain exhibited significantly higher final bacterial loads and penetrated into the brain more frequently than the Δlatch mutant (Figure 9C). Histologic examination of brain tissue from mice infected with the Δlatch mutant showed normal brain morphology with no signs of inflammation or injury (Figure 9D), whereas mice infected with WT GBS showed meningeal thickening, tissue destruction and neutrophil infiltration (Figure 9E and 9F).
The SRR proteins of GBS are thought to be important both for colonization of the female genital tract, and for the pathogenesis of invasive diseases, such as sepsis and meningitis. Expression of Srr1 has been shown to enhance the attachment of bacteria to vaginal and cervical epithelial cells in vitro, and to facilitate genital colonization in mice [30]. These interactions may be mediated in part by the binding of Srr1 to cytokeratin 4 on the surface of these epithelial cells. Studies in vitro indicate that the Srr1 interacts with cytokeratin 4 to promote bacterial attachment to the cell surface [14], [30]. However, binding can be blocked by sWGA, suggesting that the glycosylated serine-rich domains may also be involved in the interaction of Srr1 with cytokeratin 4 [14]. Strains expressing Srr1 are also more virulent in animal models of meningitis, as compared with their isogenic, srr1-deleted variants [7], [8]. Expression of Srr1 enhances GBS binding to hBMEC, which is likely to be an essential step for initiating central nervous system invasion and meningitis [7].
Our results now demonstrate that Srr1 promotes the adherence of GBS to human fibrinogen, and that this process is likely to be important for the pathogenesis of meningitis. Binding occurs via the interaction of Srr1-BR with the C-terminus of the fibrinogen Aα chain. This appears to be a specific event, requiring the entire Srr1-BR, and amino acids 283–410 of the Aα chain. Although Srr1 has limited primary sequence similarity to other known fibrinogen binding proteins, our secondary structure analyses indicate that Srr1-BR is likely to have a conformation resembling that of ClfA and possibly other related proteins, such as SdrG of Staphylococcus epidermidis. These and a number of other Gram-positive bacterial adhesins are thought to bind fibrinogen through a “dock, lock, and latch” (DLL) mechanism [24]–[26], as described above. Deletion of the predicted latch-like domain of Srr1 significantly reduced fibrinogen binding by the recombinant protein, as well as by bacteria, suggesting that Srr1 binding occurred by a comparable mechanism. If so, this would be the first example of a streptococcal DLL adhesin. Notwithstanding these similarities, there are some notable differences between Srr1 and its staphylococcal counterparts. For example, while Srr1 binds the Aα chain of fibrinogen, ClfA recognizes the C-terminus of the γ chain, and SdrG binds the N-terminus of the β chain [24], [25], [27]. Although both Srr1 and ClfB bind the C-terminus of the Aα chain, their binding sites on fibrinogen appear to differ [27], [28], [31]. A recombinant peptide representing the Aα chain binding site for ClfB (AA283–347) did not inhibit Srr1-BR binding to fibrinogen (Figure S7). Conversely, a peptide containing Aα chain residues 348–410 effectively blocked Srr1-BR binding, but no effect on ClfB binding to fibrinogen. These findings suggest that, while the binding of Srr1 to the Aα chain has some features in common with ClfB, the interactions of these adhesins with fibrinogen must also differ significantly. Further understanding of the precise basis for Srr1 binding to fibrinogen, and whether it occurs via a DLL mechanism, will require solution of its crystal structure.
Srr1 binding to fibrinogen was also important for the attachment of GBS to hBMEC in vitro. Binding of GBS to brain endothelium was reduced by deletion of the putative latch domain of Srr1, and was significantly enhanced by adding human fibrinogen, at concentrations (20 µg/ml) well within those found in whole blood (2–4 mg/ml) [32]. These findings indicate that the Srr1-fibrinogen binding is a relevant process for CNS invasion, and indeed we found that in mice with experimental meningitis, the latch deletion was also associated with significantly reduced levels of bacteria, mortality, and inflammation within the CNS. Of note, levels of the bacteria within the bloodstream were not altered by the above mutation, further indicating that the virulence properties associated with Srr1 and fibrinogen binding are specific to CNS infection.
FbsA and FbsB are two additional fibrinogen binding proteins of GBS that have been characterized [33], [34]. These proteins appear to be structurally unrelated to Srr1 or other known fibrinogen binding proteins. FbsA and FbsB can bind fibrinogen directly in vitro, although their binding sites on fibrinogen have not been identified. FbsA can also enhance the attachment of GBS to hBMEC [35]. However, FbsA alone is not sufficient for cell invasion, but appears to require FbsB for this process [36]. The contribution of FbsA and FbsB, and their interactions with fibrinogen to virulence is not well-defined. Neither protein has been examined for its role in the pathogenesis of meningitis. Deletion of fbsA was associated with decreased virulence in an animal model of septic arthritis and septicemia [37]. However, neither active nor passive immunization with FbsA or FbsA-specific antibodies resulted in protection against subsequent infection [37], suggesting that the virulence properties of FbsA may be unrelated to fibrinogen binding. Two other GBS proteins (the fibronectin binding protein Fib and a predicted ABC transport protein SAG0242) have been shown to bind fibrinogen, but neither the mechanisms for protein binding, nor the biologic importance of these interactions, have been addressed [33].
In summary, our results show that Srr1 mediates the binding of GBS to fibrinogen, and that this interaction is likely to occur via a DLL-like mechanism, involving the C-terminus of the fibrinogen Aα chain. It is the first streptococcal adhesin for which this type of binding has been identified, indicating that DLL binding may be a generalized mechanism for attachment by Gram-positive organisms. In addition, Srr1-fibrinogen binding appears to be important for the adherence to brain endothelium and the development of meningitis Given that Srr1 or its homolog Srr2 appear to be expressed by most clinical isolates of GBS, this interaction may prove to be a promising candidate for novel therapies targeting bacterial virulence.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee of San Diego State University (Animal Welfare Assurance Number: A3728-01). All efforts were made to minimize suffering of animals employed in this study.
Purified human fibrinogen was obtained from Haematologic Technologies. Rabbit anti-fibrinogen IgG was purchased from Aniara. Rabbit anti-Srr1 IgG was generated using purified Srr1-BR protein (NeoPeptide).
The bacteria and plasmids used in this study are listed in Table S1 and S2. S. agalatiae strains were grown in Todd-Hewitt broth (Difco) supplemented with 0.5% yeast extract (THY). All mutant strains grow comparably well in vitro (data not shown). Escherichia coli strains DH5α, BL21 and BL21(DE3) were grown at 37°C under aeration in Luria broth (LB; Difco). Appropriate concentrations of antibiotics were added to the media, as required.
Genomic DNA was isolated from GBS NCTC 10/84, using Wizard Genomic DNA purification kits (Promega), according to the manufacturer's instructions. PCR products were purified, digested, and ligated into pET28FLAG to express FLAG-tagged versions of Srr1-BR (amino acids [AA] 303–641), the amino terminus of Srr1-BR (AA 303–479), the carboxy terminus of Srr1-BR (AA480–641) or the latch deletion of Srr1-BR (AA 303–628). Untagged Srr1-BR and Srr1-BRΔlatch were cloned into pET22b(+) (Novagen). The plasmids were then introduced to E. coli BL21(DE3) for over-expression. Proteins were purified by either Ni-NTA (Promega) or anti-FLAG M2 agarose affinity chromatography (Sigma-Aldrich), according to the manufacturers' instructions.
cDNAs encoding the Aα-, Bβ- and γ-chains of human fibrinogen were generously provided by Professor Susan Lord (University of North Carolina at Chapel Hill) [38]–[40]. The full length and truncated forms of chains were amplified and cloned into pMAL-C2X (New England Laboratory) to express MalE-tagged versions of the chains. Plasmids were then introduced to E. coli BL21 by transformation. All recombinant proteins were purified by affinity chromatography with amylose resin, according to the manufacturer's instructions (New England Biolabs).
Purified human fibrinogen and recombinant fibrinogen chains were separated by electrophoresis through 4–12% NuPAGE Tris-Acetate gels (Invitrogen) and transferred onto nitrocellulose membranes. The membranes were treated with casein-based blocking solution (Western Blocking Reagent; Roche) at room temperature, and then incubated for 1 h with FLAG-tagged Srr1-BR (0.5 µM) suspended in PBS-0.05% Tween 20 (PBS-T). The membranes were then washed three times for 15 min in PBS-T, and bound proteins were detected with mouse anti-FLAG antibody (Sigma-Aldrich).
Purified fibrinogen (0.1 µM) was immobilized in 96-well microtiter dishes by overnight incubation at 4°C. The wells were washed twice with PBS and blocked with 300 µl of a casein-based blocking solution for 1 h at room temperature [41], [42]. The plates were washed three times with PBS-T, and FLAGSrr1-BR, FLAGSrr1-BR-N, FLAGSrr1-BR-C or FLAGSrr1-BRΔlatch in PBS-T was added over a range of concentrations. The plates were then incubated for 1 h at 37°C. Unbound protein was removed by washing with PBS-T, and the plates were incubated with mouse anti-FLAG antibodies diluted 1∶4000 in PBS-T for 1 h at 37°C. Wells were washed and incubated with HRP-conjugated rabbit anti-mouse IgG diluted 1∶5000 in PBS-T for 1 h at 37°C. The dissociation constant KD for Srr1 binding was calculated using Prism software v. 4.0 (GraphPad).
For inhibition assays, the wells containing immobilized with fibrinogen (0.1 µM) were pretreated with rabbit anti-fibrinogen or rabbit IgG for 30 min, followed by washing to remove unbound antibody prior to the addition of FLAGSrr1-BR. In addition, FLAGSrr1-BR was coincubated with anti-Srr1 IgG or purified untagged Srr1-BR proteins on the wells immobilized with fibrinogen. After washing out unbound proteins, bound FLAGSrr1-BR was then assessed as described above.
hBMEC were fixed with 4% paraformaldehyde and fibrinogen was stained with rabbit anti-fibrinogen IgG (1∶1000) and Alexa Fluor 488 conjugated goat anti-rabbit IgG (Invitrogen). Coverslips were mounted on glass slides using Vectashield (Vector labs) and visualized with a confocal laser scanning microscope (Leica Microsystems).
Overnight cultures of GBS were harvested by centrifugation and adjusted to a concentration of 106 CFU/ml in PBS. Purified fibrinogen (0.1 µM) was immobilized in 96-well microtiter plates as described above, and then incubated with 100 µl of GBS suspension for 30 min at 37°C. The wells were then washed to remove unbound bacteria, and then treated with 100 µl of trypsin (2.5 mg/ml) for 10 min at 37°C to release the attached bacteria. The number of bound bacteria was determined by plating serial dilutions of the recovered bacteria onto THB agar plates as previously described [41].
The human brain microvascular endothelial cell line (hBMEC) was developed and kindly provided by Kwang Sik Kim (Johns Hopkins University) [43], [44] and cultured as previously described [45]. Bacterial adherence assays were performed as described [46]. In brief, bacteria were grown to mid-log phase and then added to confluent hBMEC monolayers at a multiplicity of infection (MOI) of 0.1. After 30 min incubation, monolayers were washed 6 times with PBS to remove non-adherent bacteria, lysed and plated on THB agar to enumerate the bacteria. Bacterial adherence was calculated as (recovered CFU/initial inoculum CFU)×100%. In indicated experiments exogenous fibrinogen (20 µg/ml) was added directly to bacteria and incubated 1.5 hours with rotation at 37°C prior to addition to hBMEC monolayers.
GBS cell wall extracts were prepared by treatment with spheroplasting buffer (500 units/ml mutanolysin, 20 mM Tris, 10 mM MgCl2·6H2O, and 0.5 M raphinose), as described previously [47], [48]. Proteins were separated by SDS-PAGE with 3–8% Tris-Acetate gels (Invitrogen) under reducing conditions and then were transferred to nitrocellulose membranes. After blocking with casein based blocking reagent (Roche), the membranes incubated with either 1) anti-Srr1-BR IgG (1∶3000) following by incubation with anti-rabbit IgG (1∶10,000); or 2) biotin conjugated wheat germ agglutinin (WGA; Vector Labs) (0.2 µg/ml) followed by incubation with HRP conjugated streptavidin (0.2 µg/ml).
A murine model of hematogenous GBS meningitis has been described previously [46]. Outbred 6- to 8-week old male CD-1 mice (Charles River Laboratories; 10 mice per group) were injected via the tail vein with 5×107 CFU WT GBS (NCTC 10/84) or GBSΔlatch mutant. At 24 h post GBS injection, blood was collected via tail vein (20 µl) and plated on THB agar to determine the bacterial load in the bloodstream. Mouse survival was accessed over time. At the time of death, or at 78 h post infection, blood and brain tissue were collected aseptically from mice after euthanasia. Bacterial counts were in blood and tissue homogenates were determined by plating serial 10-fold dilutions on THB agar. Brain sections were also embedded in paraffin and stained with hematoxylin and eosin (H&E).
Amino acid similarity was compared using PSI-BLAST and secondary structure was determined by the prediction servers (PHYRE and HHPRED) [19], [49], [50].
Data were expressed as means ± standard deviations and were compared for statistical significance by the unpaired t test.
|
10.1371/journal.pcbi.1005769 | A unifying Bayesian account of contextual effects in value-based choice | Empirical evidence suggests the incentive value of an option is affected by other options available during choice and by options presented in the past. These contextual effects are hard to reconcile with classical theories and have inspired accounts where contextual influences play a crucial role. However, each account only addresses one or the other of the empirical findings and a unifying perspective has been elusive. Here, we offer a unifying theory of context effects on incentive value attribution and choice based on normative Bayesian principles. This formulation assumes that incentive value corresponds to a precision-weighted prediction error, where predictions are based upon expectations about reward. We show that this scheme explains a wide range of contextual effects, such as those elicited by other options available during choice (or within-choice context effects). These include both conditions in which choice requires an integration of multiple attributes and conditions where a multi-attribute integration is not necessary. Moreover, the same scheme explains context effects elicited by options presented in the past or between-choice context effects. Our formulation encompasses a wide range of contextual influences (comprising both within- and between-choice effects) by calling on Bayesian principles, without invoking ad-hoc assumptions. This helps clarify the contextual nature of incentive value and choice behaviour and may offer insights into psychopathologies characterized by dysfunctional decision-making, such as addiction and pathological gambling.
| Research has shown that decision-making is dramatically influenced by context. Two types of influence have been identified, one dependent on options presented in the past (between-choice effects) and the other dependent on options currently available (within-choice effects). Whether these two types of effects arise from similar mechanisms remain unclear. Here we offer a theory based on Bayesian inference which provides a unifying explanation of both between and within-choice context effect. The core idea of the theory is that the value of an option corresponds to a precision-weighted prediction error, where predictions are based upon expectations about reward. An important feature of the theory is that it is based on minimal assumptions derived from Bayesian principles. This helps clarify the contextual nature of incentive value and choice behaviour and may offer insights into psychopathologies characterized by dysfunctional decision-making, such as addiction and pathological gambling.
| Standard theories of decision-making assume that the incentive value of an option should be independent of options presented in the past and options available during choice [1–4]. These theories are fundamentally challenged by empirical evidence showing that expectations (derived from past experience) about upcoming options change value attribution and choice behaviour [5–14]. For example, in a series of recent experiments from our lab [8–10], participants made choices in blocks (i.e. contexts) associated with one of two distinct, but partially overlapping, reward distributions. Participants’ choices were consistent with attributing a larger incentive value to rewards (common to both contexts) in blocks associated with low compared to high average reward. In other words, the incentive value of a reward increased when the average was lower. In addition to the average reward of a context, evidence from a similar task indicated that reward variance within a given context also exerts an influence on incentive value [11]. These findings highlight contextual effects exerted by expectations about options (induced, for example, by options available during previous choices); namely, between-choice contextual effects.
In addition, the empirical literature has highlighted contextual influences elicited by options available during choice; namely, within-choice context effects [6, 15–20]. Standard theories of decision-making assume that the incentive value of an option should be independent of other options available during choice [1–4]. This implies that the choice proportion between two options, comprising a more valuable and a less valuable option, should be unaffected by the introduction of a third [2]. However, a recent study [6] has shown that this choice proportion follows a U-shape function, which diminishes as the value of a third option approaches the value of the target options–and starts increasing thereafter (Fig 1A). This is hard to reconcile with standard theories and represents a form of within-choice context effect, whereby the value of an option is affected by other options available during choice.
In this task, it is unnecessary to compare options across different attributes (single-attribute decisions; [6]). However, other forms of within-choice context effect have been observed when options are defined by the same set of attributes that have to be traded of against each other (multiattribute decisions; [15–20]. For example, consider a binary choice between a high-quality and expensive car A versus a low-quality and cheap car B (Fig 1B). Imagine the values of the attributes are such that an agent is indifferent about the two options (i.e., the higher price of car A is exactly compensated by its quality), resulting in the same probability of choosing options A and B. What happens if a third option is also available? Standard models (based on the assumption that values are independent of other options) predict that the choice probability difference will remain zero, independent of a third option. However, empirical data highlight a so-called similarity effect [20–23], whereby preference for an option over a second option–which is equally preferable during binary decisions–increases if a third option is available that is similar to the second option (Fig 1B). In our example, the choice probability difference between car A and car B will be positive when a third low-quality and cheap (similar to car B) car C is also available. A form of influence called the attraction effect [15, 24–26] has also been found with the availability of a third option that is characterized by a low score for one attribute and an intermediate score for the other (Fig 1B). The presence of such a third option favours the option with a high score for the attribute for which the third option has an intermediate score. In our example, the choice probability difference between car A and car B will be positive when a third medium-quality and expensive car D is also available. Finally, empirical data are consistent with a so-called compromise effect [17, 25, 27]. This applies when the choice set includes two options scoring high in one attribute and low in another plus a third option characterized by intermediate scores for both attributes. While the three options are equally preferred (i.e., are chosen an equal amount of times) if presented in pairs during binary choices, when they are all available, a preference for the option characterized by intermediate scores is observed (Fig 1B). For instance, although during binary choices an average-price and average-quality car E is not preferred over car A or over car B, car E will be favoured when presented together with both car A and car B.
Several explanations have been proposed to account for contextual effects on incentive value and choice, with most models focusing on within-choice context effects during multiattribute decisions [16–20, 27, 28]. Other theories have been proposed to explain between-choice context effects [29–31], and disregard within-choice effects. We are aware of a single attempt to encompass both between-choice and within-choice effects, though restricted to non-multiattribute decisions for the latter type of effects [6]. Whether models developed to explain a certain class of context effects generalise to other effects remains unclear–and a unifying account encompassing all known context effects is lacking. Developing a parsimonious account would represent an important theoretical advance, as it would explain diverse empirical phenomena with the same underlying principles.
The goal of the present paper is to describe a unifying theory, referred to as Bayesian model of context sensitive value (BCV) that explains between-context and within-context effects, in single and multiattribute decisions. This theory represents a generalization of a recent model developed to explain between-choice contextual effects [11]. The key idea is that agents build a generative model of reward within a context and, every time a new reward or option is presented, use Bayesian inference to invert this model to form a posterior belief about the underlying reward distribution. Incentive value is computed during this belief update and corresponds to a precision-weighted reward prediction error. The advantage of this theory relies on its grounding upon simple normative principles of Bayesian statistics. In addition, the model can explain between-choice context effects [8–11] and makes specific predictions that have been confirmed empirically.
In brief, BCV proposes that the incentive value of a stimulus (or option) corresponds to the change in reward expected (in any given context) when the stimulus is presented. This makes precise predictions about choice under ideal (Bayesian) observer assumptions (with a minimal number of free parameters). Crucially, predictions include specific forms of context effects, and raise a question of whether these predicted effects are consistent with empirical findings.
In this paper, we applied BCV to multi-alternative choice (considering both single and multiattribute decisions) and ask whether the model predicts the context-effects found empirically. We first present a theoretical extension of BCV applicable to decisions in which multiple options are available and can be characterized by multiple attributes. We show that predictions derived from the model are remarkably similar to empirical findings on within-choice contextual effects, both during non-multiattribute and multiattribute decisions. We next review BCV in relation to between-choice context effects and describe how the model can also explain these empirical findings. On this basis, we offer the model as a principled description of between and within-choice context effects.
The idea behind BCV is to establish a link between theories of value and normative accounts of brain functioning based on Bayesian statistics [32–37]. The Bayesian brain framework rests on the idea that an agent builds a model of the processes generating sensory cues. The generative model comprises a set of random variables (i.e., hidden states or causes of sensory outcomes) and their causal links (i.e., probabilistic contingencies). The variables can be separated into hidden and observable variables, the former representing the latent causes of observations, and the latter representing sensory evidence or cues. Sensory evidence conveyed by observable variables is combined with prior beliefs about hidden causes to produce a posterior belief about the causes of observations. The application of this logic has proved effective in explaining several empirical phenomena in perception [32–37]. For instance, psychophysical data indicate that human perception depends on integrating different perceptual modalities (e.g., visual and tactile) in a manner consistent with Bayesian principles [38], where evidence is weighted by the precision of sensory information. Furthermore, process theories that mediate Bayesian inference (e.g., predictive coding) have a large explanatory scope in terms of neuroanatomy and physiology [39].
Inspired by a recent framework that conceptualises planning and choice as active inference [40–45], our core proposal is that Bayesian inference drives the attribution of incentive value to reward, and this in turn determines choice.
In a previous work, we have developed a version of BCV applicable to conditions where past options elicit context effects by shaping expectancies before a reward is presented ([11]; see below). However, our previous formulation did not consider conditions where multiple options (potentially characterized by multiple attributes) are available. Here we generalize BCV to encompass conditions in which multiple options are available and options can be characterized by multiple attributes. We define a multi-attribute option un (e.g., car A or car B) as a contract that yields reward amount Ri,n relative to each attribute i (e.g., price or quality):
un={Ri,n}i=1,…,InwithRi,n∈R
(1)
An option set u is the set of options currently available:
Setu={un}n=1,…,Nwithunanoption∀n
(2)
The expected value (EV) of an option un corresponds to:
Run=∑iRi,n
(3)
For example, the total reward for car A is equal to the reward associated with price plus the reward associated with quality. BCV assumes that an agent builds a generative model of the reward amounts Ri,n (Fig 2). Specifically, an agent believes that, for each attribute i, reward amounts Ri,n across options are sampled from the same population. To distinguish among attributes, we assume that an agent believes that an independent population of reward amounts is associated to each attribute. For example, if two attributes characterize options, two independent populations of reward amounts are considered by the agent (Fig 2).
Formally, for each attribute i, the average of the population of the reward amounts Ri,n is represented by a random variable Ci, which is assumed to be sampled from a Gaussian distribution with prior mean μCi and uncertainty (variance) σCi2:
Ci∼N(μCi,σCi2)
(4)
The agent assumes that μCi and σCi2 are known but that Ci is not directly observable and therefore needs to be inferred from observing the different instances of reward amounts Ri,n of options for the attribute i. This is realized in the generative model by treating Ci as a hidden cause of Gaussian variables Ri,n with mean Ci and uncertainty σRi2:
Ri,n∼N(Ci,σRi2)
(5)
On the basis of the generative model, for each attribute i, the agent can estimate C^i=N(μ^Ci|Ri,σ^Ci|Ri2), namely the posterior belief about the variable Ci (i.e., the average reward amount relative to the attribute i; the hat symbol indicates estimates of unknown quantities), given the observation of all reward amounts of all options available for the attribute i, represented by the set Ri. In other words, an agent assumes that there is an average reward for each attribute which is unknown but can be estimated based on the reward amounts.
According to Bayes’ rule, the posterior belief of Ci can be calculated by considering the associated Ri,n sequentially in any order. We propose such sequential belief updating for BCV, even if options (and the associated reward amounts) are presented simultaneously, and we assume that the order of options considered is random (with potentially different orders for different attributes). For example, when three options characterized by two attributes are available (represented by R1,1, R1,2 and R1,3 for attribute one, and R2,1, R2,2 and R2,3 for attribute two), inference can involve computing, in order, P(C1|R1,1), P(C1|R1,1, R1,3) and P(C1|R1,1, R1,2 R1,3) for attribute one, and P(C2|R2,3), P(C2|R2,1, R2,3) and P(C2|R2,1, R2,2 R2,3) for attribute two. In the example above, an agent may consider first car A and next car B when estimating the average reward for price, and first car B and next car A when estimating the average reward for quality.
The rationale behind sequential belief updating is that the brain is equipped with a limited computational capacity, which precludes the instantaneous (and parallel) evidence accumulation, and hence requires the processing of one option after another. A similar evidence accumulation process is implicit in some theories of perceptual and value-based decision-making (e.g., [16, 46, 47]). Below, we will show that this evidence accumulation, in the form of sequential Bayesian belief updating, endows agents with the right sort of sensitivity to context. Formally, if Ri,n is the reward amount considered first during belief updating, in relation to attribute i, the posterior mean μ^Ci|Ri,n is [48]:
μ^Ci|Ri,n=μCi+σCi2σCi2+σRi2(Ri,n−μCi)
(6)
The posterior uncertainty σ^Ci|Ri,n2 is:
σ^Ci|Ri,n2=σCi2−σCi2σCi2+σRi2σCi2
(7)
The crucial proposal we advance is that the incentive value Vi(Ri,n)–attributed to a reward amount Ri,n in relation to the attribute i and associated with option un–is central to belief updating (see Eq 6) and corresponds to a precision-weighted prediction error [49]; namely, to the difference between Ri,n and the prior mean μCi, multiplied by a gain term which depends on the uncertainty of that attribute σRi2 and the prior uncertainty σCi2:
Vi(Ri,n)=σCi2σCi2+σRi2(Ri,n−μCi)⇒μ^Ci|Ri,n=μCi+Vi(Ri,n)
(8)
Within BCV, incentive value imbues reward and associated options with behavioural relevance, by favouring either approach to (for positive incentive values) or avoidance of (for negative incentive values) these reward amounts and options.
This implies two fundamental forms of contextual normalization. First, a subtractive normalization is exerted when μCi is different from zero. For example, if we assign positive and negative numbers to rewards (i.e., Ri,n > 0) and punishments (i.e., Ri,n < 0) respectively, their corresponding incentive values will change in sign, depending on whether punishment (i.e., μCi < 0) or reward (i.e., μCi > 0) is expected a priori. Small rewards may appear as losses in contexts where large rewards are expected. Second, a divisive normalization depends on considering the gain term σCi2σCi2+σRi2. This implies that the positive and negative value of profits (i.e., Ri,n > μCi) and losses (i.e., Ri,n < μCi) are magnified by a large gain term, when we have precise beliefs about the average reward of the population.
Sequential Bayesian belief updating means that inference proceeds by considering one reward amount at a time. If Ri,n is considered at step t+1 and Ri,t is a set containing all reward amounts already seen up until step t for attribute i, then a posterior mean μCi|Ri,t,Ri,n is obtained at step t+1 equivalent to (Bishop, 2006):
μ^Ci|Ri,t,Ri,n=μ^Ci|Ri,t+σ^Ci|Ri,t2σ^Ci|Ri,t2+σRi2(Ri,n−μ^Ci|Ri,t)
(9)
Implying a value for the reward amount Ri,n:
Vi(Ri,n)=σ^Ci|Ri,t2σ^Ci|Ri,t2+σRi2(Ri,n−μ^Ci|Ri,t)
(10)
For each attribute i, incentive values are accumulated in memory until inference is completed (i.e., all reward amounts have been considered). We can assume that inference proceeds in sequence or in parallel across attributes; however, this has no impact on incentive values, as the agent believes that attributes are associated with independent reward populations (formally: P(C1,C2,…,CI|R) = P(C1|R1), P(C2|R2),…,P(CI|RI)).
When all attributes for an option un have been considered, we assume that the incentive value of the option corresponds to the sum of the incentive values of associated reward amounts:
V(un)=∑i=1Vi(Ri,n)
(11)
Inference proceeds until, for all attributes i, the posterior expectation about rewards μ^Ci|Ri is evaluated and, at this point, a choice is realized following a softmax rule based on the incentive values of the available options [2].
In summary, BCV is based on the following assumptions:
Below these assumptions are discussed in detail. Assumptions (i), (iii), and (iv) are implicit in adopting a Bayesian scheme. Assumption (vii) is based on a standard approach in which incentive values are summed and a softmax choice rule is adopted. Assumption (ii) captures the notion of multiple attributes, in other words it enables an agent to link rewards to their attributes. Assumption (vi) (sequential belief updating or evidence accumulation) reflects the real world constraint that people have to evaluate available options and rewards one by one. In other words, agents cannot magically and instantaneously assimilate all the options on offer–they have to accumulate evidence for the underlying payoffs by evaluating each in turn. This notion plays a central role since we will see that context effects emerge because a reward is contextualized by previous rewards encountered during inference. This underlies assumption (v) that associates incentive value with a precision-weighted prediction error–a central construct in Bayesian inference. Heuristically, this scheme implies that an option is more likely to be selected if it increases expectations of reward, and will be avoided if it decreases expectations. In other words, an option is more likely to be selected if it suggests the situation is better than indicated by options considered previously during belief updating.
Note that a Bayesian perspective may suggest that incentive value corresponds to a posterior belief–rather than a precision-weighted prediction error. As an example, this would imply that the value of the same dish will be perceived as ‘higher’ in a ‘better’ restaurant. However, empirical data are consistent with the opposite notion that (adopting the same example) the value of the same dish is perceived as lower in a better restaurant [6–14]. This evidence motivated our proposal that incentive value corresponds to a precision-weighted prediction error, and not to a posterior belief.
In sum, BCV provides a principled explanation for how Bayesian inference, assigning a key role to prior expectation and uncertainty, might underlie value computation and choice. The key role of uncertainty is reflected in the precision-weighting of prediction errors. The hypothesis we entertain here is that the mechanisms postulated by BCV may be general and explain multiple forms of context effects. We have previously applied BCV to explain between-choice context effects; namely, those elicited by options presented in the past [11]. Here, we explore the possibility of applying the same model to within-choice effects, which arise when multiple options are available. In what follows, we will consider single and multiattribute choices under this Bayesian formalism.
Here, we apply BCV to explain within-choice contextual influences during non-multiattribute decisions. These comprise choices in which trading-off different attribute is not required, as for instance when options are defined by a single attribute. Consider first how different prior expectations μC (i.e., the prior expectation over the average reward of the attribute) and reward uncertainty σR2 affect the choice between two options characterized by a single attribute (Fig 3). We can examine the predicted proportion of choosing a better option (associated with high reward RH) compared to a worse option (associated with low reward RL), as a function of prior expectation μC and reward uncertainty σR2. Classical theories predict a flat function because they do not model an influence of the prior mean μC and reward uncertainty σR2 [1–4]. In contrast, BCV predicts bell-shape functions over prior expectations, that peak at the prior mean of μC = (RH–RL)/2 (Fig 3). In this setting, the reward uncertainty σR2 determines the width of the function (larger uncertainties produce narrower functions). Below, we analyse conditions where more than two options are available–and within-choice context effects come into play.
Classical decision-making models predict that, during choice, the choice ratio between two options should not be affected by the reward associated with a third option [1–4]. However, a recent study has challenged this hypothesis, highlighting within-choice context effects [6]. Adopting a choice task in which three options were available during choice, this study showed that the choice proportion between a more valuable and a less valuable target option diminished as a third option value increased towards the value of the target options (Fig 1A). After this point, the choice proportion started increasing (Fig 1A).
Here, we examine the implications of applying BCV in this scenario. Fig 4 illustrates the predictions of BCV of the ratio of choices of the two target options (a better target option RH and a worse target option RL) as a function of the reward of a third option R3 and as a function of the agent’s prior belief about the average option reward μC and about the reward uncertainty σR2. This figure shows that all these variables exert an influence. First, for certain values of reward uncertainty σR2 and prior mean μC, the reward of a third option R3 influences the choice proportion between the two target options according to a U-shape function, in a way that is consistent with empirical findings (Fig 1A). Second, the impact exerted by the reward of a third option R3 decreases as the reward uncertainty σR2 increases. In other words, within-choice context effects emerge only with small reward uncertainty σR2. This can be explained by the fact that a small σR2 magnifies reward prediction error (RPE), enabling contextual effects to emerge. Third, when the reward uncertainty σR2 is sufficiently small, the prior mean μC comes into play. Overall, a larger prior mean μC increases the choice proportion between the two target options (independently of the reward of a third option R3). Furthermore, the prior mean μC exerts a modulatory influence on the effect of the reward of a third option R3, as the effect exerted by R3 is enhanced with a larger prior mean μC. Note that context effects exerted by R3 are obtained with μC = 0, which can be considered a default value for this parameter.
Collectively, these simulations provide proof of principle that BCV can explain within-choice contextual effects in single-attribute decisions that are remarkably similar to those seen in empirical studies [6]. In what follows, we now extend the explanatory scope of BCV to multiattribute problems.
Empirical studies of multi-attribute decisions have highlighted three forms of effects, including the similarity [20–23], attraction [15, 24–27], and compromise effect [17, 25, 27]. Here, we apply BCV to multi-attribute decisions and ask whether the predictions that emerge from the model reproduce the context effects found empirically. To this aim, we consider two options (e.g., the two cars A and B described above) defined by two attributes (e.g., price p and quality q). Considering the reward amounts of car A, we assign Rp,A = 1 to price (low scores indicate high price) and Rq,A = 10 to quality. Conversely, when considering the reward amounts of car B, we assign Rp,B = 10 to price and Rq,B = 1 to quality. We now consider the choice probability difference between option A and option B as a function of the reward amounts Rp,K and Rq,K of a third option K.
Empirical evidence is hard to reconcile with standard models of choice, which predict that the choice probability difference between option A and option B should not depend on the value of a third option K. Fig 5A summarises the empirical findings by plotting the probability of choosing A minus the probability of choosing B as a function of the attributes of a third option K. This graph shows conditions in which the choice probability difference is bigger or smaller than zero, illustrating both a similarity and an attraction effect. Specifically, a similarity effect favours option A when option K is good in price and bad in quality (top-left of the graph), and favours option B when option K is bad in price and good in quality (bottom-right of the graph). An attraction effect favours option A when option K is bad in price and has an average quality (bottom-middle of the graph), and favours option B when option K has an average price and is bad in quality (middle-left of the graph).
We can now apply BCV to model choices in this scenario by analysing the influence on the choice probability (difference between option A and option B) of the prior mean μC (we use an equal prior mean for both attributes price and quality; formally: μCp = μCq), the reward uncertainty σR2 (we use an equal reward uncertainty for both attributes price and quality; formally: σRp2=σRq2), and the reward amounts Rp,K and Rq,K, associated with price and quality respectively, of option K.
Fig 5B illustrates the choice probability difference (between option A and option B) with prior mean μC = 0 and reward uncertainty σR2=0.1. Focusing on areas of the graph where a similarity effect can be tested (i.e., top-left and bottom-right), we see that the similarity effect is reproduced by BCV. Moreover, focusing on areas of the graphs where an attraction effect can be tested (i.e., bottom-middle and middle-left), we can see that this effect can also be explained by BCV. Collectively, these simulations provide proof of principle that, for some sets of values of the prior mean μC and of the reward uncertainty σR2, BCV explains both a similarity and an attraction effect. Note that these effects are obtained with μC = 0, which can be considered a default value for this parameter.
Fig 6A and 6B examines the effects of adopting other values of the prior mean μC (fixing the reward uncertainty σR2 to 0.1) in this scenario. This figure shows that an attraction effect is obtained when the prior mean μC is smaller (μC = −2 in our simulation), but no similarity effect emerges. Conversely, a similarity effect is evident when the prior mean μC is larger (μC = 2 in our simulation), but the attraction effect vanishes.
Fig 6C and 6D illustrates the choice probability difference (between option A and option B) for different values of the reward uncertainty σR2 (the prior mean μC was fixed to zero). We can see that both similarity and attraction effects are not detectable when reward uncertainty σR2 is high. For smaller values of uncertainty, a similarity effect emerges but there is no attraction effect. Both effects can be obtained only when the reward uncertainty σR2 is sufficiently low (Fig 5B). This highlights the role of the reward uncertainty σR2 in determining the degree of contextual effects.
In summary, our analyses show that, when simulating multi-attribute decisions with BCV, similarity and attraction effects emerge for appropriate values of the prior mean μC and the reward uncertainty σR2. The first parameter regulates the balance of the two effects, as an attraction effect (but no similarity effect) is obtained when the prior mean μC is small, while a similarity effect (but no attraction effect) is obtained when the prior mean μC is large. Both effects emerge for intermediate values of the prior mean μC, including a prior mean μC = 0, which is a default value for this parameter. The reward uncertainty σR2 plays a key role too, because context effects vanish when this parameter is high. Decreasing levels of reward uncertainty σR2 reveal a similarity effect first and then an attraction effect. These results indicate that the similarity and attraction effects arise naturally from BCV, without any ad-hoc assumptions–and under natural values of model parameters (prior mean μC reward uncertainty σR2).
A compromise effect [17, 25, 27] has been observed when the choice set includes two options scoring high in one attribute and low in another, in addition to a third option with intermediate scores for both attributes. Crucially, the three options are equally preferred (i.e., are chosen an equal amount of times) if presented in pairs during binary choices. However, when they are available altogether, a preference for the option characterized by intermediate scores is seen. We model this scenario by manipulating the distance between attributes for two options A and B, namely assigning Rp,A = 5 − d and Rq,A = 5 + d for option A, and Rp,B = 5 + d and Rq,B = 5 − d for option B, where the proximity parameter d varies (across simulations) from zero to four. To represent the option with intermediate scores for both the two attributes, we assign Rp,K = 5 and Rq,K = 5.
Fig 7A, 7B and 7C shows the prediction of BCV using these settings during binary choices between option K and option A, using different parameters for the prior mean μC and the reward uncertainty σR2. The results indicate that the choice probability difference is always zero, irrespective of the values of the proximity parameter d or the parameters of the model (prior mean μC and reward uncertainty σR2). Fig 7D, 7E and 7F shows the choice probability difference between option K and option A, when option B is also available. For certain values of the parameters (prior mean μC and reward uncertainty σR2), this difference is zero with d = 0 and increases with the proximity parameter d. This effect disappears when reward uncertainty σR2 is too large or when the prior mean μC is too small. Overall, these results show that the compromise effect emerges naturally from BCV, without any ad-hoc assumptions and under default values of the parameters (prior mean μC and reward uncertainty σR2). Interestingly, these simulations predict a correlation between the compromise effect and the proximity parameter d, reflecting differences between the intermediate and extreme options. This phenomenon is predicted by another model of the compromise effect [19] but remains to be validated empirically.
In summary, these simulations provide proof of principle that BCV predicts within-choice contextual effects during multiattribute decisions that are remarkably similar to those seen in empirical studies. In other words, the similarity, attraction and compromise effects seen empirically are all emergent properties of BCV. In the next section, we turn from within choice effects and consider between-choice context effects.
To characterize between-choice context-effects [11], BCV uses the same generative model as above, characterized by a prior belief μC (here we consider only options defined by a single attribute) over reward (with uncertainty σC2) and by an observation of reward amount R (with uncertainty σR2). Here, the generative model is extended to include a Gaussian observation variable O that reflects contextual information provided before an option is presented (Fig 8A). This depends on the hidden cause C and is endowed with uncertainty σO2 (as for the reward amount):
O∼N(C,σO2)
(12)
As above, we assume that an agent infers the posterior expected reward of options afforded by a given context, based on the reward amount but also now on contextual information (i.e., μ^C|O,R). Since the latter is provided before the option, we assume that the agent infers μ^C|0 first and then μ^C|O,R, when the option is presented. Assuming a prior mean equal to zero μC = 0, then:
μ^C|0=σC2σC2+σO2O
(13)
And the posterior uncertainty:
σ^C|O2=σC2−σC2σC2+σO2σC2
(14)
The mean of the posterior distribution P(C|O,R) corresponds to:
μ^C|O,R=μ^C|0+σ^C|O2σ^C|O2+σR2(R−μ^C|0)
(15)
Implying the following incentive value for the option:
V(R)=σ^C|O2σ^C|O2+σ^R2(R−μ^C|0)
(16)
This shows that, other things being equal, information about context (reflected in the value of O) induces subtractive value normalization. For instance, when contextual cues O supports a larger reward, μ^C|0 will be larger and hence the reward prediction error (i.e., R−μ^C|0) will be smaller.
An extension of this generative model is illustrated in Fig 8B, where contexts are organized hierarchically. Combining the influence of reward expectancies within a hierarchy allows the generative model to explain the impact of context at multiple levels. For instance, the value attributed to a certain dish may depend on the reward distribution associated with a restaurant (a more specific context), integrated with the reward distribution associated with a city (a more general context). In detail, a higher-level prior belief about the average reward amount of options (e.g., at the level of the neighbourhood) is represented by a Gaussian distribution with mean μHC equal to zero and uncertainty σHC2, from which a value HC is sampled. Contextual information about HC is provided and represented by HO that is sampled from a Gaussian distribution with mean HC and uncertainty σHO2. A lower-level belief about the average reward amount of options (e.g., the restaurant) is represented by a (Gaussian) distribution with mean HC and uncertainty σLC2, from which a value LC is sampled. Contextual information about LC is provided and represented by LO, which is sampled from a Gaussian distribution with mean LC and uncertainty σLO2. A reward is obtained and sampled from a Gaussian distribution with mean LC and uncertainty σR2.
We propose that agents infer the posterior expectation μ^LC|HO,LO,R P(LC|HO,LO,R) sequentially by estimating μ^HC|HO, μ^LC|HO, μ^HC|HO,LO and finally μ^LC|HO,LO,R. This produces an equation for incentive value with the following form (see Materials and Methods for derivation):
V(R)=K(R−τLOLO−τHOHO)
(17)
Three normalization factors are implicit here. The first (τLOLO) is a subtractive normalization factor proportional to the value LO observed at the low contextual level. The second (τHOHO) is a subtractive normalization factor proportional to the value HO observed at the high contextual level. The terms τ represent gain-dependent effects and describe the relative precision of information conveyed by the low-level (τLO) and high-level (τHO) observations. Finally, a third factor (K) implements divisive normalization and depends on a gain term which includes reward uncertainty (see Materials and Methods for details).
In recent studies [8–11], we have investigated the nature of contextual influence on incentive value that depends on reward expectations established before choice presentation (between-choice effects). In these studies, we have used a simple decision-making task, where participants had to repeatedly choose between a sure monetary reward and a fifty-fifty gamble. These options comprised double the sure monetary reward and a zero outcome, ensuring that the two options had equivalent expected reward or value (EV). Across blocks, we manipulated the distribution of EVs, such that these distributions overlapped. We analysed choice behaviour with EVs common to both contexts to examine whether incentive value attributed to the objective EV changed according to BCV predictions.
In one experiment (Fig 9A and 9B; [8, 9]), in different blocks, the sure monetary gain was drawn from one of two distinct, but partially overlapping, distributions of rewards (low-average and high-average context). Choice behaviour was consistent with attributing a larger incentive value to common EVs in the low average compared to high-average context. This and similar evidence [5–14, 50] suggests that incentive values are, to some extent, rescaled to the average reward expected in a given context, such that they increase (resp. decrease) with smaller (resp. larger) average reward expectations.
These data fit within predictions of BCV. In addition, BCV postulates a between-choice influence of expected reward variance on incentive values (Fig 9C and 9D). In a recent study [11], we used the same gambling task described above and manipulated contextual variance on two levels; one associated with blocks where two target trial EVs were presented (low-variance context), and another with blocks where the same two target trial EVs plus a larger and a smaller EV were presented (high-variance context). Crucially, this ensured that the two contexts had equivalent average reward but different variance. BCV predicts that the incentive value of the smaller target trial EV will be lower in the low-variance compared to the high-variance context, and the incentive value of the larger target trial EV will be higher in the low-variance compared to the high-variance context. In other words, BCV predicts a larger value difference between the two target trial EV in the low compared to high-variance context. This derives from the gain term, which depends on contextual reward variance. Specifically, low variance magnifies the reward prediction error and hence further reduces the value of rewards that are lower than expected and enhances the value of rewards that are larger than expected. We have previously provided data that are consistent with this prediction [11].
This latter study supports the hypothesis that between-choice reward variance influences incentive value consistent with BCV. In the same study (Fig 9E and 9F; [11]), we also reported that between-choice context effects can be expressed at different hierarchical levels, in line with predictions of BCV. Participants played a computer-based task, where two decks of cards (representing a low-level context) appeared. Each card was associated with a monetary reward, and decks contained cards with different average rewards. A card was drawn from a selected deck and participants had to choose between half of the card reward for sure and a gamble between the full reward and a zero outcome, each with 50% chance. Two sets of decks (representing a high-level context) alternated in a pseudo-random way. The empirical data showed that the lowest incentive values were attributed when both high-value decks and deck-sets were simultaneously presented, while the highest incentive values were attributed when low-value decks and deck-sets were simultaneously presented. Intermediate incentive values were attributed when decks and deck-sets had one high value and the other low value.
Collectively, these empirical studies provide evidence consistent with between-choice contextual effects on incentive value that depends on beliefs about the average reward and variance expected across choices at multiple hierarchical levels. Furthermore, the empirical findings endorse the predictions derived from BCV.
We advance BCV as a unifying theory of contextual effects in value-based choice under the normative principles of Bayesian statistics. BCV assumes that the brain calls on Bayesian inference to invert a generative model and compute (independently for each attribute) the average reward based on observing different reward amounts of options that are available in a given context. Our key proposal is that incentive value emerges during this inferential process, and corresponds to a precision-weighted reward prediction error. Here, we show that these principles are sufficient to explain a wide range of between-choice and within-choice contextual influences; in the latter case encompassing both single and multiattribute effects. To our knowledge, this is the first time a theory has been applied to the full range of context effects.
An important advantage of BCV is its grounding in normative principles of Bayesian statistics [32–37]. Several arguments have been made in support of a Bayesian approach. These are based on a formal and clear definition of the functions that motivate cognitive processes, which are formulated as Bayesian inference and learning. This allows BCV to establish a direct link with Bayesian schemes in other domains–a step towards formulating a unifying theory of brain function. Remarkably, we show that the same basic processes postulated by BCV can be applied to a wide range of conditions in which contextual effects on value and choice are involved. Beyond explaining the available empirical evidence, this scheme can generate new hypotheses (see below). Indeed one of our previous studies [11] was motivated by testing predictions arising out of our initial formulations of BCV.
BCV is associated with planning as inference and active inference [40–45]. The basic idea is that an agent considers the rewards on offer as samples drawn from a population. The latter is not known directly, but can be inferred based on the rewards on offer. Heuristically, agents are interested in inferring how much reward is available on a given trial, which they estimate by combining prior expectations with observations of available rewards. On this view, agents primarily aim to infer–and not maximize–the reward; implying that utility-maximization is an emergent process. We argue that an advantage of this perspective is that it offers a normative interpretation of contextual effects, which emerge from the inferential treatment offered here.
Although our theoretical treatment is grounded in Bayesian inference one might argue that the Bayesian gloss is unnecessary to understand the particular inferential mechanisms we have called upon [51]. To a certain extent, there is tautology in Bayesian explanations for behaviour. This follows from the complete class theorem (i.e., for every loss function and behaviour there is a prior belief that renders the behaviour Bayes optimal) [52]. In other words, in principle, everything is Bayes optimal under some priors. This means that the interesting questions reduce to the form of prior beliefs that constitute a subject’s generative model. Our focus has been on the form of these models and the particular role of precision weighting in belief updating and choice. The results of our analysis are consistent with empirical data on several forms of context effect, and hence may contribute to a clarification of the computational principles at play.
In BCV incentive value, and in turn choice behaviour, emerges from Bayesian belief updating. Under continuous state space models of the hidden causes of reward values, belief updates and incentive value can be cast as precision-weighted (reward) prediction error. A possibility consistent with BCV is that action is steered by (precision-weighted) prediction errors and is oriented to error cancellation, with approach and avoidance responses elicited by positive and negative prediction errors, respectively. The crucial role of prediction error highlights a perspective in which incentive value is inherently relative with respect to reward expectation. Eliciting approach and avoidance behaviour in response to positive and negative prediction errors can be conceived as a basic error-cancellation process (crystallized during evolution of biological organisms), which is a core tenet of active inference schemes.
BCV postulates that the two fundamental determinants of incentive value are prediction error and relative precision. A prediction error is determined by the difference between the observed and expected reward which, in BCV, derives from integrating different expectations under contextual uncertainty. Relative precision depends on the (relative) precision or prior confidence–and ensures that the prediction error is normalised and (Bayes) optimally weighted in relation to uncertainty about both context and reward cues. BCV predicts precision exerts an influence in two ways. First, at high hierarchical levels, precision determines the optimal integration of multiple contextual representations–as it mandates that contexts characterized by a high precision (greater reliability) will exert more influence on reward expectancy. For instance, if we assume that subjects have very precise beliefs about the low-level context (e.g., the card deck in the final experiment on between-choice context effects), then the effect of the high-level (e.g., the deck set) will disappear. Formally, this is because in the hierarchical model the low-level context constitutes a Markov blanket for the posterior expectation about the reward option (Bishop, 2006). In other words, the effect of the high-level context tells us that if subjects are using a hierarchical model, there must be posterior uncertainty about the low-level context. Heuristically, even though they can see which deck they are currently playing with, they still nuance their expectations about this deck based upon the deck-set from which it came. Second, at the lowest hierarchical level, precision determines the gain assigned to the prediction error and hence is a direct determinant of incentive value.
Within BCV, the ratio between reward uncertainty and prior uncertainty determines the gain term (or relative precision) which is used for belief updating (see Eq 6). This means that manipulating the prior uncertainty produces exactly opposite effects compared to manipulating the reward uncertainty, meaning that varying one during simulations is sufficient for testing the predictions of the model (above, we manipulated reward uncertainty and kept the prior uncertainty constant). Thus BCV has only two parameters; namely, the prior mean and reward uncertainty. The role of the latter is straightforward, as context effects are allowed only with small reward uncertainty, and the size of these effects decreases with this reward uncertainty. The role of the prior mean is more complex: for instance, a large prior mean permits a similarity effect but interferes with an attraction effect, while a small prior mean allows an attraction effect but interferes with a similarity effect. Notably, all contextual effects are expressed when setting the prior reward expectation to zero, which can be considered the default value. In short, relying on only two parameters endows BCV with simplicity and constrains the predictions that can be derived, making BCV easy to validate or falsify (see below).
We have shown that the principles underlying BCV can explain a wide range of empirical findings on the context sensitivity of value-based choice. Several previous accounts have focused on a single context effect, especially during multiattribute decisions. Some models have been developed explicitly for explaining the similarity effect [20, 53–55], other models for explaining the attraction effect [56, 57], and other models for the compromise effect [27]. However, a shortcoming of these models is their inability to explain all three effects within a single formal framework. More recently, adopting connectionist architectures, the multi-alternative decision field theory [16, 58, 59] and the leaky competing accumulator [19, 59, 60] have been able to reproduce all three effects (see also [61]). The first model [16, 58] is based on a process modelling attentional switches across attributes and a comparator mechanism which, for the attribute under attention, computes the difference between the reward of each option and the mean reward across options. The second model [19, 59, 60] is similar, except that the comparator applies a non-linear asymmetric (loss-averse) value function to the difference. Although these models fit remarkably with empirical literature and shed light on the neural mechanisms underlying choice, we argue that BCV presents several advantages. First, it is based on normative principles of Bayesian inference. This constrains the model in terms of empirical predictions. In other words, the similarity, attraction and compromise effect are implicit in the way the model works. In fact, these effects arise when defaults parameters are used. Second, BCV is a more parsimonious model; as the number of free parameters is much lower (essentially, the prior mean and the reward uncertainty). Third, without any further assumptions, BCV applies to a wider range of phenomena including single-attribute decisions and also accounts for between-context effects. Overall, while previous connectionist models are informative especially at the implementation level, BCV helps clarify context sensitivity at the algorithmic and computational level.
The concept of wealth in expected utility theory [3] and status quo in prospect theory [62] have been recently re-casted in terms of average expected reward [29]. This formulation opens the possibility of context effects dependent on changes in reward expectation. In line with this view, empirical evidence indicates a between-choice context effect that depends on the average contextual reward (as for example inferred from past choices), consisting in attributing larger incentive values in contexts characterized by lower reward. A similar idea has inspired decision by sampling theory [14, 31], which evokes a few basic cognitive processes to explain choice behaviour. According to this model, each choice option elicits retrieval from memory (in the form of random sampling) of stimuli encountered in the past, especially those associated with the current context. A set of binary comparisons follows between the option and the samples, and the number of comparisons in which the option is favoured over each sample is recorded. This number corresponds to the incentive value of the option and is computed for all options available, hence determining their relative preference. Since samples are drawn from memory, they depend on past experience and therefore reflect the distribution of options and outcomes characterizing the environment of an agent. This model can account for an attribution of larger incentive value to the same reward in contexts where lower compared to higher reward is expected before options are provided. This effect is explained by a decreased likelihood, in the former compared to the latter context, of sampling stimuli from memory that are preferred to rewards common to both contexts (assuming a recency effect in memory sampling; [14, 31]. BCV extends these views by appealing explicitly to Bayesian principles (i.e. Bayesian belief updating and evidence accumulation), with implications for empirical predictions. For instance, contrary to BCV and empirical findings, it remains unclear whether these previous models can account for between-choice contextual influence of reward variance or any within-choice contextual effects.
Divisive normalization theory [6, 63–68] has been proposed recently to explain both between-choice and within-choice contextual effects during single attribute decisions. Divisive normalisation was first proposed in the sensory domain to explain phenomena such as neural adaptation within the retina to stimuli of varying intensity [63]. There is evidence that similar principles can explain higher-order cognitive processes, such as selective attention and perceptual decision-making [63, 69]. Recently, divisive normalisation has been extended to contextual adaptation effects in value-guided choice [6], and proposes that incentive value corresponds to the reward divided by the average reward of past or current choices. This can explain contextual influences elicited both within-choice effects during non-multiattribute decisions and between-choice effects that depend on the average contextual reward. Though this scheme relies on a normalization scheme similar to BCV, different empirical predictions arise. It remains unclear whether this divisive normalization scheme is able to explain between-choice effects deriving from reward variance, and can explain data on multi-attribute choices. In addition, BCV, but not divisive normalization theory, is based on normative principles of Bayesian statistics. However, an attractive aspect of divisive normalization theory is the explicit connection with mechanisms characterizing biological neural processes [63]. A similar connection can be motivated for BCV, given several proposals showing how Bayesian inference (the framework of BCV) is compatible with neuronal processes [49, 70, 71].
The manner in which BCV conceptualizes incentive value is similar to recent economic models that postulate incentive value is adapted to the statistics of the expected reward distribution [29, 30]. These theories can be broadly classified into those based on subtractive normalization, which assume that incentive value corresponds to the reward minus a reference value [29], and those based on divisive normalization, assuming that incentive value corresponds to the reward divided (or multiplied) by the range of an expected distribution of rewards [30]. An important difference between BCV and these theories is the derivation of the former but not the latter from normative assumptions of Bayesian inference. From Bayesian belief updating, BCV derives the proposition that incentive value corresponds to precision-weighted prediction error, hence implying both a subtractive normalization to the expected reward and a divisive normalization with respect to the reward uncertainty. Importantly, these predictions are not ad hoc but derive from Bayesian assumptions, distinguish BCV from other models, and have been recently supported empirically [11]. In addition, while these recent economic models focus on between-choice context effects, BCV is more general as it can reproduce within-choice effects in both single and multiattribute decisions.
Like BCV, a recent proposal has interpreted multi-attribute within-choice effects based on the notion that perception of reward is stochastic [72]. The idea is that, for each attribute, an agent forms noisy observations of reward amounts and of the ordinal positions of the reward amounts. Multi-attribute effects can then be obtained by integrating these two observations [72]. Though there are analogies between BCV and the model of Howes et al. [72], we emphasize several important differences. First, the latter does not employ a Bayesian framework, since it is not based on integrating prior beliefs and observations, nor it is based on optimal weighting of different sources of information (as in multi-sensory integration). Second, the model of Howes et al. [72] has been applied to aspects of multi-attribute effects (such as the impact on reaction times), which remain to be explored with BCV. On the other hand, the model of Howes et al., [72] remains to be explored in relation to within-choice effects involving a single attribute and in relation to between-choice effects.
Specific empirical predictions can be derived from BCV, and here we highlight some of these. Standard economic theories assume that choice should be independent of whether options are presented simultaneously or sequentially. However, the latter case remains largely to be investigated. BCV may inspire this investigation, as it predicts that a higher value will be attributed to an option after presentation of lower value options. This because BCV proposes a sequential belief updating in which options considered so far contextualize the option observed now. Other predictions involve interactions regarding between- and within-choice effects. For example, consider the example above in which an agent usually evaluates equally car A (expensive and high quality) and car B (cheap and low quality). One may design an experiment where participants are first exposed to a set of cars having a fixed level of quality and varying on price. BCV predicts that this manipulation would determine a lower reward uncertainty for quality compared to price. In other words, quality would become more salient than price, predicting a preference for car A over car B. In addition, BCV predicts other forms of interactions regarding between- and within-choice effects dependent on manipulations of the reward uncertainty and the prior mean (see above), which also remain to be explored empirically. Finally, BCV may be relevant for research on the neural underpinnings of decision-making. A main aspect of this theory is the idea that incentive value corresponds to a precision-weighted reward prediction error. Interestingly, reward prediction error is reflected in activity of brain regions involved in reward processing [73]. BCV raises the possibility that a stimulus which elicits a stronger prediction error response in the brain will be attributed a higher incentive value.
There are shortcomings to BCV, though we argue that the same framework may be fruitfully used to address some of these shortcomings. A shortcoming of our current formulation assumes that model parameters are given. In reality, these parameters need to be learned in the first place. Questions about the mechanisms that might underpin learning of generative models adopted for Bayesian inference are still largely open, though substantial contributions exist, particularly in the context of structure learning [74–80]. A second shortcoming is that here we have assumed that choices occur after inference has considered all observations. An important extension of BCV is a consideration that action tendencies actually develop during evidence accumulation, and this speaks to models of choice that focus on action dynamics, sequential policy optimisation and reaction times [16, 46, 47]. Another important extension of BCV would be to generalize to domains outside incentive value computation. Context effects similar to those observed in value-based decision-making have been reported in many other conditions during perception and judgement [81–84]. Notably, multi-attribute context effects have been recently shown outside incentive value computation [85, 86], suggesting that they may derive from a general way in which the brain works [61].
We offer BCV as a unifying theory of contextual effects during choice behaviour based on Bayesian normative principles. BCV predictions are in line with available empirical evidence about context sensitivity seen empirically both within and between-choice. These different effects are explained using the same simple set of principles, invoking minimal assumptions. We argue that strengths of this model are its foundation on normative principles, simplicity, the link with other influential models of brain function, and the ability to explain a wide range of empirical data. This theory may help clarify the nature of incentive value attribution and choice behaviour. This is particularly prescient when trying to understand ecological phenomena and psychopathologies characterized by dysfunctional choice, such as addiction.
Here we derive Eq 17 from the generative model shown in Fig 9B. A higher-level contextual variable (e.g., a neighbourhood containing several restaurants) is represented by a Gaussian distribution with mean μHC equal to zero and uncertainty σHC2, from which a value HC is sampled. Sensory evidence about HC is provided and represented by HO which is sampled from a Gaussian distribution with mean HC and uncertainty σHO2. A lower-level contextual variable (e.g., one of the restaurants) is represented by a (Gaussian) distribution with mean HC and uncertainty σLC2, from which a value LC is sampled. Sensory evidence about LC is provided and represented by LO, which is sampled from a Gaussian distribution with mean LC and uncertainty σLO2. A reward is obtained and sampled from a Gaussian distribution with mean LC and uncertainty σR2. The posterior distribution P(LC|HO,LO,R) can be inferred sequentially in the order P(HC|HO), P(LC|HO), P(LC|HO,LO), and P(LC|HO,LO,R). The posterior mean of P(HC|HO) is:
μ^HC|H0=σHC2σHC2+σHO2HO
(18)
And the posterior uncertainty:
σ^HC|H02=σHC2−σHC2σHC2+σHO2σHC2
(19)
The posterior mean of P(LC|HO) is equal to μ^HC|H0 (μ^LC|H0=μ^HC|H0), while the posterior uncertainty is:
σ^LC|H02=σ^HC|H02+σLC2
(20)
The posterior mean of P(LC|HO,PO) is:
μ^LC|HO,LO=μ^LC|H0+σ^LC|HO2σ^LC|HO2+σLO2(LO−μ^LC|H0)
(21)
And the posterior uncertainty:
σ^LC|H0,LO2=σ^LC|H02−σ^LC|HO2σ^LC|HO2+σLO2σ^LC|H02
(22)
The posterior mean of P(LC|HO,LO,R) is:
μ^LC|HO,LO,R=μ^LC|HO,LO+σ^LC|HO,LO2σ^LC|HO,LO2+σR2(R−μ^LC|HO,LO)
(23)
Finally, with few rearrangements, we obtain the following incentive value for a reward offer:
V(R)=σ^LC|HO,LO2σ^LC|HO,LO2+σR2(R−σ^LC|HO2σ^LC|HO2+σLO2LO−σLO2σ^LC|HO2+σLO2σHC2σHC2+σHO2HO)
(24)
This equation implements three normalization factors: (i) a subtractive normalization factor (σ^LC|HO2σ^LC|HO2+σLO2LO) proportional to the value LO observed at the low contextual level, (ii) a subtractive normalization factor (σLO2σ^LC|HO2+σLO2σHC2σHC2+σHO2HO) proportional to the value HO observed at the high contextual level, (iii) a divisive normalization factor (σ^LC|HO,LO2σ^LC|HO,LO2+σR2) that captures the weighting dependent on the (relative) reward uncertainty. If we define the three factors as τLO and τHO and K respectively, we obtain Eq 17.
|
10.1371/journal.ppat.1007105 | Caspase-11-dependent pyroptosis of lung epithelial cells protects from melioidosis while caspase-1 mediates macrophage pyroptosis and production of IL-18 | Infection with Burkholderia pseudomallei or B. thailandensis triggers activation of the NLRP3 and NLRC4 inflammasomes leading to release of IL-1β and IL-18 and death of infected macrophages by pyroptosis, respectively. The non-canonical inflammasome composed of caspase-11 is also activated by these bacteria and provides protection through induction of pyroptosis. The recent generation of bona fide caspase-1-deficient mice allowed us to reexamine in a mouse model of pneumonic melioidosis the role of caspase-1 independently of caspase-11 (that was also absent in previously generated Casp1-/- mice). Mice lacking either caspase-1 or caspase-11 were significantly more susceptible than wild type mice to intranasal infection with B. thailandensis. Absence of caspase-1 completely abolished production of IL-1β and IL-18 as well as pyroptosis of infected macrophages. In contrast, in mice lacking caspase-11 IL-1β and IL-18 were produced at normal level and macrophages pyroptosis was only marginally affected. Adoptive transfer of bone marrow indicated that caspase-11 exerted its protective action both in myeloid cells and in radio-resistant cell types. B. thailandensis was shown to readily infect mouse lung epithelial cells triggering pyroptosis in a caspase-11-dependent way in vitro and in vivo. Importantly, we show that lung epithelial cells do not express inflammasomes components or caspase-1 suggesting that this cell type relies exclusively on caspase-11 for undergoing cell death in response to bacterial infection. Finally, we show that IL-18’s protective action in melioidosis was completely dependent on its ability to induce IFNγ production. In turn, protection conferred by IFNγ against melioidosis was dependent on generation of ROS through the NADPH oxidase but independent of induction of caspase-11. Altogether, our results identify two non-redundant protective roles for caspase-1 and caspase-11 in melioidosis: Caspase-1 primarily controls pyroptosis of infected macrophages and production of IL-18. In contrast, caspase-11 mediates pyroptosis of infected lung epithelial cells.
| Burkholderia pseudomallei is a bacterium that infect macrophages and other cell types and causes a diseases called melioidosis. Inflammasomes are multiprotein complexes that control activation of the proteases caspase-1 and caspase-11 resulting in production of the inflammatory mediators IL-1β and IL-18 and death of infected cells. Mice deficient of caspase-1 or caspase-11 are more susceptible to infection with B. pseudomallei or the closely related B. thailandensis. Here we show that absence of caspase-1 completely abolished production of IL-1β and IL-18 as well as death of macrophages infected with B. thailandensis. In contrast, in the highly susceptible caspase-11-deficient mice, IL-1β and IL-18 production and macrophages death were not significantly affected. Rather, absence of caspase-11 abolished death of infected lung epithelial cells. Taken together, our results show that caspase-1 and caspase-11 have non-redundant protective roles in melioidosis: Caspase-1 primarily controls cell death of infected macrophages and production of IL-18. In contrast, caspase-11 mediates cell death of infected lung epithelial cells.
| Burkholderia pseudomallei is a Gram-negative flagellated bacterium that causes melioidosis, a diseases endemic to South-East Asia and other tropical regions and the most common cause of pneumonia-derived sepsis in Thailand [1, 2]. Due to global warming and increased international travel, cases of melioidosis are increasingly being reported outside the endemic areas. B. pseudomallei infection can be contracted through ingestion, inhalation, or subcutaneous inoculation and leads to broad-spectrum disease forms including pneumonia, septicemia, and organ abscesses. Although not pathogenic to humans, Burkholderia thailandensis possesses several of the B. pseudomallei virulence factors, causes morbidity and mortality in mice, and is often used as a model for melioidosis [3–5]. Following infection of macrophages and other non-phagocytic cell types, Burkholderia is able to escape the phagosome and invade and replicate in the host cell cytoplasm. Macrophages and IFNγ have been shown to play a critical role in protection from melioidosis [6–8]and several B. pseudomallei virulence factors have been identified. Analysis of mouse strains with different susceptibility to B. pseudomallei infection indicates that the early phases of the infection are crucial for survival, emphasizing the necessity for better understanding of innate immune responses during melioidosis.
Burkholderia has been shown to activate TLR2, TLR4, and TLR5 in epithelial reporter cell line [9]. Interestingly, while Myd88-/- mice are highly susceptible to B. pseudomallei infection [10], Tlr4-/- mice have similar resistance to wild type (WT) mice but Tlr2-/- mice showed reduced mortality [11] indicating that MyD88-dependent pathways may play opposite role in melioidosis. This notion is supported by our previous works that showed that IL-18 was protective in melioidosis while IL-1β was deleterious because of excessive neutrophils recruitment to the lung and tissue damage due to release of neutrophil elastase [12, 13].
Caspase-1 has been shown to be protective against Burkholderia infections [14]. Production of IL-1β and IL-18 in melioidosis is regulated by activation of caspase-1 downstream of the NLRP3 inflammasome while activation of the NLRC4 inflammasome triggers the pyroptotic cell death process [12, 15]. A potentially confounding factor that affects all the works that examined the role of caspase-1 in melioidosis is that those studies relied on caspase-1-deficient mice that also lacked caspase-11. The non-canonical inflammasome composed of caspase-11 (encoded by Casp4) has also been shown to play a protective role in melioidosis [16] by recognizing cytoplasmically-located LPS [17, 18]. This process is dependent on priming of macrophages with interferons or TLR ligands. The mechanism through which caspase-1 and caspase-11 initiate pyroptosis is by cleaving gasdermin D, a cellular protein that open pores in the cell membrane [19, 20].
As for infections by most intracellular bacteria, IFNγ is an essential component of the innate immune response to B. pseudomallei and its absence results in severely decrease resistance to the infection [6–8]. The antimicrobial properties of IFNγ are mediated by several effector mechanisms operating in a variety of cell types. Among the thousand IFN-stimulated genes, IFN-induced GTPases, iNOS, and NADPH oxidase are the most studied and effective antimicrobial effector mechanisms of macrophages. Recently it has been proposed that in the early phase of Burkholderia infection caspase-11 may act as an IFNγ-inducible effector mechanism because of its reliance on the IL-18-IFNγ axis for priming, which would place non canonical inflammasome downstream of canonical caspase-1 activation [15].
The function of IFN-inducible effector mechanisms such as iNOS, ROS, Guanylate binding proteins, and caspase-11 in melioidosis has been examined previously [15, 16, 21, 22]. While iNOS and several GBPs do not seem to be required to survive a lethal infection with B. pseudomallei or B. thailandensis, absence of NADPH oxidase or caspase-11 renders mice significantly more susceptible. For this reason we decided to determine the relative contribution of caspase-11 and NADPH oxidase to the protection conferred by IFNγ against B. thailandensis infection. We also revisited the role of caspase-1 independently of caspase-11 using recently generated bona fide caspase-1-deficient mice [23]. The results presented here support the following conclusions: first, the main protective role of caspase-1 in melioidosis is to trigger pyroptosis in macrophages and production of IL-18. Second, caspase-11’s function during B. thailandensis infection is to mediate pyroptosis in lung epithelial cells, rather than in macrophages. Finally, the protective action of IFNγ is mediated by ROS independently of caspase-11.
We and others have previously shown that caspase-1 plays a critical role during intranasal or intraperitoneal infection with Burkholderia [12–14]. However, those results were obtained using caspase-1-deficient mice that also lacked caspase-11 due to a passenger mutation in the 129 mouse strain. “Pure” caspase-1-deficient mice have been recently generated [23] and for this reason we decided to reexamine the role of caspase-1 and caspase-11 in melioidosis. Analysis of the survival of mice infected intranasally with B. thailandensis showed that both Casp1-/- and Casp11-/- mice were significantly less resistant than WT mice and became moribund within 5 days of infection and had to be euthanized (Fig 1A). The bacterial burdens in different organs were measured at different time points post-infection (p.i.) and in mice infected using different doses and revealed the relative susceptibility of Casp1-/-, Casp11-/-, and Casp1-/-/Casp11-/- mice (Fig 1B). Mice lacking both caspases had the highest amount of bacteria while mice deficient in caspase-11 were clearly more susceptible than Casp1-/- mice. This pattern has been previously observed in a study that compared Casp11-/- mice to mice that lacked both ASC and NLRC4 (used as surrogate for pure caspase-1 deficient mice) [15]. In agreement with the function of caspase-1 in the generation of the mature form of IL-18, this cytokine was absent in BALF or serum of Casp1-/- mice while was detected at the same level in Casp11-/- or WT mice (Fig 1C). The levels of IL-1β, a deleterious factor in melioidosis [12, 13], were decreased in Casp11-/- mice compared to WT mice. However, when Casp11-/- mice were infected with a lower inoculum and examined at different time points p.i., IL-1β (like IL-18) was detected in their BALF at the same level, or even higher, than in WT mice (Fig 1D). Interestingly, neutrophils were present in the BALF of Casp11-/- mice in significantly reduced number than in WT mice 48 and 72 hours p.i. (Fig 1E). This phenotype, however, was not observed 24 hours p.i., was not due to impaired production of TNFα, IL-6, KC, or MCP-1, and correlated with the increased bacterial burdens in organs (S1 Fig). Whether absence of caspase-11 negatively affects recruitment, survival, or permanence of neutrophils in the infected lung cannot be determine at present and will be the focus of future studies. Impaired neutrophil recruitment has been previously observed in Casp11-/- mice during lung infection with K.pneumoniae [24].
Analysis of bone marrow-derived macrophage (BMM) cultures infected with B. thailandensis confirmed that IL-1β and IL-18 production is mediated by caspase-1 and not caspase-11 (Fig 2A). Induction of pyroptosis in these cells was primarily dependent on caspase-1 with negligible contribution of caspase-11 (Fig 2B). For these experiments BMM were primed O/N with IFNγ. However, when bone marrow-derived dendritic cells (S2 Fig) or BMM were primed with IFNγ concomitantly to infection (see below), caspase-11 contribution to pyroptosis was modest but statistically significant. Intracellular B. thailandensis replication in BMM inversely correlated to occurrence of pyroptosis with maximal bacterial count in Casp1-/- and Casp1/Casp11-/- cells (Fig 2B). B. thailandensis replication in Casp11-/- cells was moderately higher than in WT cells. Interestingly, IFNγ priming of cells significantly decreased bacteria replication in all strains emphasizing the importance of inflammasomes-independent microbicidal mechanisms activated by IFNγ (see below). Although B. thailandensis has been shown to serve as a useful model for melioidosis, it is not pathogenic in humans and differs in many aspects from the virulent B. pseudomallei. Therefore, it was important to determine the role of caspase-11 even during infection with B. pseudomallei. As shown in S3 Fig, two light emitting B. pseudomallei clinical isolates robustly replicated in Casp1-/-/Casp11-/- macrophages but to a much lower degree in WT and Casp11-/- cells, again indicating the prominent role of caspase-1, rather than caspase-11, in restriction of intracellular replication of Burkholderia in macrophages.
Taken together, these results conclusively indicate that the increased susceptibility of Casp1-/- mice to melioidosis is likely due to their inability to produce IL-18 and trigger pyroptosis in myeloid cells. In contrast, the reason for the susceptibility of Casp11-/- mice remains unclear but does not appear to be due to lack of IL-18 or gross inability to trigger pyroptosis in macrophages.
The fact that pyroptosis of B. thailandensis-infected macrophages and cytokine processing in these cells appeared mostly dependent on caspase-1 with minor involvement of caspase-11 raised the question of why Casp11-/- mice appeared so susceptible to melioidosis. Caspase-11 function has been extensively studied in myeloid cells but its role in non-hematopoietic cells has been mostly neglected. To increase our understanding of the role played by Caspase-11 during melioidosis we performed bone marrow transplant experiments. As shown in Fig 3A, the bacterial burden in different organs of WT mice reconstituted with Casp11-/- bone marrow cells was significantly higher compared to WT mice receiving WT cells, as expected for a hematopoietic role for caspase-11. However, Casp11-/- mice reconstituted with WT bone marrow cells still had organ bacteria burdens much higher than WT mice indicating that caspase-11 also plays a protective role in the radio-resistant cell compartment. Analysis of bone marrow and spleen cells indicated complete and equally effective reconstitution by both genotypes. (S4 Fig). However, confirming what observed in Casp11-/- mice (Fig 1E), severely decreased neutrophil numbers were detected into the lung of mice reconstituted with caspase-11-deficient bone marrow (Fig 3B and S4 Fig). These results suggest that caspase-11 plays a protective role not only in hematopoietic cells but also in radio-resistant cell types.
B. thailandensis can infect several cell types including lung epithelial cells [25, 26]. The mouse lung epithelial cell line TC-1 is often used to study the lung innate immune response to bacteria [27, 28]. TC-1 cells were readily infected with B. thailandensis (S5A Fig). To test whether caspase-11 can be activated in lung epithelial cells infected with B. thailandensis, TC-1 cells were incubated with Biotin-VAD-FMK, a cell permeable caspase pseudosubstrate that irreversibly binds to active caspase [29]. As shown in Fig 4A, caspase-11 could be pulled down from B. thailandensis-infected TC-1 cell lysates using streptavidin agarose. The B. thailandensis bsaZ mutant, which is unable to escape the phagosome, activated caspase-11 to a much lower degree. Caspase-11 expression in TC-1 cells was strongly induced by TNFα/IFNγ while expression of the canonical inflammasome components NLRP3, NLRC4, ASC, Caspase-1 was not detectable (S5B Fig) suggesting that caspase-11 may be the only pathway available in TC-1 cells to trigger pyroptosis. This was confirmed by knocking-out caspase-11 gene in TC-1 cells using CRISPR/CAS9 technology (Fig 4B). While the control TC-1 cells underwent pyroptosis upon infection with B. thailandensis, the caspase-11-deficient TC-1 cells were resistant to B. thailandensis-induced cell death (Fig 4C). Intracellular B. thailandensis replication proceeded unrestrained in TC-1 cells lacking caspase-11 but was significantly restricted in control TC-1 cells (Fig 4C). TC-1 cells expressed IL-18 mRNA upon treatment with TNFα/IFNγ though IL-18 secretion could not be detected in conditioned culture supernatants (S5C Fig). To test whether cell death of lung epithelial cells occurs in vivo, we exposed WT or Casp11-/- infected mice to a green fluorescent compound that stains the nucleus of necrotic cells in situ. Mice were euthanized shortly thereafter and the lungs were fixed in paraformaldehyde. Histological sections were counterstained with the epithelial marker EpCAM and analyzed by immunofluorescence microscopy to visualize necrotic epithelial cells in situ. As shown in Fig 4D and 4E, significantly decreased cell death of lung epithelial cells was observed in lung sections of Casp11-/- mice compared to WT mice. It was recently shown that pyroptotic cells trap intracellular bacteria and are phagocytosed by neutrophils [30]. Our preliminary results (S5D Fig) suggest that pyroptotic epithelial cells encounter the same fate and are phagocytosed by neutrophils and macrophages. Taken together, these results suggest that one of the most critical functions of caspase-11 in melioidosis is to control pyroptosis of lung epithelial cells.
IL-18 is a potent inducer of IFNγ, a cytokine required to survive infection with B. thailandensis [12, 31]. In agreement with this activity of IL-18 and with previous works from our and others labs, IFNγ production was severely decreased in Casp1-/- mice but not Casp11-/- mice (Fig 1C) suggesting lack of IFNγ as a possible mechanism to explain the protective effect of IL-18. Confirming this hypothesis, administration of recombinant IFNγ significantly reduced organ bacteria burdens in Il18-/- mice infected with a lethal dose of B. thailandensis (Fig 5A) showing that IFNγ is sufficient to mediate the protective action of IL-18. However, IL-18 performs other functions and, therefore, we asked whether IFNγ was necessary for the protective action of IL-18. As shown in Fig 5B, administration of recombinant IL-18 significantly decreased organ bacterial burdens in WT mice but not in Ifngr1-/- mice. IL-18 treatment induced IFNγ in both mouse strains (Fig 5C). Taken together, these results demonstrate that IFNγ is necessary and sufficient to mediate the protective action of IL-18.
We next turned our attention on the role of IFNγ in melioidosis. A number of studies including from our group [6–8, 12], have demonstrated the protective role of IFNγ during B. thailandensis infection though the mechanism of protection remains undefined. IFNγ is known to activate several microbicidal mechanisms that are critical for killing intracellular bacteria, including different families of GTPases, NRAMP1, NADPH oxidases and iNOS. Recent work suggested that caspase-11 should also be considered as an IFNγ-inducible mechanism [15]. Because both ROS and caspase-11 have been shown to be protective against B. thailandensis infection while iNOS, NRAMP1, or GBP do not appear to play a significant role in the innate immune response against this bacterium [15, 21, 22, 32], we decided to determine to what degree the protective action of IFNγ in melioidosis is mediated by either ROS or caspase-11.
As shown in Fig 6A, administration of recombinant IFNγ to intranasally infected mice significantly reduced the organ bacterial burdens in WT, Casp1-/-/Casp11-/- and Casp11-/- mice but not in Cybb-/- mice (deficient in the gp91 subunit of the NADPH oxidase), suggesting that production of ROS is an essential microbicidal mechanism triggered by IFNγ against B. thailandensis.
The importance of ROS as mediator of IFNγ protection was also observed in culture of BMMs infected with B. thailandensis (Fig 6B). Intracellular bacteria replication was drastically higher in Casp1-/-/Casp11-/- and Casp11-/- macrophages compared to WT cells and correlated with the decreased pyroptosis in cells lacking either caspase. Intracellular B. thailandensis replication was also elevated in Cybb-/- macrophages but this was not due to decreased pyroptosis, which was not significantly different than in WT cells. Treatment with IFNγ significantly restricted bacteria replication in WT, Casp1-/-/Casp11-/-, and Casp11-/- cells but, importantly, not in Cybb-/- cells. The decreased bacteria replication was not due to higher rate of pyroptosis, which was not significantly affected by treatment with IFNγ. Taken together, these results suggest that production of ROS plays a predominant role in the antimicrobial action of IFNγ.
Confirming the protective role of NADPH oxidase downstream of IFNγ, ex vivo generation of ROS was significantly impaired in neutrophils or macrophages obtained from the BALF of infected Il18-/- or Ifngr1-/- mice 14 hours or 48 hours p.i. (Fig 7A). The organ bacterial burdens in these mice inversely correlated with the amount of ROS produced (Fig 7B). Importantly, administration of the antioxidant N-acetyl-cysteine (NAC) dissipated the protective effect of exogenous IFNγ administration (Fig 7C), again reinforcing the notion that ROS is a major mediator of the protective effect of IFNγ in melioidosis. NAC treatment had no effect on the production of IL-1β or IL-18, whose BALF levels correlated with the bacterial burdens (Fig 7D).
The innate immune response to lung infection with Burkholderia species has been examined in a few papers but much remains to be learned. Here we have analyzed the role of the canonical and non-canonical inflammasomes and of the IL-18-IFNγ axis in a mouse model of melioidosis. We and others have previously shown that processing and secretion of the mature form of IL-1β and IL-18 in response to Burkholderia infection was dependent on caspase-1 [12, 14, 15]. The caveat of those studies is that they were performed using mice that also lacked caspase-11. The recent generation of bona fide caspase-1-deficient mice [23] allowed us to examine for the first time the role of this caspase independently of concomitant absence of caspase-11. Our data conclusively demonstrate that processing and secretion of IL-1β and IL-18 in response to B. thailandensis infection in vivo or in vitro is completely dependent on caspase-1 but unaffected by absence of caspase-11. The fact that IL-1β secretion is not reduced in absence of caspase-11 also indicates that activation of the NLRP3 inflammasome, which we previously showed exclusively controls IL-1β and IL-18 secretion in response to Burkholderia species infection [12, 13], does not occur as a consequence of caspase-11-mediated pyroptosis and potassium efflux, as in other circumstances [33]. Our results also show that pyroptosis of macrophages infected with B. thailandensis or B. pseudomallei and restriction of intracellular bacteria replication is primarily mediated by caspase-1 with minor involvement of caspase-11. Previous works have attributed a more prominent role to caspase-11 in the pyroptosis of B. thailandensis-infected BMM [15]. It should be noted that those studies relied on Asc-/-Nlrc4-/- cells or Casp1-/-/Casp11-/- cells reconstituted with transgenic human caspase-4 as proxy of bone fide caspase-1 deficient cells, two models that may not faithfully represent caspase-1 absence. Our results also indicate that experimental variables, such as the length of IFNγ priming, may lead to discordant conclusions regarding the involvement of caspase-11 in the pyroptosis of myeloid cells. In fact, it has been proposed that caspase-11 may function as a back-up mechanism to trigger pyroptosis in situations where caspase-1 may be inactivated [34].
The most important result of our study was obtained through bone marrow adoptive transfer experiments and the analysis of the role of caspase-11 in epithelial cell. Our data show that while cell death in B. thailandensis-infected macrophages occurred primarily through caspase-1, with caspase-11-dependent pathway playing a secondary role, pyroptosis of lung epithelial cells was exclusively dependent on caspase-11 and efficiently restricted intracellular B. thailandensis replication in these cells. Interestingly, lung epithelial cells do not express canonical inflammasome components and therefore depend exclusively on caspase-11 for induction of pyroptosis. It is surprising to observe that Casp1-/- mice that are unable to release mature IL-18/IL-1β or trigger pyroptosis in myeloid cells appear as susceptible (if not more, Fig 1B) as Casp11-/- mice that are sufficient for both functions. At face value, this result would attribute equal importance to pyroptosis triggered by caspase-11 in epithelial cells and to that triggered by caspase-1 in myeloid cells plus IL-18 production, a notion previously underappreciated. Thus, caspase-11 dependent pyroptosis of infected lung epithelial cells may be the main protective mechanism triggered by the non-canonical inflammasome in melioidosis. The non-canonical inflammasome was recently shown to restrict S. typhimurium replication in intestinal epithelial cells [35]. Caspase-11 has also been shown to control pyroptosis of endothelial cells during endotoxemia-induced lung injury [36]. Thus, it is conceivable that activation of caspase-11 in cell types other than myeloid or epithelial cells may also play a protective role in melioidosis, an issue we will examine in future studies.
Extending our previous work, we show here that IL-18’s protective action in melioidosis is exclusively dependent on its ability to induce IFNγ. This is an important result because IL-18, in addition to being a strong inducer of IFNγ, also has many other activities. In fact, it has been shown that IL-18 can protect from Streptococcal infections independently of IFNγ [37]. The fact that IFNγ appeared to be indispensable to survive B. thailandensis infection prompted us to investigate the downstream effector mechanisms triggered by IFNγ and responsible for the protection. We concentrated on caspase-11 and the NADPH oxidase because both pathways were already known to provide protection from B. thailandensis infection and because both are IFNγ-inducible, though it was unclear which one contributed more prominently to the IFNγ protective effect. Our data show that in vitro and in vivo the protective action of IFNγ is dependent on production of ROS through the NADPH oxidase system while caspase-11 was dispensable. A previous study has concluded that IFNγ primes caspase-11 in vivo to protect from melioidosis [15]. Although that study ruled out contributions from iNOS and GBP encoded on chromosome 3, the role of NADPH oxidase was not examined. Moreover, that study used a strain of B. thailandensis that has been passaged into Casp1-/-/Casp11-/- mice to acquire higher virulence and used the intraperitoneal infection route, rather than the intranasal one, as in our study. Although it is clear that caspase-11 priming is a necessary step for the function of this molecule, it should be pointed out that several inflammatory stimuli, including TLR agonists produced by B. thailandensis, can prime caspase-11 as efficiently as IFNγ. Interestingly, it was shown that human caspase-4 does not require IFNγ priming in vivo [15].
The results presented here also indicate that caspase-11 may control recruitment of neutrophils to the infected lung. A similar observation has been previously reported during infection with K. pneumoniae [24]. The reason for the impaired inflammatory response of Casp11-/- mice is unclear and actively pursued in our lab. Preliminary analysis failed to detect defective production of the main neutrophil-specific chemotactic factors. It is conceivable that chemotactic alarmins released by pyroptotic epithelial cells may be the missing factor in Casp11-/- infected mice. Whether neutrophils are effective against Burkholderia species is an unresolved issue. We and others have shown that neutrophils are not very effective against this bacterium and that excessive neutrophil recruitment to the infected lung becomes deleterious due to tissue damage caused by release of neutrophil elastase [12, 13, 38]. For these reasons, we think it is unlikely that the high susceptibility to melioidosis of Casp11-/- mice is primarily due to the observed defective neutrophil recruitment.
In conclusion, our results identify non-redundant mechanisms activated by the canonical and the non-canonical inflammasomes that confer host protection in melioidosis: Caspase-1-dependent activation of the IL-18-IFNγ-NADPH oxidase axis and pyroptosis in myeloid cells and caspase-11-dependent pyroptosis of infected lung epithelial cells.
All the animal experiments described in the present study were conducted in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All animal studies were conducted under protocols approved by the Rosalind Franklin University of Medicine and Science Institutional Animal Care and Use Committee (IACUC #B14-17). All efforts were made to minimize suffering and ensure the highest ethical and humane standards.
C57BL/6J, B6.SLJ, Il18-/-, Casp1-/-/Casp11-/-, Cybb-/-, and Ifngr1-/- mice were purchased from Jackson lab. Casp11-/- mice were provided by Vishva Dixit (Genentech) and Casp1-/- mice by Mohamed Lamkanfi (VIB Belgium). All mouse strains were on C57BL/6J genetic background and were bred under specific pathogen-free conditions in the RFUMS animal facility. Age-(8–12 weeks old) and sex-matched animals were used in all experiments. Experimental groups were composed of at least 5 mice, unless stated otherwise.
B. thailandensis E64 was obtained from ATCC. B. pseudomallei K96243 and 390b are clinical virulent isolates. Bacteria were grown in Luria broth to mid-logarithmic phase, their titer was determined by plating serial dilutions on LB agar, and stocks were maintained frozen at -80°C in 20% glycerol. For mice infections, frozen stocks were diluted in sterile PBS to the desired titer. Mice were anesthetized using isoflurane and the infectious doses were applied to the nares in 50 μl total volume PBS. Recombinant murine IL-18 (MBL, Nagoya, Japan) was delivered intranasally (1 μg) 6 hours prior to bacterial infection. Two additional IL-18 treatments (1 μg/each) were administered by intraperitoneal injections at 12–15 hours intervals before euthanasia. Recombinant murine IFNγ (Pepro Tech, NJ, USA) was administered by intraperitoneal injections (2 μg) once daily for two days. In other experiments, 1 μg IFNγ was administered by intraperitoneal injections once daily for two days in the presence or absence of 10 mg N-acety-L-cysteine (NAC, Sigma) delivered at 12–15 hours intervals for 2 days. All cytokines were diluted to desired concentrations with PBS and PBS alone was applied as control.
Organs aseptically collected were weighted and homogenized in 1 ml PBS. Serial dilutions were plated on LB agar plates containing Streptomycin (100 μg/ml) using the Eddy Jet Spiral Plater (Neutec). Bacterial colonies were counted 24 hours later using the Flash & Grow Automated Bacterial Colony Counter (Neutec).
BALF were collected from euthanized mice by intratracheal injection and aspiration of 1 ml PBS. Cytokines levels in tissue culture conditioned supernatants, BALF, or sera were measured by ELISA using the following kits: MCP-1, IFNγ, TNFα, KC, IL-1α, IL-1β, IL-6 (eBioscience), and IL-18 (MBL Nagoya, Japan).
Cells obtained from BALF were counted and stained with anti-CD11b, anti-CD11c, anti-F4/80, anti-Ly6G, anti-NK1.1 and acquired with a LSRII BD flow cytometer. For reactive oxygen species (ROS) measurement, BALF were collected and immediately spun down at 300 × g for 10 minutes to collect cells. Cells were loaded with 7 μM freshly prepared 2',7'-dichlorodihydrofluorescein diacetate, H2DCFDA (Molecular Probes) in PBS at 37 °C for 30 minutes. Cells were stained with anti-CD11b, anti-CD11c, anti-Ly6G, and anti-F4/80 for 10 minutes, washed twice with PBS, and immediately acquired using a LSRII BD flow cytometer. The ROS level was assessed by MFI of DCF (an oxidized product of H2DCFDA) using FITC-channel. Data was analyzed using FlowJo (TreeStar, OR, USA) software.
Cell lysates were separated by SDS-PAGE, transferred to PVDF membranes, and probed with anti-Caspase-11 antibody (Abcam, ab180673), anti-β-Actin antibody (Cell signaling, 4967), anti-Cyclophilin B antibody (Abcam, ab178397). HRP-conjugated anti-Rabbit IgG antibody (Sigma, A0545) was used as secondary antibody. Immunoblots were developed using ECL method and exposed to X-ray film.
Release of LDH in tissue culture media, a reflection of pyroptosis, was measured using the Roche Cytotoxicity Detection Kit (Roche Applied Science, 11644793001). BMM or TC-1 cells were plated in 48-well plates. Bacteria were added to the cell culture and the plates were centrifuged at 300x g for 10 minutes to maximize and synchronize infection and incubated for 30 minutes (BMM) or 2 hours (TC-1) at 37°C. Cells were washed with PBS to remove extracellular bacteria and medium containing kanamycin and gentamicin (200 μg/ml each) was added to inhibit extracellular bacteria growth. Media were collected at 4 and 8 hours post infection for LDH measurement. Cells were lysed in PBS-2% saponin-15% BSA and serial dilutions of the lysates were plated on LB agar plates containing streptomycin (100 μg/ml). For real time cell culture infection with B. pseudomallei, BMM infected with strain 390B or K96193 (MOI 10) in black 96-well plates were incubated in a 37° C, humidified, 5% CO2 atmosphere IVIS Spectrum camera system. Images were captured every 10 min for 10 hr using capture settings of 1 min with medium binning. Grid ROI measurements of Total Flux (p/s) per well were extracted for plotting luminescence of viable bacteria as a function of infection time. All work with B. psuedomallei was performed under biosafety level-3 (BSL3/ABSL3) containment according to policies and standard operating procedures approved via the University of Louisville Committee on Biocontainment and Restricted Entities, The University of Louisville has been approved for select agent work by the Centers for Disease Control and Prevention.
Bone marrow from 8-weeks old B6.SJL (CD45.1) mice or Casp-11-/- mice (CD45.2) was harvested and 106 bone marrow cells were injected intravenously into lethally irradiated (1040 rad) B6.SJL or Casp11-/- mice (8-weeks of age). Chimeric mice were infected five weeks later. Peripheral blood, bone marrow cells, and splenocytes were stained with anti-CD45.1 and anti-CD45.2 (Biolegend) and analyzed by flow cytometry to confirm the efficiency of bone marrow reconstitution.
TC-1 were kindly provided by Thomas Kawula (Washington State University) and grown in RPMI1640-10% FCS. To target caspase-11 gene, TC-1 cells were transfected using Effectene reagent (Quiagen) with Caspase-11 CRISPR/CAS9 KO plasmid and caspase-11 HDR plasmid (Santa Cruz sc-419462 and sc-419462-HDR). TC-1 cells were also transfected with Control CRISPR/CAS9 plasmid (sc-418922) as control. Cells positive for GFP and RFP expression were sorted using BD FACSAria II cell sorter and single cell clones isolated. Expression of caspase-11 in different clones was measured by RT-PCR and immunoblot. One clone was selected that lacked caspase-11 protein expression and produced an aberrant Casp4 mRNA that lacked exons 3, 4, and 5 (S6 Fig).
TC-1 cells were seeded into 10 cm tissue culture dishes and infected when confluent with B. thailandensis or the bsaZ mutant (MOI 500) in a final volume 5 ml in the presence or absence of 100 ng/ml IFNγ. Three hours later, cell monolayers were extensively washed with D-PBS to remove extracellular bacteria and incubated for another 4 hours in medium containing gentamicin and kanamycin (200μg/mL). Biotin-VAD-FMK (15 μM, Santa Cruz) was added to the culture medium one hour before lysing cells in RIPA buffer. Active caspase-11 was pulled down by incubation overnight at 4 °C with streptavidin agarose (Sigma Aldrich) and analyzed by Western Blot with rabbit anti-caspase-11 antibody (Abcam, ab180673)
Mice infected with B. thailandensis (5x105 CFU) for 48 hours were administered intranasally with the Image-iT DEAD Green viability stain (Invitrogen, 1 nmole in 50 μl saline) and euthanized 30 minutes later. Lung were perfused and fixed in 4% paraformaldehyde/PBS and embedded in paraffin. Four-micron sections were stained with anti-EpCam antibody (Abcam, ab213500) followed with Alexa Fluor-647 to label Clara cells and Alveolar Type II pneumocytes and visualized with a Nikon Eclipse 80i Microscope equipped with photometrics coolsnap ES2 imaging system. For quantification of dead cells, up to 400 EpCAM-positive cells were counted in six random fields and scored for nuclear positivity to the green fluorescent viability stain.
All data were expressed as mean ± S.D. Survival curves were compared using the log rank Kaplan-Meier test. Mann-Whitney U test, One-way ANOVA Tukey Post-test, or unpaired t-test were used for analysis of the rest of data as specified in the figure legends. Significance was set at p<0.05.
|
10.1371/journal.pcbi.1004350 | An Evolutionary Approach for Identifying Driver Mutations in Colorectal Cancer | The traditional view of cancer as a genetic disease that can successfully be treated with drugs targeting mutant onco-proteins has motivated whole-genome sequencing efforts in many human cancer types. However, only a subset of mutations found within the genomic landscape of cancer is likely to provide a fitness advantage to the cell. Distinguishing such “driver” mutations from innocuous “passenger” events is critical for prioritizing the validation of candidate mutations in disease-relevant models. We design a novel statistical index, called the Hitchhiking Index, which reflects the probability that any observed candidate gene is a passenger alteration, given the frequency of alterations in a cross-sectional cancer sample set, and apply it to a mutational data set in colorectal cancer. Our methodology is based upon a population dynamics model of mutation accumulation and selection in colorectal tissue prior to cancer initiation as well as during tumorigenesis. This methodology can be used to aid in the prioritization of candidate mutations for functional validation and contributes to the process of drug discovery.
| Evolutionary dynamic models have been intensively studied to elucidate the process of tumorigenesis. One key aspect of studying tumorigenesis is to distinguish the “driver” mutations providing a fitness advantage to cancer cells against neutral “passenger” or “hitchhiking” mutations. Many statistical models to address this question have been developed. Evolutionary models, however, add another layer of complexity by taking into account the process of mutation accumulation and selection within the tissue. Here we present a novel approach combining both statistical and evolutionary thinking to identify driver mutations in cancer genomes using cross-sectional mutation data. Our method considers the process of mutation accumulation and selection before and during colorectal cancer initiation. This work demonstrates the importance of using evolutionary population dynamic models to study driver events of tumorigenesis.
| Cancer cells often harbor hundreds to thousands of genetic changes [1–3]. Many of those changes represent neutral variation that does not influence cancer development; such mutations are called passenger or hitchhiking mutations [1, 4, 5]. A few alterations, however, are essential for driving tumorigenesis. Those changes are known as driver mutations and increase the reproductive fitness of the cancer cell [6–8]. The identification of such mutations is of crucial importance for drug discovery because they represent promising targets for therapeutic intervention.
There is a growing literature of mathematical and statistical approaches to this question [9–18]. In particular, several recent contributions utilized evolutionary models. Tomasetti et al [18] investigated a multiphase model of cancer initiation and tumor growth and studied the mutational burden of tumors across various tissue types. They found that the mean number of somatic mutations in tumors of self-renewing tissues is correlated with patient age at diagnosis. In their model, the time of tumor initiation is chosen so the results fit to incidence data. Another evolutionary modeling framework was considered by Bozic et al [10]. In this work, the mutation accumulation process during the tumor growth (post-initiation) phase was modeled. The authors proposed a formula relating the number of driver mutations to the total number of mutations in the tumor, and applied this methodology to experimental data to infer the selective advantage conferred by typical somatic mutations. This model was based on the assumption that each driver mutation leads to the same selective advantage over the parent cell fitness. Finally, McFarlane et al [13] argued that passenger mutations can also accumulate albeit weak deleterious effects, eventually resulting in oncogenic phenomena.
Adopting not evolutionary, but data-driven approaches, statisticians and computational biologists have been successful at designing methodology to distinguish driver from passenger mutations. For example, MuSiC [19] compares the mutation rates of genes against the background mutation rates using both the Neymanian likelihood ratio test and the Fisherian combined p-value criterion. Youn and Simon [20, 21] further incorporated heterogenous functional impacts of mutations at different nucleotide positions—for instance, missense mutations are considered have less impact than frameshift indels. Later on, using a more ad hoc approach, Vogelstein et al. [22] developed a method based on the patterns of mutation frequencies in each gene: they required that > 20% of mutations in a putative cancer gene are located at recurrent positions and are missense, and that > 20% of mutations in a putative tumor suppressor genes are inactivating. MutSigCV [9], improved from the original MutSig method [1], corrected for population and genomic heterogeneity of mutation rates using a high-dimensional approach to decrease the false discovery rate when calling driver mutations based on mutation frequency. Similarly, DrGaP [23] took into account the length of protein-coding regions, transcript isoforms, variation in mutation types, different background mutation rates, redundancy of the genetic code and others and used a likelihood ratio test to obtain significance levels. Moreover, Multi-Dendrix, DriverNet, MuSiC, and MEMo [12, 17, 19, 24–26] were also developed to identify driver pathways using network-based approaches. An integrated meta-analysis using multiple methods can be accessed through DriverDB [27].
Here we describe a computational approach designed to identify alterations that act as drivers during tumorigenesis. We first designed a mathematical model of the evolutionary processes of mutation accumulation both in healthy tissue during the phase prior to tumor initiation, as well as during the clonal expansion phase of the tumor. One novel aspect of our model is the inclusion of flexible mutational fitness distributions during both phases of mutation accumulation; each mutation may confer a random effect on the reproductive fitness of a cell, drawn from a fitness distribution. We tuned this model to specifically describe disease progression in colorectal cancer by combining literature-based estimates of biological parameters with epidemiological data on the incidence of colorectal cancer as well as pre-cursor conditions. We then sought to identify driver mutations by considering a hypothetical neutral mutation at any locus in the genome and following its progression through our evolutionary framework to determine the likelihood of observing this mutation at a detectable frequency in a significant portion of patients with this cancer type. This quantity is called the “Hitchhiking Index” and can be used to reject the hypothesis that any particular candidate mutation is neutral, thus identifying potential driver mutations. Since the likelihood of acquiring each candidate gene mutation may vary across the genome, we stratified mutation rates into three large groups: those with high, intermediate and low mutation rates [9]. By helping to identify mutations that are potential driver mutations during tumorigenesis, this methodology can be used to aid in the prioritization of candidate mutations for functional validation.
Our computational approach is based upon the evolutionary dynamics of the accumulation of driver and passenger mutations in a population of cells (Fig 1.(a)). There are two phases (Fig 1.(b)): a pre-initiation phase and a clonal expansion phase. During the pre-initiation phase, the first driver mutation has not yet emerged and the population is maintained at a homeostatic cell number. Cells proliferate according to a stochastic process: at each time step, a cell is chosen at random proportional to its fitness to divide, and its offspring replaces another randomly chosen cell. During each cell division, a mutation may emerge with probability u. Each mutation confers an additive change to the fitness of the daughter cell; this additive change is chosen from a mutational fitness distribution which is approximated discretely by a mutational kernel M1. The survival of the resulting mutant clone is dependent on its relative fitness, as well as any subsequent mutations it may accumulate. This phase of the methodology is designed to model the behavior of stem cells within a crypt of the colon.
If a cell in the population has accumulated a sufficiently large fitness to counter the homeostatic mechanisms of the compartment, then the second phase of clonal expansion begins (Fig 1.(b)). The cell number in the tumor is now described by a multi-type stochastic branching process: at each time step, a cell is chosen proportional to fitness to divide (possibly with mutation) or chosen at random to die. This initiated cancer cell carries one or more driver mutations that confer a larger growth than death rate; the population of cells thus grows on average exponentially. Each time a cell in this population divides, a mutation may again arise with probability u. Once again, mutations confer an additive change to the fitness of the daughter cell; this additive change is chosen from a mutational fitness distribution which is approximated discretely by a mutational kernel M2. This mutational kernel is not, in general, the same as the mutational kernel from the pre-initiation phase (M1), since there is no reason to assume that the mutational fitness distribution in a normal compartment of healthy cells during the pre-initiation phase should be the same as the distribution in a rapidly expanding cancer clone. However, we utilize the same family of mutation kernels, noting that the shape parameters may differ between phases. As mutations accumulate and undergo clonal expansion in the model, the number of cell types grows and the tumor population becomes more heterogeneous. Each additional driver mutation increases the fitness of the cell lineage, such that the rate of expansion accelerates (Fig 1.(b)). Clonal growth continues until the tumor reaches its detection size, Nd (Fig 1.(b)).
We then utilize this underlying evolutionary model to determine the probability q that a particular candidate gene is found in a detectable frequency of the tumor, conditioned on the null hypothesis that alterations of this gene are selectively neutral. This probability q is then used to derive the probability of obtaining the observed frequency of alterations in the sample set, given the null hypothesis. More specifically, when sampling tumors from Y patients, we calculate the “Hitchhiking Index”, which specifies the probability of detecting a certain mutation in at least α% of Y patients, as
H = ∑ k = ⌈ α Y ⌉ Y P ( X = k ) ,
where X is binomially distributed as Binom(Y,q) and ⌈αY⌉ is the smallest integer greater than αY. This index provides a tool for rejecting the null hypothesis. If for example H < 0.001 for a given observed α and Y, we can reject the assumption of neutrality of a mutations in a given gene of interest.
The parameters of this model (e.g. cell turnover rates and mutation kernels in both phases, population size during homeostasis and at detection) must be specifically tuned for each cancer and sample type. In addition to the evolutionary processes within the tissue, the value of the Hitchhiking Index also depends upon the relative mutation rate of each candidate gene, detection sensitivity of the sample set, and the observed alteration frequency in the sample set. Note that the Hitchhiking Index provides a method for rejecting the null assumption of neutrality; however, failure to reject this assumption does not necessarily imply that the gene is neutral.
Let us provide further details of underlying mathematical framework that models the pre-initiation phase, which was first introduced in [28]. During this phase we consider a small population of cells of constant size N which describes a homeostatic compartment of cells at risk for accumulating mutations leading to cancer initiation, described by a multi-type Moran process [29]. Each cell on average divides every D days. Thus, at rate N/D (i.e. time between events are i.i.d exponential random variables with mean D/N), the process undergoes division events. During each event, a cell is chosen at random to die, and proportional to its fitness an individual is chosen to reproduce. Specifically, if there is a single cell with fitness s and N−1 cells with fitness 1, then the cell with fitness s is chosen to reproduce with probability s/(s + (N−1)). During each cell division event, an (epi)genetic alteration may occur with probability u < < 1; thus alterations arise in the compartment of cells at rate Nu.
The fitness effects of individual alterations are random variates drawn from a mutational fitness landscape governed by the mutation kernel M1(⋅, ⋅). Here M1(x,y) represents the probability that a cell with fitness x produces a daughter cell with fitness y (i.e. M ( x , y ) = f ψ x ( y - x )). If y > x, then the fitness of the daughter cell is advantageous as compared to the fitness of its parent cell; if y < x, it is disadvantageous, and if y = x, it is neutral. The type space is discretized into fitness bins to aid in computational tractability, and thus the kernel M is a finite-state transition matrix. In this work we utilize a general family of mutation kernels that have exponentially decaying tails on the positive and negative sides with shape parameters α and β, respectively. Note that α = β = 0 represents the uniform distribution case, and for α,β > 0 the mutational fitness distribution has a mode at 0 (neutral mutations). This process continues for as long as the fitness of all cells is within the homeostatic range [a,b] for a = 1−1/N and b = 1+1/N; these values were chosen since they signify the boundaries for neutral evolution [30]. Also note all sample paths will result in cancer initiation prior to death.
In the event that a cell with a sufficiently large fitness emerges, it can escape homeostatic mechanisms in the compartment and initiate clonal expansion. It has been shown that when 3Nu(logN + γ) ≪ 1, where γ is the Euler-Mascheroni constant, the time between mutational events is much larger than the time it takes for a mutation to take over or go extinct in a population of cells (see, e.g. [28]). In the application of colorectal cancer considered in this work, this condition holds, supported from the parameterization discussed in the section Model parameterization. Therefore, on the timescale of interest, the population moves between various homogenous states and we can approximate this process by a Markov process Z(⋅), where Z(t) represents the fitness of the homogeneous compartment at time t. The process Z jumps whenever a cell harboring a novel non-neutral mutation reaches fixation in the compartment, and takes values in the space of all possible fitness values dictated by the fitness landscape. This process closely approximates the behavior of cellular fitness values in a small compartment for the vast majority of time.
Given that a mutation of fitness y arises in a population of cells with fitness x, the probability that the mutation takes over the population is given by rx,y = (1−x/y)/(1−(x/y)N). By symmetry we have rx,x = 1/N. With this information, we define the intensity matrix for the Markov process Z, denoted by Q = Q(x,y),
Q ( x , y ) = { u r x , y M 1 ( x , y ) N / D , y < b u N M 1 ( x , y ) / D , y ≥ b , .
Mutational events transforming a cell with fitness x to a cell with fitness y occur at rate uM(x,y)N/D, and of those a fraction rx,y reach fixation (i.e. 100% frequency) in the entire population. At the lower end of the fitness range a, there is a reflecting boundary such that any fitness below a is immediately replaced by a cell with fitness a; this behavior is implemented in the mutation kernel M. By definition, we have Q(x,x) = −∑y ≠ x Q(x,y), and the states x ≥ b are absorbing so Q(x,y) = 0 for all y and x ≥ b. The summation is over all y in the discrete fitness space.
The process Z represents the dynamics of the fitness of a healthy compartment of cells over time; this process is then conditioned upon the event that cancer initiation occurs during a human lifetime [28]. This is achieved by creating an additional lifetime process, L, that is run simultaneously with Z in a second dimension. The lifetime process has a single absorbing state representing death of the patient, and transition rates between intermediate stages are tuned using mortality statistics in the United States. Details of this tuning process can be found in [28]. We are then interested in the joint process (Z,L) conditioned on Z hitting its absorbing state (fitness greater than b) prior to L hitting its absorbing state (death). This set of sample paths describes the cancer initiation paths which may lead to tumor diagnosis prior to death.
To consider the fate of a particular candidate mutation ‘A’ in our data set, suppose that this mutation arises with probability u0 per cell division. Under our null hypothesis, we assume mutation A is neutral and confers no selective advantage/disadvantage; thus its presence does not alter the evolutionary outcome of the sample path. Conditioned on initiation prior to death, we compute the probability of mutation A arising in the initiating cancer cell. To do this, we analyze the amount of time the (Z,L) process spends in each state of the two-dimensional state space. Details of this derivation are provided in the Methods.
Once a cell in the pre-initiation phase has acquired a fitness value greater than b, the second clonal expansion phase of the model commences. This phase is modeled by a continuous time multi-type birth and death process, initiated by the cell from the pre-initiation phase that has accumulated a sufficiently large fitness to break free from the Moran process and initiate clonal expansion. The initial branching process has birth rate b and death rate d, where b > d. During this expansion phase, each time a cell divides, it has a probability u of mutating and selecting a new random birth rate from a fitness distribution, and probability u0 of obtaining the specific candidate mutation A without any change in birth rate. The birth rate of a mutated daughter cell is the parental fitness plus a random variable selected from a distribution specified by the mutation kernel M2. The type space is once again discretized into fitness bins to aid in computational tractability, and the kernel M2 is a finite-state transition matrix, accordingly.
We then aim to determine the probability that mutation A is present in a significant fraction of the final population size at detection, Nd, conditional upon the event that the expansion process was initiated and reached detection size during a human lifetime. To study this event we first utilized analytical calculations to determine the probability of tumor initiation prior to death, in the first phase of the model, by solving the linear system described in [28] using a biconjugate gradient stabilized method. We next performed event-driven Monte-Carlo simulations of the two-dimensional Markov process describing the fitness of the crypt and the lifetime state conditional on initiation prior to death [28]. To account for the fact that sample paths of interest may be rare, the probability of initiation calculated in the previous stem was used to perform a Doob h-transform of the process (Z,L) conditioned on initiation prior to death. Thus we only simulated sample paths that lead to initiation, saving computational time.
At the time of initiation for each sample path, the time of initiation as well as the status of mutation A in the initiating cell is recorded. These initial conditions are used to begin a stochastic simulation of the clonal expansion process. During each event in the simulation, a cell is chosen to divide based on its relative fitness and abundance. During each cell division, a mutation occurs with probability u and the outcome of that mutation is selected from the mutation kernel Mc. The tagged mutation A arises with probability u0 ≪ u. Naturally, a certain fraction of paths in the expansion phase die out at early times due to stochastic fluctuations. For the remaining paths, the simulation is halted when the total population hits the detection size Nd, and then the abundance of mutation A in the total population of tumor cells is recorded.
To apply this framework to analyze genomic data from any specific cancer type, the main challenge is to determine the probability q that a particular mutation of interest, mutation A, arises during the clonal expansion phase and eventually makes up a significant fraction of the final population size, M. This outcome can occur via two possible scenarios: (1) mutation A arises and reaches 100% frequency during the pre-initiation phase, so that all cells of the resulting tumor have mutation A (Fig 1.(c)), and (2) mutation A is not present in the initiating cell but arises during clonal expansion (Fig 1.(d)). Recognizing these two mutually exclusive possibilities suggests an interesting question: which is the more likely path out of these two scenarios? The answer to this question depends on the parameters of the evolutionary model, which might vary from cancer type to cancer type. We investigated colorectal cancer in particular, through the approach outlined in the following.
In this work we have developed a novel methodology to identify driver mutations from cross-sectional tumor sequencing data, based on an evolutionary model of tumorigenesis. We developed the ‘Hitchhiking Index,’ which represents the probability of observing alterations in a particular gene in a certain fraction of the patient sample set, under the null assumption that the gene is not a cancer driver. This index takes into account the impact of a number of important parameters on the statistical power of the conclusion: the sample detection threshold (sensitivity of the sequencing method), patient sample size, and variable mutation rates across the genome. The underlying evolutionary model is designed and parameterized for colorectal tumorigenesis, but can be generalized to other cancer types. Here we have not incorporated full pathway information of the gene of interest, but the model can easily be adapted to group genes together into pathways and analyze selection dynamics at the pathway level instead of at the individual gene level. The Hitchhiking index is calculated for any particular candidate gene, using the observed patient sequencing data, and can be used to identify candidate genes as potential drivers.
We applied our methodology to analyze TCGA data for colorectal cancer, considering heterogeneous mutation rates measured per cell division which are inferred from baseline mutation rate estimates and relative changes from that rate across the genome as determined by MutSigCV [9]. We built upon MutSigCV by incorporating heterogeneous mutation rate estimations into our model: (1) We specify an underlying evolutionary dynamic model to describe the processes generating mutations to calculate the probability of a mutation being a driver event; and (2) by controlling for the bias introduced by DNA replication timing, gene expression and higher-order chromatin structure, we infer the relative mutation rate per cell division compared to the cross-sectional mutation rate. Remarkably, we found that any gene that is mutated in at least 10% of cells in the tumor is most likely to have arisen prior to clonal expansion of an initiated cell clone. Utilizing the Hitchhiking Index analysis, we obtained a list of 43 genes identified as potential drivers. In comparison to a recent analysis utilizing MutSigCV [9], our methodology identified other colorectal cancer related genes such as COL12A1, MLL2, FAT4m and ARID1A. Recent studies support the crucial role of these genes in the development of colon cancer [36–39].
One caveat to our approach is that it is unclear how to choose the threshold Hitchhiking Index value; similar to a statistical p-value, the choice of threshold at which to reject the null hypothesis is largely a matter of choice. Since this index depends upon the sample size and detection sensitivity of the method, it would be difficult to compare absolute values of the Hitchhiking Index across different sample sets. Note that if the Hitchhiking Index for a particular gene is above the rejection threshold, we do not conclude that the gene is necessarily a passenger—our methodology provides only the probability of observing the data, conditioned on the assumption of neutrality. Also, note we used the same u and u0 in both phases. While the precise values are unknown, it is possible that the mutation rates can be higher in the latter phase, since it is possible that the initiating mutation causes an increase in the mutation rate itself, e.g., a mutation that reduces the effectiveness of DNA repair. This scenario has the potential to alter the creation rate of passenger mutations and will be the topic of future investigations. As such, our current work excludes consideration of colorectal cancers with microsatellite instability, a deficiency of the mismatch repair (MMR) pathway that leads to increased point mutation rates across the genome. Additionally, hereditary forms of colorectal cancer are also not explicitly considered and will be investigated in future work. Furthermore, the presented model does not include all possibilities for alternative initiation mechanisms. Such a model would not be very useful since it would address mutually exclusive evolutionary trajectories; instead, we have presented one possible model of the evolutionary process leading to tumorigenesis. A novel feature of this model is the formulation of the initiation event as an accumulation of a sufficiently large fitness advantage in the initiating cell through a flexible series of mutational events rather than a specific set or number of hits. Because of this flexibility, a multitude of mutational pathways can lead to initiation in our model, and in particular it can be used to consider the situation in which each of these events is disruption of a particular pathway. The approach can also modified to incorporate the more traditional view that a specific set of hits is required to initiate cancer (by specifying instead a discrete distribution for the mutational fitness landscape) and making this landscape dependent on the current mutation status.
Here, we have utilized a specific, tunable evolutionary model of mutation accumulation in cancer to develop a novel statistical test for identifying driver mutations from cross-sectional genomic data of cancer sample sets. We have opted for a somewhat more flexible approach to modeling the process of mutation accumulation and initiation. For instance, we have considered mutational heterogeneity in a coarse manner, by grouping genes into three different categories with different baseline mutation rates per cell division. A more complete model could in principle use different baseline mutation rates for each categories of DNA replication timing, gene expressions and other genomic features, even including the difference between transitions and transversions [15, 40]. In contrast to the model by Tomasetti et al [18], in which all mutations prior to initiation are considered to be selectively neutral and the time of tumor initiation is set by epidemiological data, here we have assumed that mutations conferring random fitness advantages can arise during the constant population size phase, and that tumor initiation occurs as a result of accumulating sufficiently many advantageous mutations to escape homeostasis. Consequently, in our model the timing of cancer initiation is random, and correlated with the process of mutation accumulation. We have carried the same modeling framework through to the tumor growth phase, in which cells may accumulate mutations conferring a spectrum of fitness changes. Consequently, in contrast to the model by Bozic et al [10], we assume that driver mutations may be variable in number and lead to variable fitness effects and that tumors may alternatively have many drivers with small selective advantage or a few drivers with large selective advantages. These differing modeling choices reflect a rich set of hypotheses about the underlying evolutionary dynamics of mutation accumulation in cancer; more modeling and experimental effort is needed to investigate the perspectives and relative strengths of these and many other models. Several important conclusions, however, seem robust: first, mathematical analyses of the evolutionary processes in cancer suggest that the majority of mutations found in tumor sequencing efforts arise prior to cancer initiation; and second, mathematical frameworks of evolution and mutation accumulation in cancer can be exploited to extract important biological information from genomic sequencing data.
Let m(t) denote the number of cells carrying mutation A that are present in the compartment at time t. We define the stopping times τ = inf{t ≥ 0 : Z(t) ≥ b} and σ = inf{t ≥ 0 : L(t) = d}. We are interested in finding
μ A ( x , r ) = P ( x , r ) [ m ( τ ) > 0 | τ < σ ] , (1)
where μA(x,r) represents the probability that mutation A is present in the compartment at the time of initiation starting from state (x,r). Between jumps of the two-dimensional process (Z,L), a random number of neutral mutations can reach fixation within the compartment of cells. Let Tj and Tj+1 be the jump times of (Z,L), and for simplicity denote (Z(Tj),L(Tj)) = Xj. During the transition from Xj to Xj+1, the compartment can accumulate Yj(Xj) neutral mutations. Define
ρ ( x , r ) = u M 1 ( x , x ) / D u M 1 ( x , x ) / D - Q ( x , x ) - S .
The numerator represents the rate at which neutral mutations which eventually reach fixation arise within the compartment, and the denominator represents the total rate at which fixating mutations arrive and the time process changes. With this definition, Yj(x,r) is distributed like a geometric random variable with
Prob ( Y j ( x , r ) = n ) = η ( x , r ) n ( 1 - η ( x , r ) ) ,
which gives
Prob ( Y j ( x , r ) > 0 ) = ρ ( x , r ) .
By conditioning on the first step we can see that μA(⋅,⋅) satisfies the following equation for each possible fitness x in [a,b],
μ A ( x , r ) = ρ ( x , r ) + ∑ y Q r ( x , y ) μ A ( y , r ) u M 1 ( x , x ) / D - Q ( x , x ) - s - Q x ( r , r + 1 ) μ ( x , r + 1 ) u M 1 ( x , x ) / D - Q ( x , x ) - S ,
where we note that μA(x,r) = 0 for those fitnesses that lie outside of [a,b]. Therefore we can find μA(⋅,⋅) by solving the linear system.
We tested the sensitivity of the model predictions to varying parameters. First, we studied the sensitivity to the parameters for which we have no experimental data-based estimates: the shape parameters of mutational fitness distribution and the birth and death rates of the initiating cell. We also investigated the sensitivity to the background mutation rate u, and in particular studied the impact of an increased mutation rate during the clonal expansion phase. For all of these sensitivity analyses we confined our parameter variation to the ranges in which the model is consistent with the population-level epidemiological data. In particular, we required that there is a significant incidence of aberrant crypt foci over a lifetime (10–90 percent of the population), that the average age at diagnosis is between 70–79 yrs, and that the lifetime risk of colon cancer is around 6–10%.
We first investigated the sensitivity of our model predictions to the shape parameters of the mutational fitness distribution during the pre-initiation phase of the model. Keeping all other parameters constant, we varied α1 and β1. To determine the allowable ranges of these parameters, we first studied the probability of initiation during a lifetime as a function of α1 and β1. Note that this is the probability that a single crypt leads to initiation during a lifetime, and there are 15 million crypts in the average human colon. Thus, in order to ensure that the average probability of aberrant crypt foci is between 10 and 90 percent in the population, we require that the probability of initiation from a single crypt is between 0.7×10−8 and 6×10−8. In Fig 5.a the probability of initiation for a single crypt is plotted for example ranges of parameters α1,β1. The intermediate colors (between dark blue and red) represent admissible initiation probabilities. Thus we see that for any given α1, there is a small range of β1 that can give rise to the correct range of initiation probabilities. We then investigated the compatibility of these α1,β1 combinations with the other incidence data. We found that only α1 in the range 170 ∼ 180 can give rise to the correct overall lifetime cancer incidence rates within the range 6–10%. Furthermore, for α1 = 170, β1 must fall within the narrow range 115 ∼ 120; larger values of β1 result in lifetime incidence above the allowable range, and lower values of β1 result in too low an incidence of aberrant crypt foci in the population Similarly, for α1 = 180, β1 must lie within the range 180 ∼ 185 to match the incidence data constraints.
Next we used the model to determine the sensitivity of the results, as determined by the Hitchhiking Index, to variations to α1 and β1 within these allowable ranges. For example, Fig 5.b shows the number of patients out of 100,000 samples in which cells harboring mutation ‘A’ make up a threshold frequency in the final tumor cell population, for varying α1 and β1. We observed only modest differences in q, which would translate to negligible differences in the Hitchhiking Index. Therefore, within the constraints of matching the observed incidence data, the Hitchhiking Index is robust to varying the shape parameters of the mutational fitness landscape during the carcinogenesis phase.
We also investigated the sensitivity of the results to the shape parameters of the mutational fitness distribution during the clonal expansion phase of the model. Fig 5c. demonstrates the impact of varying α2 and β2 up to 60 percent from the original values on the Hitchhiking Index; we found that the Hitchhiking Index is not particularly sensitive to these parameters.
We then investigated the model’s sensitivity to the growth rates of the first cell initiating clonal expansion (b,d). We varied b first to determine the impact of the net growth rate on the Hitchhiking Index. Variation of this parameter leads to overall lifetime cancer incidence rates that fall outside the range of our incidence data; this suggests that the net growth rate during the clonal expansion phase within our model should not vary significantly from the fitted value. There is, however, the possibility that both b and d vary in such a way that the net growth rate remains conserved, for example if both b and d are increased or decreased by the same amount. These variations might lead to small differences in the model predictions.
|
10.1371/journal.pntd.0005756 | Wetlands, wild Bovidae species richness and sheep density delineate risk of Rift Valley fever outbreaks in the African continent and Arabian Peninsula | Rift Valley fever (RVF) is an emerging, vector-borne viral zoonosis that has significantly impacted public health, livestock health and production, and food security over the last three decades across large regions of the African continent and the Arabian Peninsula. The potential for expansion of RVF outbreaks within and beyond the range of previous occurrence is unknown. Despite many large national and international epidemics, the landscape epidemiology of RVF remains obscure, particularly with respect to the ecological roles of wildlife reservoirs and surface water features. The current investigation modeled RVF risk throughout Africa and the Arabian Peninsula as a function of a suite of biotic and abiotic landscape features using machine learning methods. Intermittent wetland, wild Bovidae species richness and sheep density were associated with increased landscape suitability to RVF outbreaks. These results suggest the role of wildlife hosts and distinct hydrogeographic landscapes in RVF virus circulation and subsequent outbreaks may be underestimated. These results await validation by studies employing a deeper, field-based interrogation of potential wildlife hosts within high risk taxa.
| Rift Valley fever (RVF) is a vector-borne zoonotic disease that imparts a substantial burden to the economy and public health of pastoralist communities across the African continent and Arabian Peninsula. Furthermore, RVF is also an emerging pathogen of growing global concern. Knowledge of the epidemiological and ecological factors that influence the geographic distribution of RVF outbreaks and determine risk for humans and animals is incomplete. The current study examined the distribution of RVF outbreaks from 1998 to 2016 and modeled their occurrence as a function of climate, surface water, land cover, livestock density, wild mammalian species richness, and human migration. The results indicate that wetlands, Bovidae species richness, and sheep density were associated with increased risk of RVF outbreaks. Our findings contribute to improved understanding of the spatial and ecological dynamics of RVF risk with a particular emphasis on the distribution of wetlands and potential wildlife reservoirs in designing RVF surveillance programs.
| Rift Valley fever (RVF) is an emerging, vector-borne viral zoonosis that causes significant morbidity in humans and their livestock. The etiologic agent, Rift Valley fever virus (RVFV), is a Phlebovirus in the Bunyaviridae family and is transmitted by several mosquito species that facilitate viral maintenance (Aedes spp.) or amplification (Culex spp.) [1,2]. Human infections are invariably asymptomatic or mild in early stages, however, severe cases can manifest as hemorrhagic fever or encephalitis [3,4]. Sheep, goats, and cattle experience fetal abortions as a result of RVFV infection, and the disease contributes to substantive economic losses to pastoralist communities during outbreaks [5,6].
Historically, most outbreaks in humans and domestic animals have occurred in the African continent, and in eastern Africa these typically follow periods of excessive rain in poorly draining arid or semi-arid landscapes [7]. The resultant flooding is conducive to the breeding and hatching of infected mosquitoes which transmit the virus to ruminant hosts followed by eventual secondary transmission to other ruminants and humans [5]. In more recent years, RVF has progressively expanded east into the Arabian Peninsula, with outbreaks in Saudi Arabia and Yemen [8].
The epidemiology and infection ecology of RVFV is complex and our knowledge of these incomplete. Several species in two distinct mosquito genera transmit infection. As primary vectors, Aedes mosquitoes maintain RVFV transovarially during dry periods; during these times there are little to no reported human or livestock infections. Aedes mosquito population explosion following wet periods leads to localized transmission to mammalian hosts [9,10]. Following this, Culex mosquitoes can expand (amplify) transmission to more dispersed livestock and human populations distant from the areas of local Aedes transmission [2,11,12]. Once RVFV becomes amplified in livestock, ongoing human infection occurs primarily through zoonotic transmission as a result of direct or indirect contact with animal tissues and body fluids, such as occurs during slaughtering or through performing obstetrical procedures on infected animals. Transmission from mosquitoes that feed on infected animals is also a viable though less important source of human infection [1,13].
While the role of vectors in RVFV infection ecology is well-established, the extent to which wildlife contributes to transmission as possible maintenance or amplification hosts is not well understood. Field investigations suggest that wild ruminants and rodents are the most likely RVFV reservoirs [14]. Nevertheless, data from these field surveys are limited, so definitive mammalian natural reservoirs for RVFV are not described [11,14].
The landscape epidemiology of RVFV is also incomplete with respect to abiotic systems of influence. For example, periods of excessive rain are strongly associated with RVF outbreaks in East Africa [7,15–18], however very little is known regarding the interaction between climate and terrestrial or hydrogeographic profiles in mediating RVF outbreaks [19]. In addition, there has been a lack of attention to land cover characteristics, which have the potential to influence mosquito habitat, sylvan reservoir habitat, and the movement of domestic livestock through the landscape. Finally, anthropogenic influence, such as human migration, may introduce novel, or increase existing, exposures among pastoralist and/or other rural and peri-urban communities [20].
The study sought to expand our current understanding of RVF epidemiology and infection ecology by investigating the role of diverse hydrogeographic features and wild Bovidae and Muridae species richness in delineating the landscape suitability of future outbreaks across the African continent and Arabian Peninsula.
Occurrence data for RVF outbreaks in humans and livestock animals were obtained from the ProMED-mail electronic surveillance system. This surveillance system is maintained by the International Society of Infectious Diseases and provides near real time and archival documentation of formal and informal reports of infectious diseases [21]. The database was searched using the keywords “rift valley fever”, “rift valley fever virus”, “rvf”, and “rvfv”. Only those reports documenting RVF outbreaks in humans or livestock in unique locations were included (i.e. duplicate outbreaks were not included). One hundred and three reports of laboratory confirmed, geolocated outbreaks of RVF in humans and livestock were documented by the ProMED system between January 1, 1998 and August 31, 2016. Google Maps was used to capture the geographic coordinates for each outbreak and cross-checked against Open Street Map. Centroids of the reported outbreak locations were recorded to a spatial resolution of 4km2.
To test our landscape suitability model (see Statistical Analysis section), a second source of RVF outbreak data were obtained from the World Organization for Animal Health (OIE). OIE maintains an official biosurveillance mechanism for RVF in livestock. These data have been archived since 2004 and can be accessed via the World Animal Health Information System (WAHIS) web portal [22]. Reports included the location of each event by place name, the date, type of livestock affected, and the number of infected animals identified. Between January, 2005 and August, 2016 a total of 50 "immediate notification" and subsequent “follow-up” RVFV outbreak reports were submitted to OIE. The geographic coordinates for these events were obtained with Google Maps as above. Outbreaks from ten of these reports could not be located within this coordinate reference system. This left a total 40 OIE reports with 102 unique outbreak occurrences. Twenty-three of the OIE documented outbreaks were also recorded in the ProMED surveillance and therefore were not included in this testing dataset to prevent inflation of model performance. Thus, the final OIE sample of 79 was used for model testing.
Altitude and four Bioclim climate rasters were obtained from the WorldClim Global Climate database and used as climate indicators for this investigation [23]. Aggregate spatio-temporal weather station data between 1950 and 2000 were used to calculate the mean temperature during the hottest and coldest quarters, and the mean precipitation during the wettest and driest quarters, and extracted as 30 arc second (approximately 1 km2) resolution rasters [24].
Vegetation cover was assessed using the MODIS-based Maximum Green Vegetation Fraction (MGVF), which is a data product from the United States Geologic Survey's Land Cover Institute [25]. The MGVF records the percentage of green vegetation cover per pixel as a function of the normalized difference vegetation index at a resolution of 1 km2[26]. Rasters were obtained at two time points, years 2001 and 2010, and the difference between them calculated to determine vegetation loss over this 10 year period. Change in MGVF over this time period was considered a more robust representation of vegetation cover than mean MGVF, and therefore more appropriate in assessing its influence on RVF landscape suitability.
The Global Lakes and Wetlands Database [27] was used to define surface water. This raster was derived from three discrete components. The first two comprised vector data of polygons. Component 1 represented lakes with area ≥ 50 km2 and controlled water reservoirs with volume ≥ 0.5 km3, while component 2 represented all surface water with area ≥ 0.1 km2. The third component combined and rasterized the polygon data from the first two components, while supplementing the wetland data. The final 1 km2 raster based on component three was used here. The surface water categories were: lake, controlled water reservoir, river, freshwater marsh, swamp, coastal wetland, brackish, bog, or intermittent wetland [28]. The surface water types were extracted and new distance rasters created. Distance was calculated in the QGIS geographic information system using the proximity function to produce separate 1 km2 resolution rasters for each water category[29]. Pixel values in these rasters convey the distance in kilometers between a given pixel and the nearest pixel occupied by each unique category of surface water. In this way the models can incorporate a spectrum of proximity to diverse hydrogeography across the metacontinent (see Statistical Analysis section).
Net human population migration was obtained as a 30 arc-second raster from the Socioeconomic Data and Applications Center (SEDAC), which is part of the National Aeronautics and Space Agency's Earth Observing System Data and Information System [30]. This raster describes the net change (increase vs. decrease) in persons per km2 from the period 1990 to 2000 [31].
The global densities of cattle, sheep, and goats were represented as 1 km2 resolution rasters from the Gridded Livestock of the World (GLW) [32]. The GLW also classified ruminant livestock production systems by system (livestock-only, mixed rain fed, and mixed irrigated) and climate regime (Hyper-arid, Arid, Humid, and Temperate/Tropical Highlands) comprising 12 production system categories plus one additional category classified as Urban [33].
Rasters of Bovidae and Muridae species richness at 1 km2 resolution were acquired from the International Union for Conservation of Nature (IUCN) and Center for International Earth Science Information Network (CIESIN)[34].
Finally, all species of Aedes mosquitoes observed across the geographic range of RVF outbreaks were extracted from the Global Biodiversity Information Facility (GBIF)[35]. There were 215 field observations of Aedes mosquitoes geolocated within the African continent and Arabian Peninsula. However, of these 215 mosquito observations, 151 observations recorded the genus only without species designation, while 57 were Ae. africanus, and 7 were Ae. albopictus. As such, there was not sufficient species representation in the GBIF to produce valid models of the ecological niche of Aedes vectors. One generic ecological niche model of Aedes mosquitoes was included in an exploratory analysis, but this contributed very little to the loss function when modeling RVF landscape suitability (see Statistical Analysis section), further suggesting that the mosquito data were insufficient to include in the current investigation. Similarly, this analysis did not include potential Culex amplification vectors as there were too few GBIF specimens across the region and those that were present were of too diverse an ecological and behavioral spectrum to be pooled for analysis.
This study used maximum entropy (Maxent) machine learning to model the landscape suitability of RVF outbreaks in human and livestock hosts across Africa and the Arabian Peninsula at a resolution of 4 km2. In the current study, risk is defined explicitly as the probability of landscape suitability to RVF outbreaks. Machine learning in general, and Maxent in particular, is analytically appealing because a specific model form is not assumed. Instead algorithms create rule-based data partitions that optimize homogeneity between predictors and outcomes [36]. Further, the Maxent machine learning algorithm does not require the locations of RVF outbreak absences which are effectively unknowable [37,38].
The full Maxent model (based on ProMED data) comprised the following landscape features: mean dry quarter precipitation; mean warm and cold quarter temperature; change in vegetation cover; proximity to the surface water features; wild Bovidae and Muridae species richness; cattle, sheep, and goat densities; and net human migration between 1990 and 2000. Correlation between most of the landscape factors acquired for this study was low. However, there were a few exceptions (wet quarter precipitation, ruminant production systems, swamp, and altitude), all of which were correlated with several other landscape factors and provided generally redundant information. Therefore, these factors were dropped from the original 22 predictors acquired from our data sources described above. Ten thousand background points were sampled, weighted according to human population density to adjust for any potential sampling bias in RVF occurrences derived from ProMED. A value of 1.0 was selected for the regularization parameter, to correct for overfitting of the model predictions. The Maxent models were trained using five-fold cross-validation. This approach divides the training set into k = 5 subsets, iteratively fits the model to 4-subset combinations, and then tests against the 5th. Each of the five subsets included approximately 20 RVF outbreaks selected randomly from the total number of available observations in the training dataset (ProMed; n = 103 outbreaks).
Landscape features used in the full Maxent model were ranked according to their permutation importance, which randomly permutes the values of the landscape factors between background and presence points in the training dataset. This is preferred over the direct percent contribution to the loss function because it is non-heuristic and more robust to any residual correlation in assessing the influence of individual features on RVF landscape suitability [37,39].
Finally, as a robust evaluation of prediction error, the trained models were tested against the data obtained from OIE. The difference in model predictions based on the training and testing data was used to assess the model prediction error, which was reported as the area under the curve (AUC). The models were fit using the maxent function (dismo package; v. 0.9–3) setting the distribution to Bernoulli [38,40,41]. All analyses were performed using R statistical software version 3.1.3 [42].
The distributions of RVF outbreaks captured by ProMED and OIE are presented in Fig 1. The clustering of these outbreaks in the Sahel and in eastern and southern African is demonstrated, as is the more recent emergence of RVFV in the Arabian Peninsula.
All landscape features used in the ecological niche modeling are presented separately for the abiotic (climate, vegetation change, and surface water) and biotic (livestock densities, Bovidae and Muridae species richness, and human population migration) features in Figs 2, 3 and 4.
The predicted landscape suitability of the African continent and Arabian Peninsula to RVF outbreaks is presented in Fig 5. High risk landscapes were identified in Mauritania extending eastward into the Sahel, as well as in large portions of Sudan, Kenya, Tanzania, South Africa, Madagascar, and a corridor adjacent to the Red Sea in Saudi Arabia and Yemen. Moderate landscape suitability was predicted for northern parts of the Maghreb, the Horn of Africa and the broader Arabian Peninsula.
The Maxent model identified proximity to intermittent wetlands (permutation importance = 18%), Bovidae species richness (11.7%), sheep density (11%), dry quarter precipitation (10.2%), and proximity to freshwater marsh (9.1%) as the most influential features to RVF landscape suitability (Fig 6). Muridae species richness was not as influential to suitability as Bovidae richness, but was impactful in the model with 8.6% permutation importance.
Response curves for these features and RVF outbreak risk are presented in S1 Fig. Increasing wild Bovidae species richness was associated with an increase in landscape suitability, as was sheep density up to an approximate average of 100 animals per km2, after which it decreased and remained constant at 250 animals per km2. Muridae richness demonstrated a V-shaped relationship with high landscape suitability in areas of low and high species richness. Close proximity to intermittent wetlands and freshwater marsh was associated with greater suitability to RVF outbreaks, as was low precipitation during the driest quarter. The model performed well when tested against the OIE data with the AUC equal to 83%.
This is the first study to explore and identify the influence of diverse surface water types on RVF outbreaks at a continental scale. The distribution of hydrogeographic features, particularly intermittent wetlands, contributed to suitable landscape conditions for RVF outbreaks. Our model also indicated that Bovidae species richness, sheep density and (to a lesser extent) Muridae species richness were predictive of RVF landscape suitability. This supports field observations that suggest that wild ruminants and rodents are the most likely wild reservoirs for RVFV [14]. These two important landscape features, i.e. wetland hydrogeography and Bovidae species richness, are novel contributions to our understanding of the epidemiology and infection ecology of RVF outbreaks in the African continent and the Arabian Peninsula.
Proximity to intermittent wetlands was particularly important to RVF landscape suitability, as identified by its permutation importance (18%). Hydrogeography has not previously been investigated across the diverse spectrum of surface water types in association with RVF occurrence in continent or country-wide analyses. One comprehensive study did find that proximity to rivers was an important determinant of landscape suitability across the African continent [43]. Our model, which is based on a larger number of epidemic occurrences, identified other wetland features–namely intermittent wetlands and freshwater marsh–as more important contributors to RVF risk. Nonetheless, the two continent-wide studies are in general agreement with respect to high risk landscapes in the Sahel, and eastern and southern Africa.
A regional study in the Ferlo area of Senegal examined proximity to transient ponds in hyperlocal settings of mixed vegetation cover and found that the juxtaposition of these small bodies of water with dense vegetation was strongly associated with positive serology in sheep and goats [44]. Moreover, this same group demonstrated that these landscape features corresponded to fluctuations in Aedes and Culex species during RVF epidemic years [45]. These kinds of transient surface water features were the most influential features to RVF landscape suitability in our model as well. Another study in the Mbeya region of Tanzania, found that proximity to Lake Malawi was very strongly associated with antibody evidence of past infection in humans [46,47], which is consistent with our finding of the importance of fixed freshwater sources such as freshwater marshes.
Our understanding of the landscape epidemiology of RVFV is, perhaps, most deficient with respect to reservoir hosts. Several potential mammalian RVFV hosts have been studied in both field and experimental settings to identify natural reservoirs in RVFV infection ecology [14]. While several species across multiple orders may be implicated as reservoirs, one review identified ruminants and rodents as likely natural reservoir(s) [14]. Evidence from the same study suggests that ruminants are more effective in virus amplification rather than maintenance and this may pose the greatest risk to humans in proximity. Consistent with this, our model demonstrated that wild Bovidae species richness was important in delineating the landscape suitability of RVF (permutation importance 11.7%). Interestingly, this landscape feature was as important as sheep density and more important than density of goats and cattle (see below). Muridae species richness was not as influential as Bovidae species richness, but the former did contribute to RVF landscape suitability (permutation importance = 8.6%). The V-shaped response curve also suggested high risk landscapes associated with low and high Muridae species richness and lower risk landscapes across intermediate species richness. While some previous work has identified the possibility of individual Muridae species as possible maintenance [14] or amplification hosts [48–54], the findings from the current study cannot attribute either role to the Muridae.
The density of sheep was strongly associated with landscape suitability to RVF outbreaks (permutation importance = 11%). This is not surprising given the high susceptibility of sheep [5] and their focus within most of the RVF outbreaks that occurred between 1998 and 2016 [55,56]. Moreover, it is through contact with animals and animal products that most human RVF occurs [13,57,58]. Goat and cattle density did not contribute substantively to RVF landscape suitability(permutation importance < 3 for both). While goats certainly have been affected in many RVF outbreaks, their lesser susceptibility relative to both sheep and cattle has been previously demonstrated [56].
Dry quarter precipitation was a moderately strong contributor to landscape suitability. Low levels of precipitation were associated with increased landscape suitability during the driest period of the year. However, as precipitation increased during this time, RVF suitability sharply decreased (S1 Fig). This may suggest that climatically drier areas are more susceptible when periodically inundated with sporadic episodes of rain [56]. Indeed, barren, arid landscapes punctuated with temporary ponds were associated with the greatest proliferation of both maintenance and amplification vector mosquitoes in West Africa [55,59]. Furthermore, regions with sensitive hydrogeographic dependency demonstrate high RVF susceptibility to precipitation variability attributable to El Nino southern oscillations [18,60]. Indeed, the latter, more recent, study identified a clear sensitivity to both seasonal climatic variation and El Nino oscillation across the whole of the African continent [60]. While the current findings may suggest climatic variations in arid to semi-arid conditions, or extremes of precipitation between dry and wet seasons, we emphasize that our study did not assess the relationship between RVF outbreaks and specific weather events.
This investigation identified landscapes suitable to RVFV beyond the historical extent of transmission. This suggests that conditions may be favorable for transmission should the virus be introduced into these naïve geographic spaces. In essence this allows us to conceive of how a realized niche in one space may be related to a potential niche in another, and what structures may be necessary or useful to prevent the realization of the niche in the novel space. For example, the introduction of RVFV from infected cattle herds to non-infected susceptible herds has been well documented in both local [61] and global [62,63] scales. More specifically, network structure and resulting dissemination of herd animals across villages has been shown to drive the introduction or re-introduction of RVFV to susceptible livestock herds [61]. Translocations of RVFV at more global scales can follow international trade in livestock [62], as well national-level trade in livestock [63]. Surveillance mechanisms should target transboundary livestock movement and mosquito control along livestock migratory corridors for effective prevention of RVF transmission. Indeed, some success has been documented with respect to cross-disciplinary, systems-based approaches to RVFV surveillance in East Africa as organized by the Armed Forces Health Surveillance Center, Division of Global Emerging Infections Surveillance and Response System Operations (AFHSC-GEIS)[64].
A perennial problem in ecological niche modeling, in general, and in emerging disease risk mapping, in particular, is the lack of independent sample data for testing trained models [65]. Typically only one sample is available, which is partitioned into training and testing datasets. A strength of the current study is that we utilized two distinct sources of RVF outbreak data; one (ProMED) to train the landscape suitability model and the other (OIE) to test the prediction error of the model. Nonetheless, the findings from the current study must be interpreted with caution due to its limitations. First, temperature and precipitation were measured as single composites over the period 1950 to 2000. Therefore, while the spatial resolution of these measurements was high (~1 km2 rasters), the analysis was simultaneously constrained by coarse temporal resolution, given that the climate rasters represent 50 year averages. Second, the number of documented RVF occurrences is small. Therefore, with only 103 ProMED-reported occurrences and 78 unique OIE-reported occurrences, this collection of RVF outbreaks may not represent the true incidence across the continent of Africa and the Arabian Peninsula. Furthermore, the model may produce an underfit representation of the ecological niche of RVF. Nevertheless, the current analysis attempted a robust validation of models trained on limited data by using a second vetted source of RVF occurrence data from OIE for testing. Third, due to a lack of adequate presence data, the current investigation was unable to evaluate the influence of Aedes and Culex mosquitoes. Because these vectors are part of the RVFV infection cycle, the lack of Aedes and Culex mosquitoes in our model necessarily demarcates a less refined picture of RVF landscape suitability. Nevertheless, the hydrogeographic features identified by the model as influential to RVF suitability correspond to surface water features documented to be most favorable to the Aedes mosquitoes [44,45,66] and Culex mosquitoes [2,59,67] and so the model still appears to identify landscapes appropriate to the relevant vector ecology.
The findings of this study will help to improve our understanding of the landscape epidemiology of RVF outbreaks. The model of RVF landscape suitability will require further development as more data become available to validate the results. Moreover, because this study is observational rather than experimental, we cannot assign causation to the associations between RVF outbreaks and landscape factors, including surface water features and mammalian hosts. More definitive causal inference demands better assessment of species in the landscapes where outbreaks are occurring, and crucially at high spatial resolution and in real time. This will require thorough field studies that combine animal serology and viral RNA detection, more detailed habitat description, observation of wild Bovidae movement patterns, wildlife-domestic animal interaction, and maintenance and amplification vectors, as well as cultural and economic drivers of the livestock industries.
This study found that proximity to wetlands in landscapes that are rich in wild Bovidae species and high in sheep density, delineate the most suitable landscapes for RVF outbreaks. Moreover, this study found that the RVF risk surface may extend to regions beyond the historical range of past zoonotic experience should the virus be introduced to these regions via livestock transport or local invasion by infected mosquitoes.
|
10.1371/journal.ppat.1006543 | Murine gammaherpesvirus M2 antigen modulates splenic B cell activation and terminal differentiation in vivo | Murine gammaherpesvirus 68 (MHV68) infection of laboratory strains of mice has provided a tractable small animal model for dissecting gammaherpesvirus pathogenesis. The MHV68 latency associated antigen M2 promotes viral latency establishment in germinal center (GC) B cells and plays an important role in virus infection of plasma cells (PCs), which is linked to virus reactivation. More recently, M2 has been highlighted as a potent immunomodulatory molecule capable of hindering both cell-mediated and humoral immunity to MHV68 infection and subsequent challenges. M2 expression in B cells results in activation of B cell receptor signaling pathways that promote proliferation, differentiation, and cytokine production—a hallmark of gammaherpesviruses. In this study, we utilized an adoptive transfer model to explore the biological consequence of M2 expression in activated B cells in vivo. Secondly, we engineered and validated two independent MHV68 M2 reporter viruses that track M2 protein expression in latently infected B cells during infection. Here we demonstrate that upon adoptive transfer into naive mice, M2 expression promotes activated primary B cells to competitively establish residency in the spleen as either a GC B cell or a PC, most notably in the absence of an ongoing GC reaction. Moreover, M2 antigen drives robust PC differentiation and IL10 production in vivo in the absence of other viral factors. Lastly, we confirm that M2 expression during MHV68 infection is localized to the GC compartment, which is a long term latency reservoir for gammaherpesviruses. Overall, these observations are consistent with, and extend upon previous reports of M2 function in B cells and within the context of MHV68 infection. Moreover, this work provides support for a model by which M2-driven dysregulation of B cell function compromises multiple aspects of antiviral immunity to achieve persistence within the infected host.
| Gammaherpesvirus (GHVs), which primarily infect B cells, are capable of exploiting B cell biology to achieve a stable and persistent infection for the lifetime of the host. GHV infections traffick to germinal center (GC) B cells and plasma cells (PCs), which are important immune effectors that promote the generation of protective antibodies in response to pathogens. The mechanism by which murine gammaherpesvirus 68 (MHV68) M2 latency protein activates B cell receptor signaling pathways to modulate the immune response to infection and further promote viral pathogenesis within the GC B cell and PC compartments is not completely understood. Here we demonstrate that M2 expression alone, in the absence of other viral factors, drives robust PC differentiation and IL10 production in vivo. Moreover, M2 promotes the accumulation of splenic GC B cells, which was subsequently verified as the site for potent M2 expression during latent MHV68 infection. Our work further substantiates a model in which a viral protein dysregulates B cell activation, differentiation, and cytokine production to create a permissive environment for viral persistence in the infected host. This work justifies further investigations addressing the impact of GHV latency antigen function within the GC reaction and overall host response to infection.
| Herpesvirus infections characteristically exhibit dynamic host-pathogen interactions that promote viral persistence for the lifetime of the infected host (reviewed in [1]). Gammaherpesviruses (GHVs) primarily infect and establish latency in B cells and can potentially trigger lymphomagenesis in an immunosuppressive environment. For example the human GHVs, Epstein-Barr virus (EBV) and Kaposi’s sarcoma-associated herpesvirus (KSHV), have been identified as the etiological agents of Burkitt’s lymphoma and Kaposi’s sarcoma, respectively [2, 3]. Although studies utilizing immortalized latently infected cells lines and transgenic mice have provided valuable insights into the functions GHV antigens in B cells, the narrow host cell tropism of EBV and KSHV, coupled with the lack of robust small animal models for these human pathogens, has significantly impacted research efforts with respect to viral pathogenesis studies in the infected host. Murine gammaherpesvirus 68 (MHV68), which exhibits similar genomic organization and extensive sequence homology with other GHVs, is a natural rodent pathogen that has proven to be a useful tool for studying latency, reactivation, and pathogenesis [4]. MHV68 infection of laboratory strains of mice results in a brief phase of acute replication followed by subsequent latency establishment in macrophages, dendritic cells and B cells, with the latter representing the predominant latency reservoir in vivo [5–7]. Combined with the fact that MHV68 can infect various cell lines in vitro, this model provides a robust system that can be utilized to interrogate the functional role of both host and viral factors in GHV pathogenesis.
Non-specific B cell activation and lymphoproliferation are markers commonly associated with herpesvirus infections and this phenomenon is further exploited by GHVs that encode unique latency associated antigens capable of modulating B cell signaling activity [8–10]. EBV proteins LMP1 and LMP2a are constitutive CD40 and BCR mimics, respectively, that provide latently infected B cells with survival signals in the absence of T cell help and antigen recognition [11, 12]. Transgenic expression of LMP1 or LMP2a in murine B cells results in enhanced survival, proliferation, differentiation, and immunoglobulin production [11–13]. During latent EBV infection, LMP1 and LMP2a expression is associated with naïve and germinal center (GC) B cells and their proposed function is to drive latently infected B cells through the GC reaction and into the long lived memory B cell reservoir [14, 15]. However, the ultimate impact of GHV latency antigen signaling activity with respect to B cell differentiation and viral trafficking during EBV infection is currently under debate. The MHV68 M2 latency associated antigen, although it has no obvious viral or cellular homologs, exhibits similar functions upon expression in murine B cells. M2 functions as molecular scaffold that triggers the activation of BCR signaling pathways resulting in B cell proliferation, differentiation, and immunoglobulin production in vitro [16–19]. During MHV68 infection in vivo, M2 expression enhances latency establishment in splenic GC B cells and is particularly critical for plasma cell (PC) differentiation of latently infected B cells and subsequent virus reactivation, which is a hallmark of GHV infection [20–24].
Recently, an expanding body of literature provides convincing evidence that B cells exhibit potent immunoregulatory activity via provision of IL10 in a number of auto-immune diseases and pathogen infections [25–30]. For example, B-cell derived IL10 suppressed pathogen-specific CD4+ T cell, natural killer, and neutrophil responses which correlated with decreased survival in a salmonella infection model [28]. Activation of TLR, CD40, and/or BCR signaling induces IL10 production in B cells, but the mechanisms underlying the development and function of IL10-competent B cells have yet to be fully elucidated [26, 31, 32]. The MHV68 M2 antigen drives robust IL10 expression from B cells through upregulated expression of B cell differentiation factor IRF4, which supports B cell proliferation and differentiation in vitro [17, 19]. Accordingly, M2 expression during MHV68 infection contributes to elevated IL10 serum levels, reduced antiviral CD8+ T cell activity and attenuated MHV68-specific humoral responses [19, 33]. Moreover, loss of M2 expression restored parasite-specific antibody responses that effectively rescued animals from a lethal co-infection with MHV68 and rodent plasmodium [33]. Therefore, further investigation of M2 may provide important insights into common mechanisms by which GHVs, as well as potentially other pathogens, manipulate B cell biology to alter the host response to infection.
While previous studies have reported the detection of M2 transcripts in various splenic B cell subsets following MHV68 infection, the mechanism by which M2 antigen expression dampens antiviral immunity and promotes MHV68 pathogenesis in vivo has not been rigorously addressed [5, 34, 35]. In this study, we sought to interrogate: (i) the impact of M2 expression in the context of adoptively-transferred splenic B cells; and (ii) the location of M2 antigen expression during latent MHV68 infection in vivo. Here, we show that M2 expression alone, in the absence of other viral factors, dramatically modulates B cell activity similar to that observed during MHV68 infection in mice. Moreover, these observations may represent a common mechanism by which MHV68 and other pathogens subvert the adaptive immune response by dysregulating B cell function during infection.
Extensive characterization of MHV68 M2 antigen function in vitro has revealed a wealth of information, namely that it functions as an adaptor protein by triggering the assembly of multimeric protein complexes that activate various B cell signaling pathways [16–18, 36]. In particular, retroviral transduction of LPS-stimulated primary B cells with a constitutively active M2 expression vector results in enhanced B cell proliferation and IL10 production [19]. Due to the fact that IL10-dependent cell proliferation requires a functional M2, we utilized this retroviral transduction system as a screening tool to evaluate the detection efficacy of an intracellular marker that could monitor M2 protein expression without adversely disrupting its function. To this end, an M2-mCherry transgene was generated by fusing the M2 ORF upstream of the mCherry fluorescent protein sequence with an intervening 30 amino acid F2A peptide derived from the foot-and-mouth disease virus 2A [37]. The incorporation of the F2A peptide allows for cotranslation of separate polypeptides from a single mRNA transcript [37], thus enhancing the likelihood of preserving both M2 and mCherry functionality. The M2-mCherry construct was cloned into a replication-defective murine stem cell virus (MSCV) vector upstream of an IRES-Thy1.1 cassette, which facilitates detection of transduced B cells by surface Thy1.1 expression (Fig 1A) [19]. Primary B cells were isolated from bulk splenocytes by negative selection and stimulated overnight with LPS prior to transduction with the MSCV-M2-mCherry retrovirus, or the previously described wild type MSCV-M2 and negative control MSCV-M2.stop retroviruses (the latter contains a translation stop codon at amino acid 13 of the M2 open reading frame) (Fig 1A). Following retroviral transduction, B cell cultures were harvested in triplicate at each time point and monitored for cell surface Thy1.1 expression and IL10 production. Consistent with previously published reports [17, 19], transduced cells (Thy1.1+) expressing M2 expanded to ~90% of the culture and secreted ~40ng/mL of IL10 by day five post-transduction, which was not observed in M2.stop-transduced cultures (Fig 1B and 1C). M2-mCherry expressing cultures exhibited Thy1.1+ B cell expansion and robust IL10 production nearly identical to that observed with the wild type M2 expression construct (Fig 1B and 1C), indicating that M2 protein expressed from the fusion gene possessed wild type M2 function in B cells.
Following verification of M2-mCherry function in primary B cells, we sought to evaluate the efficiency of mCherry detection with respect to cell surface marker Thy1.1, which serves as a marker for transduction efficiency [19]. We initially visualized robust mCherry expression by fluorescence microcopy in transduced primary B cell cultures expressing M2-mCherry, but not in cells expressing wild type M2 (Fig 2A). Analysis of intracellular and cell surface marker expression by flow cytometry demonstrated that the MSCV-M2-mCherry transduced (Thy1.1+) population exhibited robust mCherry expression, which was not observed in untransduced (Thy1.1-) or MSCV-M2 transduced B cells (Fig 2B). Accordingly, mCherry and Thy1.1 expression were detected at a 1:1 ratio in M2-mCherry-transduced B cells throughout the time course, while the frequency of mCherry+ population in M2-transduced B cell cultures remained at ≤5% of total cells (Fig 2C). Thus, we concluded that both M2-mCherry and M2-Thy1.1 constructs could serve as faithful reporters that preserve wild type M2 function and facilitate detection of M2 protein expression in murine B cells.
Constitutive M2 expression in B cells results in upregulated expression of several plasma cell differentiation factors leading to robust IRF4-driven IL10 production, immunoglobulin secretion, and differentiation into an activated pre-plasma memory B cell (CD19+GL7HiB220LosIgD-sIgG+CD138Lo) in vitro [17, 19, 22]. To further evaluate the consequence of M2-driven activation of B cell signaling pathways, we sought to determine the fate of M2-expressing B cells upon adoptive transfer into a naïve host. Primary splenic B cells isolated from CD45.1+ donors were stimulated with LPS and retrovirally transduced with MSCV-M2 or MSCV-M2.stop retroviruses as previously described, prior to adoptive transfer into the peritoneum of naïve C57BL/6 mice (CD45.2+). Splenocytes from adoptive transfer recipients were harvested at one and five days post-transfer and analyzed by flow cytometry to derive the absolute number of cells per spleen for the indicated populations (Fig 3A). A fraction of the adoptively transferred B cell population trafficked to the spleen and the overall recovery of CD45.1+ cells from both M2 and M2.stop adoptive transfer recipients was comparable throughout the time course (Fig 3B). However, transduced B cell numbers (CD45.1+Thy1.1+) were 7-fold higher in M2 animals as compared to M2.stop animals at D5 post transfer (Fig 3C). Interestingly, M2-transduced B cells comprised ≥80% of the total splenic CD45.1+ population at D5 post transfer while the frequency of M2.stop-transduced B cells remained at just ~10%. We also observed a modest yet significant 3-fold increase in M2-transduced B cells numbers in the B220+ fraction versus the corresponding M2.stop-transduced population at D5 post transfer (Fig 3D). Due to the dynamic nature of this system, we are unable to address whether M2 expression confers enhanced survival or proliferation—although a previous study demonstrated that both factors contribute to the expansion of M2-transduced B cells in vitro [19]. Next, we sought to evaluate whether the splenic microenvironment could further influence the differentiation state of M2-expressing B cells in vivo. Transduced B cells from the total CD45.1+ B cell population and the CD45.1+B220+ fraction were subphenotyped for cell surface markers consistent with plasma cell (PC) and germinal center (GC) B cell phenotypes, respectively (Fig 3A). The numbers of transduced B cells exhibiting either a PC (B220LoCD138Hi) or GC (B220+GL7HiCD95Hi) phenotype per spleen at D1 post transfer were indistinguishable for both conditions (Fig 3E and 3F). However, by D5 post transfer, mice receiving M2-transduced B cells exhibited a striking 208-fold and 37-fold enhancement of Thy1.1+ B cells exhibiting a PC or GC phenotype, respectively, versus the corresponding populations in mice receiving M2.stop-transduced B cells. Importantly, the observed decline in M2.stop transduced B cells exhibiting a GC phenotype from D1 to D5 post-transfer cannot be attributed to a massive loss of M2.stop-transduced B cells, as we detected a substantial Thy1.1+ population in M2.stop adoptive transfer recipients at D5 post-transfer (Fig 3C). These results clearly demonstrate that PC and GC B cell populations expressing M2 exhibit a competitive advantage over the untransduced B cell population with respect to maintaining residence in the spleen.
To further evaluate B cell differentiation status as a function of M2 concentration, adoptively transferred B cells were divided into Thy1.1 –(no M2 expression), Thy1.1 Lo/Int (low to intermediate levels of M2 expression), and Thy1.1 Hi (high level M2 expression) populations (Fig 4A), followed by analysis of cell surface marker expression for each group. Similar to previous reports, B220 expression was significantly downregulated in the M2-Thy1.1 Hi population, while M2.stop-Thy1.1+ B cells exhibited levels of B220 similar to that of the untransduced Thy1.1 –population ([17, 19]; Fig 4B and 4C). Consistent with M2-driven IRF4 expression in B cells [17], the PC marker CD138 was robustly upregulated in ca. 70% of transduced B cells expressing the highest levels of M2 (Thy1.1 Hi) and ca. 15% of B cells expressing intermediate or low levels of M2 (Thy1.1 Lo/Int), as compared to ca. 2% of the untransduced B cell population (Fig 4B). We also observed graded expression of the B cell activation marker GL7, which appeared to be highest in the M2-Thy1.1 Lo/Int population and was almost undetectable in M2.stop Thy1.1+ B cells (Fig 4). Finally, we assessed intracellular IL10 protein expression in these transduced B cell populations–which revealed robust induction of IL10 in the M2 Thy1.1 Hi population and a more modest induction of IL10 expression in the Thy1.1 lo/Int B cell population (Fig 4B). This is consistent with our previous studies demonstrating that IL10 is a downstream target of M2-induced IRF4 expression [17]. Overall, this data confirms that the observed changes in cell surface marker expression are a direct consequence of M2 expression–as opposed to the mere ability to survive within the splenic microenvironment longer than the M2.stop transduced B cell populations–and further establishes M2 as a potent factor with respect to B cell activation and differentiation. In conclusion, this is the first evidence that M2 expression alone, in the absence of an ongoing infection and other viral factors, can support the activated GC B cell phenotype and directly promote PC differentiation and IL10 production from stimulated B cells in vivo.
Given that the phenotype of adoptively transferred M2-transduced B cells was consistent with MHV68 latency reservoirs in vivo, we sought to interrogate the context of M2 expression during infection. To this end, we generated two independent M2 reporter viruses in the background of the previously described MHV68-H2bYFP virus, which marks latently infected B cells with intracellular YFP that can be detected by flow cytometry [38]. Previous studies demonstrated that YFP+ splenic B cells contain the viral genome and are capable of reactivating infectious virus, which allows us to utilize YFP expression as an internal control for latency establishment efficiency. To track M2 expression under the control of its native promoter, the M2 ORF in the context of the MHV68-H2bYFP BAC genome was substituted with either the M2-mCherry or the M2-Thy1.1 reporter construct sequences via the galK recombination system (Fig 5A) [39]. The recombinant MHV68-H2bYFP BAC clones were sequenced at the M2 locus to confirm the presence of the inserted constructs and the integrity of the viral genome was evaluated by RFLP analysis.
M2 expression during MHV68 infection in vivo enhances latency establishment and virus reactivation in a dose and route dependent manner [20, 21]. To evaluate overall viral fitness and account for any deleterious genomic alterations, we sought to quantify latency establishment efficiencies for the recombinant MHV68-H2bYFP M2 reporter viruses. C57BL/6 mice were infected at 1000 PFU via the intraperitoneal (IP) route with the parental MHV68-H2bYFP virus and two independent clones of either the M2-mCherry or M2-Thy1.1 reporter viruses and intracellular YFP expression in latently infected splenocytes was detected by flow cytometry. Consistent with previous reports, the H2bYFP virus displayed a typical variation in the frequency of latently infected B cells (%B220+YFP+) that averaged ~0.3% at 14dpi (Fig 5B and 5C) [40, 41]. M2-mCherry- and M2-Thy1.1- infected animals exhibited on average ~0.2% and ~0.3% B220+YFP+ cells, respectively, which was not significantly different to that of the parental MHV68-H2bYFP virus (Fig 5B and 5C). We subsequently evaluated virus reactivation from latently infected splenocytes upon explant into tissue culture by utilizing the previously described limiting dilution ex vivo reactivation assay [42]. The average frequency of cells reactivating infectious virus following M2-mCherry (1 in 6,309 cells) and M2-Thy1.1 (1 in 5,128 cells) infection was similar to that of parental H2bYFP (on average 1 in 6,165 cells) (Fig 6A and 6B). In conclusion, our extensive evaluation of the recombinant MHV68-H2bYFP M2 reporter viruses verify that the addition of exogenous sequences at the M2 locus did not dramatically alter M2-driven latency establishment in vivo and subsequent ex vivo virus reactivation from latently infected splenic B cells.
Previous studies characterizing the MHV68-H2bYFP virus demonstrated that a majority of latently infected B cells faithfully display either a GC or a PC phenotype [38]. Incorporation of the M2.stop mutation in the MHV68-H2bYFP background revealed that M2 expression was dispensable for access to the GC B cell compartment, but critical for viral entry into the PC compartment that serves as a major reservoir for reactivating virus [22]. Accordingly, we sought to analyze the cell surface phenotype of M2-expressing B cells during latency establishment in vivo. To facilitate a side by side comparison of the engineered MHV68-H2bYFP M2 reporter viruses, animals were infected at 1000 PFU IP with either the parental H2bYFP virus, M2-mCherry virus, or M2-Thy1.1 virus. First, we quantified the M2 reporter positive population within the latently infected B220+YFP+ population by utilizing the parental H2bYFP virus as a negative control. Following IP infection, we detected a robust population of cells that was positive for intracellular mCherry or cell surface Thy1.1 expression with the respective reporter virus, which was not present in H2bYFP-infected animals (Fig 7). Very similar results were obtained with M2-mCherry and M2-Thy1.1 reporter virus infection, which resulted in ~40% of the latently infected B cell population exhibiting detectable M2 reporter expression at 14 dpi (Fig 7). Latently infected B220+YFP+ B cells characteristically exhibit a GC phenotype during latent MHV68 H2bYFP infection, which was recapitulated in M2 reporter virus infected animals (Fig 8A, upper panels). Accordingly, ≥90% of mCherry+ and Thy1.1+ cells within the B220+YFP+ compartment reproducibly and almost exclusively exhibited a GC phenotype (Fig 8A, lower panels). Latently infected populations were also analyzed for cell surface markers consistent with a PC phenotype, which comprises ~20% of the CD3-YFP+ population (Fig 8B, upper panels). Although both reporter viruses were capable of efficiently establishing latency within the PC reservoir, we were consistently unable to detect M2 reporter activity in latently infected PCs using either the M2-mCherry or M2-Thy1.1 viruses (Fig 8B, lower panels).
We subsequently evaluated MHV68 H2bYFP M2-Thy1.1 infection following inoculation at 1000 PFU IN, which represents a more stringent route of infection, in order to reveal any potential defects with viral trafficking to the spleen. Analysis of infected splenocytes at 16 dpi via the IN route demonstrated that the M2-Thy1.1 virus exhibited no significant alterations in latency establishment or viral trafficking to the GC and PC compartments (Fig 9A–9C). M2 reporter positive cells comprised on average ~20% of the B220+YFP+ population (Fig 9D), which was modestly reduced compared to the IP route (~40%; Fig 7B). Despite a slightly lower frequency of YFP+Thy1.1+ cells, we found that M2 reporter activity was reproducibly detected within the GC, but not the PC compartment (Fig 9E and 9F), which is consistent with the results obtained following IP infections (Fig 8). In conclusion, our M2 reporter viruses system has clearly established the latently infected GC B cell compartment as a critical site of potent M2 expression during latent MHV68 infection in vivo.
Prior studies have demonstrated a critical role for MHV68 M2 with respect to viral latency establishment and virus reactivation in a mouse model of infection [20–22]. In this study, we utilized complementary methods to further elucidate the context and potential impact of M2 expression with respect to B cell function in vivo. Our adoptive transfer studies are the first demonstration that M2 drives robust PC differentiation and IL10 production in vivo in the absence of other viral factors (Figs 3 and 4). However, we were unable to detect M2 reporter activity in the latently infected PC compartment (Figs 8 and 9), which may indicate that the M2 promoter is no longer active once the cell has reached a sufficient level of IRF4 expression required for terminal differentiation into a PC [43]. Additionally, the timing and low level of M2 protein expression that is required for PC differentiation, as demonstrated in our adoptive transfer system (Fig 4), may hinder our detection of M2 protein at 14 dpi in the M2 reporter virus system (Figs 8 and 9). Although we were unable to directly confirm protein expression in the latently infected PC, the adoptive transfer data fits into a well-established model in which M2-driven IRF4 expression facilitates PC differentiation during MHV68 infection [17, 20, 22, 44]. A significant portion of latently infected PCs can be generated via the extrafollicular pathway [45] and we have shown here that M2 drives robust PC differentiation in the absence of an ongoing GC reaction (Figs 3 and 4). How do latently infected PCs contribute to GHV pathogenesis? PC differentiation initiates lytic replication of EBV and KSHV via activation of viral transactivators by PC-specific transcription factors such as XBP-1 and Blimp-1 [23, 24, 46–48]. At least one study has reported that de novo KSHV infection drives human tonsillar B cells to proliferate and differentiate into plasmablasts that functionally and phenotypically resemble multicentric Castleman’s disease [49]. Therefore, PC differentiation represents a common aspect of GHV pathogenesis and this mechanism of virus reactivation has been proposed to facilitate virus transmission and maintenance of a stable life-long infection. For MHV68, PCs have been identified as the predominant source of infectious virus production, and virus reactivation is severely impaired in the absence of M2 [22]. Moreover, PCs appear to play an important role in virus trafficking and seeding of chronic MHV68 latency reservoirs. For example, an M2-null virus exhibits significantly impaired latency establishment in the spleen at 16 dpi, despite efficient viral replication in the lungs following low dose intranasal inoculation. Additionally, long term latency maintenance at 90 dpi was severely attenuated in mice lacking splenic PCs in a conditional Blimp-1 knockout model. In combination, our studies support a model in which GHVs play a direct role in driving PC differentiation which serves to facilitate reactivation of infectious virus and reseeding latency reservoirs within the infected host.
In addition to directly promoting virus reactivation, IL10 production by PCs has the potential to suppress humoral immunity and create a more permissive environment for viral infection. IL10 is a potent immunomodulatory cytokine that impairs T cell, macrophage and dendritic cell functions in a variety of infection settings (reviewed in [50]). GHVs exploit the IL10 signaling pathway by encoding viral IL10 homologs and/or enhancing IL10 expression from B cells, which promotes B cell expansion and abrogates immune recognition and subsequent eradication of infected B cells in vitro [19, 51–54]. During primary MHV68 infection M2 functions as an immunomodulatory molecule by elevating serum IL10 levels, attenuating antiviral CD8 T cell responses, and suppressing antigen-specific responses to MHV68 and subsequent challenges[19, 33]. Here we have shown that M2-driven PC differentiation is characterized by robust CD138 expression and IL10 production (Fig 4), which is consistent with M2-driven IRF4 production in B cells [17]. Interestingly, B cell IL10 production also attenuates aspects of innate and adaptive immunity in a salmonella infection model, and IRF4HiCD138Hi PCs have been identified as a potent source of IL10 [28, 29]. Therefore, we propose that latently infected PCs could serve as one potential source of immunsuppressive IL10 production during MHV68 infection. Moreover, dysregulated BCR signaling may represent a common mechanism by which GHVs and other pathogens promote regulatory PC generation as an immune evasion tactic during infection. The ultimate significance of IL10 expression in the context of GHV infection in vivo is still under debate, and investigations are currently underway to evaluate the contribution of host-derived IL10 to MHV68 pathogenesis.
The prevailing model of GHV pathogenesis requires that the virus traverse the GC compartment in order to gain access to the long-lived memory B cell compartment (reviewed in [55]). The GC reaction represents a competitive environment wherein B cells that do not receive rescue signals triggered by antigen recognition or T cell help are subjected to death by apoptosis while high affinity B cells that successfully compete for limited T cell help are positively selected to enter the long-lived B cell compartment [56, 57]. Similar to EBV, MHV68-latently infected B cells resemble, localize and participate in ongoing GC reactions. Importantly, and in contrast to EBV, T cell help is a demonstrated requirement for expansion of latently infected GC B cells and entry to the memory B cell pool, which serves as the long term latency reservoir for GHVs [15, 40, 58, 59]. M2 transcripts have been detected in GC B cells during chronic infection [5, 34] and our M2 reporter virus system has independently verified that the latently infected GC compartment is a site of robust M2 antigen expression at the peak of MHV68 latency, which was independent of the route of virus inoculation (Figs 8 and 9). In an increasingly hostile environment, M2 antigen expression could promote latently infected B cell survival and exit from the GC as a memory B cell. In this study, we show that M2 antigen expression in stimulated B cells was able to support, at least temporarily, the activated GC phenotype in the absence viral infection (Figs 3 and 4). Moreover, M2-driven signaling promotes the formation of B-T cell conjugates in the presence and absence of specific peptide [60], which in the context of the GC reaction could provide sufficient CD40 stimulation to enhance survival and selection of latently infected B into the memory B cell pool [61–63]. Further investigation is necessary in order to determine if M2 expression in GC B cells promotes viral trafficking to the memory B cell pool in vivo. Previous reports have confirmed the presence of M2 transcripts in memory B cells [5], but these analyses are inherently misleading and more sensitive and quantitative analyses are required to correlate MHV68 transcriptional programs with specific stages of B cell differentiation. Thus, enhanced characterizations of MHV68 latency antigen function and transcriptional programs may reveal common strategies by which GHVs effectively manipulate GC B cell biology to achieve short and/or long term persistence in vivo.
In conclusion, our studies have further validated a model in which M2 antigen expression dysregulates B cell activation, differentiation, and cytokine production to simultaneously thwart immune detection and eradication and promote MHV68 pathogenesis in the infected host. While the role of M2 expression within the GC B cell remains unknown, it has great potential to significantly influence both B and T cell responses to MHV68 infection. Therefore, our studies justify continued investigations that address the impact of M2 expression with respect to the global GC response during primary and secondary infections, as this may provide important insights with respect to GHV pathogenesis and associated disease.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Emory University Institutional Animal Care and Use Committee (IACUC) and in accordance with established guidelines and policies at Emory University School of Medicine (protocol number: DAR 2003399-022419BN).
BAC-derived MHV68 viruses were reconstituted following transfection of Vero-Cre cells, a generous gift provided by David Lieb [64]. Recombinant MHV68 viruses were propagated and titered on NIH3T12s (ATCC: CCL-164) as described previously [65]. Murine embryonic fibroblasts (MEFs) utilized in ex vivo reactivation assays were isolated from day 16 C57BL/6J embryos cultured as previously described [42]. Adherent cell lines were maintained in Dulbecco’s modification of Eagle medium (DMEM) supplemented with 10% fetal bovine serum, 2mM L-glutamine, and 100U penicillin and 100mg streptomycin per mL. Primary B cells isolated from C57BL/6J mice were maintained in Roswell Park Memorial Institute (RPMI) 1640 Medium supplemented with 10mM non-essential amino acids, 1mM sodium pyruvate and 10mM HEPES.
Plasmid MSCV-M2-IRES-Thy1.1 (MSCV-M2) and MSCV-M2.stop-IRES-Thy1.1 (MSCV-M2.stop) have been previously described [19]. The mCherry protein sequence was fused to the C terminus of the M2 ORF in sequential steps as follows: the M2 sequence was amplified from MSCV-M2 using primers 5’-ctagagatctatggccccaacaccc-3’ and 5’-ctaggtttaaactctcctcgccccactc-3’ (flanking 5’ BglII and 3’ PmeI sites) and inserted into pCR-Blunt II Topo vector (Invitrogen) per manufacturer’s instructions; mCherry was amplified from pTREG-mCherry (Clontech) using primers 5’-ctaggtttaaacgtgagcaagggcgag-3’ and 5’-ctagctcgagagatcttcacttgtacagctcgtcc-3’ (flanking 5’ PmeI and 3’ Xho I BglII sites) and subsequently subcloned into to pCR-Blunt M2 vector following digestion with PmeI and XhoI. Lastly, a peptide sequence containing the AUI epitope (DTYRYI) and 30 amino acid F2A sequence [37] was inserted between the M2 ORF and the mCherry ORF using an overlapping PCR mutagenesis technique [66] with the following primer pairs: primers A&B 5’-ctagagatctatggccccaacaccc -3' and 5’-cgataggtatcctcctcgccccactcc -3'; primers C&D 5’-ggcgaggaggatacctatcgctatatacacaagcaaaagatcgttgcaccagttaagcaga ctctgaattttgacc-3' and 5’-cgtttaaactgggcccagggttggactcaacgtctccggccaacttgagcaggtcaaaattca gagtctgc-3'; primers E&F 5’-cctgggcccagtttaaacgtgagcaagggc-3' and 5’-ctagagatcttcacttgtaca gctcgtccatg-3'. PCR products AB, CD, and EF were utilized in the second ligation PCR step to generate the final M2-AU1-F2A-mCherry transgene (M2-mCherry) that was subsequently subcloned into the MSCV-IRES-Thy1.1 using BglII sites as previously described [19] to generate MSCV-M2-mCherry.
Retroviruses were produced by transfecting retroviral packaging cell line BOSC23 (ATCC) with individual MSCV vectors as previously described [41]. Primary B cells were isolated from naïve C57BL6/J mouse spleens (8–12 weeks of age) by negative selection using the EasySep Mouse B cell Enrichment Kit (Stem Cell Technologies) per manufacturer’s instructions. Following overnight stimulation with LPS at 20-25ug/mL, primary B cells were transduced with retroviruses supplemented with 5ug/mL polybrene by spinoculation at 2500rpm for one hour at 30°C. Cells were analyzed by flow cytometry using three wells per condition at days 2–5 post-transduction. Supernatants were collected and stored at -80°C for subsequent analysis by ELISA. IL10 in primary B cell supernatants was measured using the BD OptEIA Mouse IL-10 ELISA Kit (BD biosciences) per manufacturer’s instructions.
Blocking and detection antibodies were diluted in PBS supplemented with 2% FBS and 1mM EDTA. Splenocytes were blocked with anti-CD16/32 (BD bioscience) for 15 minutes on ice prior to surface staining for 30 minutes on ice. Antibodies used in this study: B220-Pac Blue, CD138-BV650 and -APC, CD45.1-FITC, CD3-PerCp, CD4-PerCp, CD8-PerCp, Thy1.1-PE, CD95-PE/Cy7, GL7-APC, and CD19-Pac Blue, -BV650, -FITC,- PerCp, -PE, -PE/Cy7, -APC, -APC/Cy7, -Alexa Fuor 594 (BD bioscience, eBioscience, or Biolegend). For intracellular cytokine staining, unstimulated cells were fixed with 4% paraformaldehyde/PBS solution after surface staining step. Cells were subsequently permeabilized using the BD Cytofix/Cytoperm Fixation/Permeabilization Kit (BD biosciences) per manufacturer’s instructions prior to staining with IL10-PE/Cy7 (Biolegend). Dead cells were labeled with fixable viability dye eFluor780 (eBioscience) per manufacturer’s instructions. Cells were analyzed on a BD LSRII flow cytometer and data was analyzed using FlowJo software.
M2-mCherry H2bYFP and M2-Thy1.1 H2bYFP bacterial artificial clones (BACs) were generated utilizing the galK selection method [39]. The M2 locus in the background of the previously described recombinant MHV68-H2bYFP genome [38] was replaced with galK gene as previously described [41]. Briefly, the galK cassette was amplified with primers flanked with 50bp sequence homology to the target sequence (5’-aggcgtgtttaaagaaaaagttatgttctgcgtta gcaccttcactgttacctgttgacaattaatcatcggca-3’ and 5’-agggggtttcaacaggcactagtctgatgaggtttcgtttt caggtaatgtcagcactgtcctgctcctt-3’) prior to electroporation of SW102 cells harboring the MHV68 H2bYFP BAC. The M2/galK intermediates exhibiting galactokinase activity were selected as previously described [41] and the presence of the desired insertion at the M2 locus was confirmed by restriction fragment length polymorphism (RFLP) analysis. M2-mCherry and M2-Thy1.1 sequences were amplified from the MSCV-M2-mCherry and MSCV-M2 vectors, respectively, with primers containing the 50bp homology arms and each cassette was electroporated into SW102 cells harboring the M2/galK intermediate. Positive recombinants were identified by PCR colony screen for the presence of the desired sequence and further evaluated by sequencing and RFLP analysis.
For experimental infections, female C57BL6/J mice at 6–8 weeks of age were purchased from Jackson labs (Bar Harbor, ME) and were infected between 8–12 weeks of age. Mice were housed and maintained at the Whitehead vivarium according to Emory University and IACUC guidelines. Mice were anesthetized with isoflourane prior to infection via the intranasal or intraperitoneal route with 1000PFU of the MHV68 viruses. Mice were sacrificed at the indicated timepoints by CO2 inhalation per AVMA guidelines and spleens were harvested and processed as described previously [38]. Splenocytes from individual mice were analyzed by flow cytometry and splenocytes from 4–5 mice per experimental group were pooled for ex vivo virus reactivation analyses. For adoptive transfer studies, female B6.SJL-Ptprca Pepcb/BoyJ mice (Jackson labs) were used as donors for primary B cell isolation and retroviral transduction as described above. At one day post-transduction,107 mock or transduced B cells were adoptively transferred into the peritoneum of naïve C57BL/6J mice at 8–12 weeks of age. Splenocytes were harvested from adoptive transfer recipients (3–4 mice/experimental group) at one and five days post transfer and analyzed by flow cytometry.
Reactivation of infectious virus from latently infected splenocytes was evaluated by utilizing a limiting dilution ex-vivo reactivation assay as previously described [67]. Briefly, infected splenocytes were pooled from 3–5 mice per condition and plated in 12 two-fold serial dilutions onto MEF monolayers in 96-well tissue culture plates. Cytopathic effect was scored for each well (24 wells/dilution) at 14–21 days post-explant.
GraphPad Prism software (San Diego, CA) was used to generate data graphs and perform statistical analyses. For line and bar graphs, the mean and standard deviation were plotted for each condition in triplicate. For scatter plots, each data point represents on animal and the horizontal bar represents the mean. Statistical significance between two conditions was determined by two-tailed unpaired t test with Welch’s correction. For three or more conditions, statistical significance was evaluated by one way analysis of variance (one way ANOVA) analysis followed by Bonferonni’s multiple comparisons post-tests. For reporting of absolute numbers of B cell populations in adoptive transfer recipients, 1 cell was added to all values in order to avoid the undefined logarithm of zero.
|
10.1371/journal.pntd.0006412 | Prevalence and risk factors of Rift Valley fever in humans and animals from Kabale district in Southwestern Uganda, 2016 | Rift Valley fever (RVF) is a zoonotic disease caused by Rift Valley fever virus (RVFV) found in Africa and the Middle East. Outbreaks can cause extensive morbidity and mortality in humans and livestock. Following the diagnosis of two acute human RVF cases in Kabale district, Uganda, we conducted a serosurvey to estimate RVFV seroprevalence in humans and livestock and to identify associated risk factors.
Humans and animals at abattoirs and villages in Kabale district were sampled. Persons were interviewed about RVFV exposure risk factors. Human blood was tested for anti-RVFV IgM and IgG, and animal blood for anti-RVFV IgG.
655 human and 1051 animal blood samples were collected. Anti-RVFV IgG was detected in 78 (12%) human samples; 3 human samples (0.5%) had detectable IgM only, and 7 (1%) had both IgM and IgG. Of the 10 IgM-positive persons, 2 samples were positive for RVFV by PCR, confirming recent infection. Odds of RVFV seropositivity were greater in participants who were butchers (odds ratio [OR] 5.1; 95% confidence interval [95% CI]: 1.7–15.1) and those who reported handling raw meat (OR 3.4; 95% CI 1.2–9.8). No persons under age 20 were RVFV seropositive. The overall animal seropositivity was 13%, with 27% of cattle, 7% of goats, and 4% of sheep seropositive.
In a multivariate logistic regression, cattle species (OR 9.1; 95% CI 4.1–20.5), adult age (OR 3.0; 95% CI 1.6–5.6), and female sex (OR 2.1; 95%CI 1.0–4.3) were significantly associated with animal seropositivity. Individual human seropositivity was significantly associated with animal seropositivity by subcounty after adjusting for sex, age, and occupation (p < 0.05).
Although no RVF cases had been detected in Uganda from 1968 to March 2016, our study suggests that RVFV has been circulating undetected in both humans and animals living in and around Kabale district. RVFV seropositivity in humans was associated with occupation, suggesting that the primary mode of RVFV transmission to humans in Kabale district could be through contact with animal blood or body fluids.
| Viral hemorrhagic fevers are known to cause high morbidity and mortality and pose a serious threat to human and animal populations in endemic countries. An outbreak of Rift Valley fever was detected in Kabale district in March, 2016 and identified the first human cases in Uganda since 1968. There was a need to perform a rapid assessment of the burden of Rift valley fever in Kabale district, identify undetected acute cases, identify risk factors associated with human disease, identify areas at high-risk or future infections, and to determine if this was a newly emerging infection or an endemic disease. Our study found the seroprevalence to be as high as 28% in humans and 36% in animals within some subcounties of Kabale district. Human seropositivity correlated with animal seropositivity, suggesting that animal to human transmission may be the predominant mode of virus spread. Our findings also suggest that this virus may have been endemic for many years prior to these human cases being identified.
| Rift Valley fever virus (RVFV) is a single-stranded RNA virus in the order Bunyavirales, and recently classified in the Phenuiviridae family [1] and the Phlebovirus genus. RVFV causes disease in humans and animals [2], and is transmitted by mosquitoes to livestock such as sheep, goats, and cattle [3]. Competent mosquito vectors include species from the Aedes, Culex, Anopheles, Eretmapodites, Mansonia, and Coquillettidia genera [4, 5]. Humans can become infected with RVFV when they come into contact with blood or body fluids of infected livestock while caring for sick animals, assisting with animal birth, or slaughtering livestock; or through bites of infected mosquitoes. Occupations at the greatest risk of RVFV infection include herdsmen and butchers [6–8]. Raw milk or meat consumption are potential sources of RVFV, although transmission via these routes has not been confirmed.
In livestock, RVFV infection can cause increased abortions and stillbirths, and high mortality in neonates and juvenile animals. As a result, RVFV outbreaks can lead to significant economic losses [3, 9]. In humans, infection can range from asymptomatic or a mild flu-like illness to more severe disease that includes hepatitis, retinitis, or encephalitis [10]. Approximately 1% of human cases develop hemorrhagic disease, and an estimated 1–2% of cases are fatal. However, during an outbreak in Saudi Arabia, case fatality was as high as 14% [10].
There is no specific treatment for RVFV infection in humans or animals, but supportive care may prevent complications and decrease mortality [2]. Currently, no RVFV vaccine is approved for use in either humans or animals in North America or Europe [3]. Some inactivated and live-attenuated vaccines have been developed and have been efficacious in animals [11, 12]. Development of human RVFV vaccines has been challenging due to the safety of the vaccine. An experimental human vaccine, TSI-GSD200, has shown some utility in laboratory workers, but has not been used extensively in other settings [11]. To better understand the utility of RVFV vaccines in a particular setting, the prevalence of disease in humans and animals must first be understood.
Rift Valley fever (RVF) outbreaks were first reported in East Africa in the 1930s [13]. In 1997–1998, large RVF outbreaks in northeastern Kenya were associated with El Niño rains and floods, resulting in many deaths in livestock and humans [14, 15]. Infection in humans was highly associated with contact with livestock and animal body fluids. RVF outbreaks were not reported again in East Africa until 2006–2007, when large numbers of humans and livestock were infected in Kenya, Somalia, Tanzania, and Sudan [16, 17]. Studies following these outbreaks reported that being a herdsman and handling or consuming products from infected animals were major risk factors for human infection. Additionally, outbreaks in East African countries mainly occurred in areas with poor soil drainage and flat lowlands that are less than 500 m above sea level [18, 19]. Also, RVF outbreaks in humans and animals following flooding have occurred in Sudan [20], Saudi Arabia [21], Yemen (2000–2001), South Africa, and Egypt [22].
Uganda is an East African nation that shares borders with Kenya and Tanzania, and to date no large RVF outbreaks have been reported since 1968, when 7 human cases occurred near Entebbe [23]. However, a 2013 serological survey of goats in Ssembabule, Mpigi, Masaka, and Mubende districts in Uganda showed a seroprevalence of 9.8%, suggesting RVFV circulation [24].
The Uganda Virus Research Institute (UVRI) in Entebbe has been implementing laboratory-based surveillance for viral hemorrhagic fevers (VHF) in Uganda since 2010, which includes testing for RVFV [25]. In March 2016, UVRI confirmed 2 acute human RVF cases in Kabale district in the southwestern region of Uganda [26]. These were the first human RVF cases identified in Uganda since 1968 [23]. Because not all human or animal RVF cases are symptomatic, RVFV infections are often undetected. Thus, UVRI, the Ugandan Ministry of Health (MoH), Ugandan Ministry of Agriculture Animal Industry and Fisheries (MAIF), and the United States Centers for Disease Control and Prevention (CDC) collaborated on a study to assess the seroprevalence of RVFV in humans and animals living in and around Kabale district. The objectives of the study were to determine the sero-prevalence of RVFV in both humans and animals in Kabale and surrounding districts, identify risk factors and high-risk areas for RVFV, determine if RVF is emerging or endemic, and identify unrecognized RVF cases that may be related to the 2016 outbreak.
Ethical approval for this study was granted from review by the UVRI Research Ethics Committee (UVRI REC: GC/127/16/03/551). Animal subjects work was conducted according to Uganda national guidelines and performed by officers from Kabale District and the Ministry of Agriculture, Animal Industries and Fisheries. The CDC National Center for Emerging and Zoonotic Infectious Diseases (NCEZID) Human Subjects office classified this project as non-research because the survey was a follow-up to the confirmed outbreak in Kabale district and the results of the study would assist the local health officials to target public health actions and interventions based on the serosurvey results (NCEZID: 032316TS).
Kabale district is located in the southwestern corner of Uganda. According to the 2014 Uganda census, it has an estimated population of 534,160 people, with the majority living in a rural setting (457,592 of 534,160; 86%) [27]. The altitude of Kabale ranges from 1,219 m to 2,347 m above sea level. Agriculture is an important source of revenue in this region. Most families own goats, sheep, cattle, and pigs. In addition to agricultural lands, Kabale district also has areas with high-altitude rain forests and savannahs.
From April 1–12, 2016, 34 locations in and near Kabale district were selected for inclusion in the serosurvey by a multidisciplinary team consisting of individuals from Kabale district, Uganda MoH, Uganda MAIF, UVRI, and CDC (Fig 1).
Four categories of people and animals were targeted for sampling based on perceived risk of RVFV infection. These were: 1) Animal slaughter house (abattoir) workers and the animals (cattle, goats, and sheep) slaughtered at the abattoir; 2) persons and animals from villages that had confirmed or probable human RVF cases; 3) persons and animals from villages considered at-risk for RVF due to geographic conditions (see below); and 4) persons and animals from randomly selected villages with no reported RVF cases. Additionally, animals were sampled from neighboring districts of Ntungamo and Kisoro. Since several studies have demonstrated increased risk of RVFV exposure in butchers, we selected this as a high-risk group for sampling. The research team worked with Kabale district health and veterinary officials to select the study sites determined to be at risk for RVF; these sites were identified based on the terrain, propensity for flooding, human and animal population density, cooperation from the community, and sharing of international borders.
When the team arrived at a selected location, local health workers assisted in recruiting participants for the serosurvey from the village by word of mouth. The questionnaire was written in English but was translated to the local language (Rukiga) and administered orally by local health workers (S1). Convenience sampling was used; all participants who presented to the sampling site were considered eligible for the study. All participation was voluntary and participants did not receive any compensation. Participants completed a consent form for the questionnaire, which was translated from English into local languages and explained by team members. Children older than 7 years were allowed to participate if a parent provided written consent. Blood samples were collected from all but 2 eligible participants identified.
Goats, sheep, and cattle that were brought to the collection site were sampled in relative proportion to the size of the herd at a given study site. Due to limited grazing areas in the Kabale region, few livestock owners maintain herds of more than 15 animals per species of cattle, sheep and goats. In cases where a livestock owner had less than 15 animals in a herd, all animals would be sampled. If an owner had more than 15 animals in a herd, only 25% of the herd was sampled. Only 3 of the households we sampled reported to have more than 15 animals per herd. We sampled all animals whose owners provided consent and presented their animals to the survey team. Both animals with a history of previous abortions or reported as having symptoms compatible with RVF, and apparently healthy animals of varied ages and sexes were sampled. Information collected about each herd included the general health of the herd, size of the herd, and grazing patterns. Because some humans and animals traveled from other subcounties to the location where the study was being administered, we did not collect both human and animal samples from all targeted subcounties.
One ~4 cc blood sample was collected from each human and animal participant for serological testing. Human specimens were tested by ELISA for anti-RVFV IgM and IgG, and animal specimens were tested for anti-RVFV IgG only. ELISA testing of both human and animal samples was performed at UVRI as previously described [28]. Human blood specimens that were IgM positive were subsequently tested by RT-PCR for RVFV-specific RNA targeting the L genome segment[29]. Briefly, following heat and detergent inactivation, specimens were tested by anti-RVFV-specific IgM and IgG ELISA using inactivated RVFV-infected Vero-E6 cell antigens, using 4 dilutions of each specimen (1/100, 1/400, 1/1600, and 1/6400). Titers and the cumulative sum optical densities of each dilution (SUMOD) minus the background absorbance of uninfected control antigen (adjusted SUMOD) were recorded. Samples were deemed positive if both the adjusted SUMOD and titer were above pre-established conservative cutoff values of ≥0.45 for IgM ELISA and ≥0.95 for IgG ELISA [28].
Questionnaire data were entered into a Microsoft Excel worksheet and analyzed using Stata 13 (StataCorp LP, College Station, TX, USA). The significance of human risk factors associated with RVFV seropositivity was first determined using bivariate analysis based on Pearson’s χ2. A p-value <0.05 was considered statistically significant. After completing the bivariate analysis, the significance of independent variables as predictors was further assessed using multivariate logistic regression. All risk factors found to be significantly associated with human RVFV seropositivity were included in the multivariate logistic regression model. A Pearson’s χ2 goodness of fit test was done on the final model [30]. To determine if animal seropositivity at the subcounty level was significantly associated with human seropositivity within that subcounty, a multivariate logistic regression was performed adjusting for risk factors found to be significant in the bivariate analysis.
A total of 655 persons participated in the serosurvey (Table 1). Participants were recruited at the Kabale town abattoir (n = 117; 18% of participants), from villages where a recent acute RVFV case had been identified (n = 237; 37%), and from villages with no recorded outbreaks (n = 293; 45%). Most participants (n = 396; 60%) were aged 20–49 years and had completed primary education (n = 360; 55%). The most common occupation listed was farmer or herdsman (n = 335; 52%), and most individuals owned animals (60%). Contact with animals was common, with 78% (n = 508) of participants reporting contact with animals in the past year.
Of all persons tested, 13% (88/655) were RVFV seropositive. Three (0.5%) persons had anti-RVFV IgM only, 78 (12%) had IgG only, and 7 (1%) had both IgM and IgG. Two individuals positive for RVFV IgM also tested positive for RVFV RNA by RT-PCR, suggesting active infection at the time of sampling. The 3 IgM-only positive individuals (a trader, a housewife, and a farmer) were all from the village in which one of the initial acute human RVFV cases was living but were not related to the initial acute RVF case. No persons under 20 years of age were RVFV seropositive, while 17% (n = 66) of individuals aged 20–49 years were seropositive (Table 2). Of individuals 50 years and older, 11% (n = 22) were seropositive. Butchers were the most likely to be RVFV seropositive, with 35% showing evidence of seropositivity. Other occupations evaluated for RVFV seropositive include farming at 10%, housewife at 8% (4/49), teacher at 18% (2/11) and trader at 12.5% (3/24). In the bivariate analysis, older age groups (χ2 = 14.4; p = 0.001), male sex (χ2 = 11.9; p = 0.001), occupation as a butcher (χ2 = 54.7; p < 0.001), history of slaughtering or butchering animals (χ2 = 23; p < 0.001), and preparing raw meat (χ2 = 13; p < 0.001) were all significiently associated with an increased risk of RVFV seropositivity (S1 Table).
In the multivariate logistic regression, being a butcher and handling raw meat were significantly associated with RVFV seropositivity, with an adjusted OR of 5.1 (95% CI 1.7–15.1; p = 0.003) and 3.4 (95% CI 1.2–9.8; p = 0.024), respectively (Table 3). Age, sex, slaughtering/butchering, and contact with animals through grazing were not significantly associated with RVFV seropositivity in the multivariate model.
Of all animals tested, 13% (133/1051) were RVFV seropositive. Seropositivity varied by species, sex, and age group among animals both in a bivariate analysis and the multivariate logistic regression (S2 Table; Table 4). Cattle showed significantly higher odds of being seropositive even after adjusting for age and sex (OR 9.1; 95% CI 4.1–20.5; p < 0.001). Adult animals and females also had significantly higher odds of being RVFV seropositive, with OR 3.0 (95% CI 1.6–5.6; p = 0.001) and OR 2.1 (95% CI 1.0–4.3; p = 0.04), respectively.
Human and animal seropositivity varied by subcounty (Figs 2 and 3), ranging 0–36% (standard deviation 11; mean 14%), whereas animal seropositivity ranged 4–28% (standard deviation 7; mean 12%) (S3 Table). Human RVFV seropositivity was significantly associated with close contact with cattle (P-value = 0.003), but shown not to be significant for contact with small ruminants (p-value = 0.06) (S1 Table). The sub-counties with the highest seroprevalence included the Kabale town council, where the main abattoir is located, and subcounties near bodies of water or wetlands.
The association between animal seropositivity and human seropositivity within a subcounty was examined using multivariate logistic regression, adjusting for contact with raw meat and occupation, because these were found to be significant risk factors in the univariate analysis. Human seropositivity within a subcounty was found to be associated with animal seropositivity, with OR of 1.1 (95% CI 1.0–1.1; p < 0.001).
Although RVFV had not been detected in Uganda since 1968, our study demonstrates that it has likely been enzootic in Kabale district for some time. Overall, we found evidence of RVFV seropositivity in 13% of humans and animals sampled. Our study also showed that butchers and those who handled raw meat were most likely to be RVFV seropositive.
Similar risk factors for RVFV seropositivity have been reported in previous studies. In a 2015 study in Kenya, LaBeaud and colleagues found that male sex, increased age, history of slaughtering livestock, history of malaise, and poor measured visual acuity were all factors for increased seropositivity [8, 31]. Although we did not find an association with sex and RVFV seropositivity after adjusting for other factors, such as occupation, we did find an association with being a butcher (i.e., someone who cuts meat either at home or at a slaughterhouse) and RVFV seropositivity. Previous studies also found that drinking raw milk may be associated with RVFV seropositivity [32, 33], but we were not able to find this association, likely because few individuals (36; 5%) reported drinking raw milk. Anecdotally, individuals reported not drinking raw milk due to concerns about Brucellosis infection. We did not find a significant association between age and seropositivity, but interestingly, no persons younger than 20 years were found to have evidence of RVFV infection. This may be because only 6 individuals below the age of 20 reported having close contact with livestock with the majority in this age group reporting to be attending school, greatly reducing their risk of exposure to potentially infected livestock.
Figs 2 and 3 show high seroprevalence, both in humans and animals, in 2 subcounties: Buhara subcounty, bordering Rwanda, and Bubare subcounty, near Kabale town. Kabale town contains the main abattoir, a likely source of RVFV infection in humans. However, Buhara and Bubare subcounties are connected by the primary North-South highway between Kabale town and the Rwandan border, a transportation corridor that could have served as a possible source of introduction of RVFV into Kabale district through livestock trade. A serosurvey of livestock conducted in Rwanda showed overall seropositivity of 16.8%, with districts closest to Tanzania showing the highest seropositivity and the 2 districts closest to the Ugandan border having the lowest seropositivity [34]. In addition, evidence of previous RVFV circulation and infection of livestock from samples collected in 2009 in goats from the Southeastern districts of Ssembabule, Mpigi, Masaka and Mubende shows a total seroprevalence of approx. 10% [24]. Although the testing methodologies employed for these samples was different than the one employed in our study, evidence of seropositivity in livestock may begin at earliest in 2009, or possibly earlier, and may suggest that RVFV was introduced into this region following outbreaks in neighboring countries and maintained through low level inter-epidemic transmission [34]. However, because there are no published reports of RVF seropositivity prior to 2009, we cannot be certain there was no widespread circulation of RVFV before that time. Our study suggests that RVFV transmission to humans in Kabale district is primarily due to exposure to the blood and body fluids of infected livestock, given that butchers and those handling raw meat were most likely to be RVFV-seropositive. Additionally, we found that human seropositivity was significantly associated with livestock seropositivity in each subcounty. Cattle density has also been previously associated with RVF seropositivity in RVFV models [35]. In our study, significantly more cattle were seropositive (27%) than goats (7%) or sheep (4%). This difference could be due to mosquito feeding behavior, as mosquitoes tend to select large and ornamented species [36]. Also, cattle are kept longer compared to goats and sheep hence are available to exposure to mosquito bites increasing their chances of being seropositive, however other factors such as mosquitoes species involved could be playing a role in Kabale district. For example, mosquitoes may have a preference for biting cattle compared to other ruminants. Furthermore, sheep and goats are usually kept indoors, especially at night, while cattle are rarely sheltered in Kabale district and thus exposed to nighttime-biting mosquitoes. Mosquitoes collected after outbreak investigations were mostly animal-specific feeders rather than human-specific (Julius Lutwama, personal communication), indicating that RVFV is primarily transmitted by mosquitoes within animal and human populations, however, humans are also infected from direct contact with infected animals. The mosquito species that were trapped in this region during the 2016 RVF outbreak include mainly Aedes and Culex species. Further investigations of the relative contribution of mosquitoes and livestock in RVFV transmission are needed.
The high seroprevalence in livestock seen in Kabale town is likely due to the main abattoir, where a majority of the animals in our survey were sampled. Animals, primarily from the neighboring subcounties, but also from more distant subcounties, are brought there for sale and slaughter. The high seroprevalence in animals can be attributed to the convergence of these animals from throughout the district in one location. This high seroprevalence in abattoir animals may be related to beef production dynamics in Uganda where older animals that are no longer of reproduction value are sold off by farmers for slaughter, hence more likely to be seropositive as stated above. The corresponding high seroprevalence seen in the abattoir workers can also be attributed to this, as well as daily occupational exposure to infected animal blood and body fluids.
Comparing our results with serological studies conducted in the neighboring countries of Kenya and Tanzania, where RVFV is endemic, provides some insight into our human serosurvey findings. In a study in coastal area of Kenya from 2009–2011 [37], RVFV seroprevalence in humans was lower (1.8%) than in our study. Similarly, the seroprevalence in humans was only 5.2% in Mbeya region in Tanzania [38], compared to 13% in our study. We think this is mainly because we collected samples after a confirmed outbreak or uptick in inter-epidemic transmission, unlike in the Kenya study. However, seroprevalence was higher in domestic ruminants in another study in Garissa, Kenya (27.6%) [39] than in our study (13%). Generally, seroprevalence in both animals and humans is expected to be higher in RVFV-endemic regions of Kenya and Tanzania than what we found in this study, but the risk factors identified were the same–mainly, contact with livestock. In a study in nearby Rwanda, the seroprevalence in cattle was 16.8%, slightly lower than 27% in our study [34]. However, these differences could be attributed to various factors, including the serological testing protocol used, which may differ in specificity and sensitivity. Additionally, our study identified local variations in RVFV seroprevalence, so the sampling strategy may also affect seropositivity results.
The extent of RVFV infection in Kabale district was greater than anticipated (as high as 36% in Bubare subcounty, and 28% in Buhara subcounty in animals) considering that the acutely identified human cases were the first to be laboratory confirmed in Uganda since 1968. The presence and percentage of RVFV-specific IgM, IgG and PCR-positive samples in persons in Kabale district indicates emergence of RVFV in Kabale is localized and that specific geographical locations may show varying levels of transmission or exposure to recently infected livestock. The level of IgG seropositivity indicates that RVFV has been circulating in the district for some time and further studies are needed to identify when this introduction may have occurred.
Our study had some limitations. Prior to initiating the survey, no reliable estimates of animal and human RVFV seropositivity existed, so we do not know how representative our samples are of the general population in Kabale district. In analyzing the correlation between human seropositivity and animal seropositivity by subcounty, we assume that seropositive humans were most likely to be exposed to RVFV within the subcounty of their residence, and did not take into account possible exposure in other locations.
Uganda has had a formal dedicated VHF surveillance program coordinated by UVRI and the MoH since 2011. During the past several years, 11 independent VHF outbreaks have been detected and confirmed by the UVRI VHF laboratory, including an outbreak of Marburg virus in 2012 that was first detected from a case in Kabale district [40]. Because of the previous experience and enhanced training and awareness, surveillance for VHF cases has become a priority, and healthcare workers were able to quickly identify the initial RVFV cases and immediately send samples for laboratory confirmation. Enhanced VHF and RVFV surveillance and health education must be continued given the potential for future RVFV re-emergence. Alongside this serosurvey, a knowledge, attitude practice (KAP) study was conducted in Kabale district and the findings of this KAP study has been used to design health education materials targeting different stakeholders [41]. For example, the education materials targeting farmers and butchers emphasize reporting of any sick animals to veterinarians, washing hands after touching raw meat or milk, cooking meat and milk thoroughly, using of mosquito bed nets and wearing of more protective clothing when working in high risk areas.
WHO recommends continued surveillance of RVFV during both outbreak and inter-epidemic periods [42]. Although RVFV rarely causes death in infected persons, the economic consequences due to loss of livestock and animal abortions can be devastating in agricultural communities [43]. Further studies are needed to fully understand the enzootic and endemic presence of RVFV in Kabale and surrounding districts. A longitudinal study is underway in Uganda to obtain samples from animals in Kabale district and throughout Uganda over several years to determine the endemicity and spread of RVFV over time. This will help determine the regions are at risk for RVFV outbreaks, so resources and surveillance efforts can be targeted to detect emergent cases and initiate any necessary control efforts.
|
10.1371/journal.pbio.1001738 | Stem Cell Transplantation in Traumatic Spinal Cord Injury: A Systematic Review and Meta-Analysis of Animal Studies | Spinal cord injury (SCI) is a devastating condition that causes substantial morbidity and mortality and for which no treatments are available. Stem cells offer some promise in the restoration of neurological function. We used systematic review, meta-analysis, and meta-regression to study the impact of stem cell biology and experimental design on motor and sensory outcomes following stem cell treatments in animal models of SCI. One hundred and fifty-six publications using 45 different stem cell preparations met our prespecified inclusion criteria. Only one publication used autologous stem cells. Overall, allogeneic stem cell treatment appears to improve both motor (effect size, 27.2%; 95% Confidence Interval [CI], 25.0%–29.4%; 312 comparisons in 5,628 animals) and sensory (effect size, 26.3%; 95% CI, 7.9%–44.7%; 23 comparisons in 473 animals) outcome. For sensory outcome, most heterogeneity between experiments was accounted for by facets of stem cell biology. Differentiation before implantation and intravenous route of delivery favoured better outcome. Stem cell implantation did not appear to improve sensory outcome in female animals and appeared to be enhanced by isoflurane anaesthesia. Biological plausibility was supported by the presence of a dose–response relationship. For motor outcome, facets of stem cell biology had little detectable effect. Instead most heterogeneity could be explained by the experimental modelling and the outcome measure used. The location of injury, method of injury induction, and presence of immunosuppression all had an impact. Reporting of measures to reduce bias was higher than has been seen in other neuroscience domains but were still suboptimal. Motor outcomes studies that did not report the blinded assessment of outcome gave inflated estimates of efficacy. Extensive recent preclinical literature suggests that stem-cell–based therapies may offer promise, however the impact of compromised internal validity and publication bias mean that efficacy is likely to be somewhat lower than reported here.
| Spinal cord injury is an important cause of disability in young adults, and stem cells have been proposed as a possible treatment. Here we systematically assess the evidence in the scientific literature for the effectiveness of stem-cell–based therapies in animal models of spinal cord injury. More studies reported effects on the ability to move (“motor outcomes”) than on sensation (“sensory outcomes”). Overall, treatment improves both sensory and motor outcomes, and for sensory outcome there was a dose–response effect (which suggests an underlying biological basis). Although more measures were taken to reduce the risk of bias than in other areas of translational neuroscience, unblinded studies tended to overstate the effectiveness of the treatment. The variability observed between the studies is not explained by differences in the stem cells used, but does seem to depend on the different injury models used to emulate human spinal cord injury. This suggests that the mechanism of injury should be an important consideration in the design of future clinical trials. Furthermore, open questions arise about the use of immunosuppressive drugs, and efficacy in female animals; these should be addressed before proceeding to clinical trial.
| Stem cells, from which all tissues can be generated, offer the potential to reconstitute tissues damaged by injury and disease. However, realising this potential will demand a detailed knowledge of the genetic and internal environmental cues that specify a cell's type, location, and interaction with its neighbours. It will also require a thorough understanding of stem cell behaviour in the context of lesioned or damaged tissues.
Stem cell transplantation was pioneered in the 1950s using haematopoietic stem cells to repopulate the bone marrow in patients with cancers of the blood and bone marrow [1]. Such is the success of this approach that an estimated 50,000 of these transplants are performed each year [2]. As understanding of stem cell biology has increased, so too has the ambition for restoring more complex tissues. In animal models, hepatocytes derived from stem cells can be engrafted into the damaged liver [3], and lineage-specific stem cells can repair damaged cornea [4],[5]. Recent studies also demonstrate the generation of artificial tissues with key features of complex solid organs including blood vessels [6], heart [7]–[9], lung [10], and kidney [11]. Even in the CNS, where the breadth of cell types and the complexity of their interactions are maximal, stem cell implants appear able to integrate into the existing circuitry [12]–[14]. In patients, lineage-specific stem cells have been reported to show efficacy in the regeneration of craniofacial bones [15] and of damaged cornea [5].
Integration into the host environment and tissue reconstruction are not the only potentially relevant biological effects of stem cells. Immunomodulatory effects of stem cells appear to reduce rejection of kidney transplants [16],[17], corneal allografts [18], and composite tissue hemi-facial allografts [19]. In the CNS, stem cells are reported to provide immunomodulatory and neuroprotective effects in models of diseases as disparate as retinopathy [20], neuronal ceroid lipofuscinosis [21], motor neuron disease [22],[23], Parkinson's disease [24], multiple sclerosis [25],[26], stroke [27]–[29], and spinal cord injury [30],[31].
There is now considerable preclinical literature on the possible benefits of stem-cell–based therapies following traumatic spinal cord injury. Stem cells may assist recovery through limitation of secondary injury, re-myelination, formation of new neuronal connections, and alteration of the inhibitory environment. However, it is unclear which type of cells and from what source are best to implant, how many are needed, whether immunosuppression should be used, and whether the implanted cells need to be modified to enhance particular desirable characteristics. It is also unclear whether the magnitude of integrative and protective effects is large enough to be potentially clinically meaningful. We also do not know whether reports of efficacy in animal models are potentially biased in favour of positive results.
Here, we report a systematic review, meta-analysis, and meta-regression of data from controlled in vivo studies testing the efficacy of stem cells as a treatment in animal models of spinal cord injury. Our objectives are (i) to establish a summary estimate of the efficacy of stem cells in animal models of traumatic spinal cord injury, (ii) to ascertain the conditions under which animal experiments demonstrate greatest efficacy, and (iii) to determine any effect of study quality on reported efficacy.
Electronic searching identified 156 full publications that met our prespecified inclusion criteria (Table S1). Forty-five different stem cell types had been investigated, from which over a third were derived from adult rats. The duration of experiments following the induction of SCI ranged from 7 d to 6 mo.
One publication [32] with two individual comparisons involving 36 animals reported the effect of autologous bone marrow stromal cells on motor score. We included this publication in the overall assessment of the prevalence of the reporting of measures taken by the original authors to reduce the risk of bias in their experiments. However, because this was the only paper to report the effects of autologous (rather than allogeneic) stem cells, we did not analyse this further, focussing instead on allogeneic stem cells.
One hundred and fifty-five publications reported the effect of allogeneic stem cells in 317 individual comparisons; 380 different motor outcomes were reported and because more than one motor outcome was reported for some individual comparisons we nested (see Methods) these into 312 individual comparisons involving 5,628 animals (Figure 1A). Six different tests were used to assess motor score: the Basso, Beattie and Bresnehan locomotor rating scale (BBB; [33]), the Basso mouse scale (BMS; [34]), the Tarlov scale [35], the forelimb placing test [36], the staircase test [37], and the mouse hind limb motor score [38]. Sixty-one sensory outcomes were reported; we excluded six outcomes that tested sensation in unaffected limbs. In 10 outcomes that used the same test at different intensities in the same cohort of animals, we only included the median intensity. Therefore, we report data on sensory outcome reported in 45 experiments nested into 24 comparisons using 473 animals (Figure 1B). In 18 cohorts both motor and sensory outcomes were reported.
We describe the reporting of study quality checklist items reported for each included publication in Table S2. All studies included in this analysis came from peer-reviewed papers; while we identified a number of potentially relevant abstracts, none of these reported data in sufficient detail to be included. One hundred and eleven of 156 publications (71%) reported compliance with animal welfare regulations, and 25 (16%) reported whether or not a conflict of interest existed.
Allocation concealment was reported in 14 of 156 publications (9%). Random allocation to treatment group (72, 46%) and blinded assessment of outcome (72, 46%) were reported more frequently in these publications than in the modelling of other neurological disorders [39]–[42], but the reporting of a sample size calculation (less than 1%) was consistent with the proportions observed elsewhere (Table 1). No publication reported all four of these measures to minimise bias.
Despite the reported benefits of hypothermia in SCI [43]–[45], in other animal models of neurological disease [46] and in humans with ischaemic neurological injury [47],[48], only 33 (21%) studies described controlling temperature during the experiments.
There were only sufficient data to assess publication bias in studies using allogeneic stem cells where outcome was measured as a motor score. Small study bias was suggested with asymmetry of the funnel plot (Figure 2A) and Egger regression (Figure 2B) but not by Trim and Fill.
As expected, our search identified a diverse range of experiments. There was substantial between-study heterogeneity for studies using allogeneic stem cells both where outcome was measured as a motor score [heterogeneity (χ2) = 9,735, 311 degrees of freedom (df), p<10−99; effect size, 27.2% improvement in outcome [95% confidence interval, 25.0%–29.4%]; 312 comparisons) and as a sensory outcome (χ2 = 183, df = 23, p<10−26; effect size, 26.3% [7.9%–44.7%]; 24 comparisons).
Systematic review and meta-analysis have helped identify biases within clinical trials [49], providing an impetus to improve standards [50]. This approach offers similar benefits for animal studies [28],[41],[51] by describing the impact of biological and experimental factors on reported efficacy in a systematic and transparent summary of all available data. This allows judgement of the extent to which conclusions are at risk of bias [52]. In this study we apply these techniques to provide a detailed systematic analysis of the animal literature describing stem-cell–based therapies in spinal cord injury.
Overall, treatment with allogeneic stem cells improves both motor and sensory outcome after spinal cord injury by around 25%, but with important differences between the two datasets. Because of the amount of data, conclusions relating to motor outcome (5,628 animals) are probably more robust than those relating to sensory outcomes (473 animals). For both outcomes there was a broad range of experimental approaches, reflected in the high levels of heterogeneity seen. This is typical for systematic reviews in animal studies and validates our choice of a random effects model, and our summary estimates should be considered as the average efficacy rather than the best estimate of a single “true” efficacy. Interestingly, improvement in sensory outcome seems to be sensitive to differences in factors relating to treatment (i.e., stem cell biology), while motor outcome appears to be more sensitive to factors relating to the lesion and the outcome measure used, and to be less dependent on the biological features of the stem cells used.
Evidence supporting a dose–response relationship for sensory outcome suggests the presence of a biologically plausible effect. We observed that prior differentiation of the implanted cells was associated with larger effects. Where the influence of cell differentiation was formally studied, a relationship with outcome was observed [53]. This suggests that optimal efficacy might be seen when cells have some lineage specificity but before final cell type commitment has occurred. For sensory outcome, studies where cells were delivered intravenously, rather than directly into the injured spinal cord, were associated with significantly larger effects. This suggests either that systemic changes may mediate the effects of stem cells or that local implantation may create additional injury that masks the benefit provided by stem cells.
We did not see a dose–response relationship for motor outcomes, even where we limited our analysis to a more homogenous subset of experiments. It may be that there is no dose–response effect or that the doses used in these experiments were all large enough to generate maximal responses. Where dose response was formally studied the authors found increasing benefit from doses as low as 10,000 implanted cells [54], and the median number of implanted cells in comparisons reporting motor outcomes was 250,000.
Immunosuppression with cyclosporine A was associated with increased efficacy in a systematic review of stem cells in focal cerebral ischaemia [28], and it is therefore interesting that in spinal cord injury both cyclosporine A and FK506 are associated with reduced efficacy. This suggests that any beneficial effect of immunosupressants in promoting the survival of transplanted cells is outweighed by other factors, such as effects on stem cell biology or intrinsic repair mechanisms. Unfortunately, because of the univariate nature of our analyses we are unable to determine a “benefit–risk ratio” for the use of immunosuppression. However, there are studies that indicate that bone-marrow–derived stem cells are able to produce compartmentalised inflammatory lesions [55],[56]. The mechanisms behind this observation are not understood, yet there are rising concerns that unwanted inflammatory-driven side effects, such as neuropathic pain, might limit the “usefulness” of gained motor function.
For motor outcome, the neurobehavioural test used (Figure 3A) accounted for most of the observed heterogeneity. The BBB locomotor rating scale was used in 70% of animals. In the more focussed analysis of rat allogeneic, midthoracic impact injury, using BBB as an outcome, studies that used other behavioural tests in addition to the BBB reported smaller effect sizes for the BBB. This may be a manifestation of outcome reporting bias; if the outcome on the BBB is smaller than expected, investigators might also report the outcome on other tests where the effect was larger; if the effect measured using the BBB was considered “sufficient,” there might be less motivation also to report outcomes using other measures, particularly if these were smaller than seen using the BBB.
Overall, there was no improvement in motor outcome where this was assessed using the staircase or mouse hind limb motor score tests. However, these accounted for a small proportion of the overall dataset, and so these results should be interpreted with caution.
Efficacy was strongly associated with both the location of and the methodology used to create the injury. The largest effect was seen with lower thoracic and lumbar lesions and when the spinal cord was lesioned by hemisection or transsection rather than contusion or compression.
The use of isoflurane anaesthesia at SCI induction was associated with substantial improvement in sensory outcome; in the overall motor analysis, there was no effect, but in the more homogenous restricted analysis, isoflurane was again associated with substantially larger effects. Again, this contrasts with findings in focal cerebral ischaemia and suggests that, despite interest in a general paradigm of “neuroprotection,” these conditions are in certain respects biologically very different. However, these findings are based on a small number of individual comparisons and should be interpreted with caution.
The sex of the experimental animal accounted for a large proportion of the observed heterogeneity in both the sensory and motor analyses. For the motor analyses, this seems to be the influence of abnormally high effect sizes reported in studies where either the sex of the animals used was not reported or where “both sexes” were used. For sensory outcome, studies using male animals led to significantly higher estimates of effect with no clear benefit detected in female animals.
Thirty percent of animals in our dataset were treated with stem cells at the time of injury. Although this may be helpful in the biological assessment of stem cell therapies, it is of limited clinical relevance. The time of administration, although important with regard to translation to a clinical setting, had no significant impact on the effects reported. This appears to be somewhat unlikely, and our findings may mask different efficacies of different stem cell approaches at different times—those with more neuroprotective characteristics perhaps being more effective when given early, and those with more influence on neuroregeneration and repair being more effective when given late.
We found that the prevalence of reporting of randomisation and blinded assessment of outcome was higher than that reported in the modelling of other neurological disorders, suggesting more rigour in the conduct of these studies [39]–[42]. Other markers of internal validity, such as sample size calculations, were rarely reported (Table 1). The lack of an a priori sample size calculation increases the risk that group sizes were increased during the experiment, in light of analysis showing borderline nonsignificant results; this is an important potential source of bias. It is of course possible that some authors had taken measures to reduce bias but did not report them; this underlines the importance of reporting guidelines [57],[58].
For the larger motor dataset, both publication bias (Figure 2B) and failure to report blinding (Figure 3H) were both associated with a significant overestimation of overall effect size; there was no apparent impact of a failure to report randomisation. In the Egger regression (Figure 2B) removal of the two most extreme data points did not change the interpretation that publication bias was present (not shown).
Stratification of the data to determine the effect of the above facets of experimentation is desirable. However, no publication randomised, blinded assessment of outcome, concealed allocation, and performed a sample size calculation and only 20 individual comparisons came from papers describing three of the four. Therefore, we subanalysed the 25% of the motor dataset that reported having both randomised and blinded.
In this subanalysis the characteristics of the animal model still have more impact than the type of cells implanted. However, there were differences, but the reductionist approach of this subanalysis does raise the possibility that these might be false positives due to loss of power. The type of cell culture medium and type of cell manipulation prior to implantation appear to have an impact, but it should be noted that in both cases it is the experiments where the precise conditions are “unknown” that report the greatest effect. There is no obvious biological explanation for this. It may be that a failure to report such details is a surrogate indication that such work is generally of lower quality, and therefore at greater risk of bias.
Immunosuppression is no longer identified as accounting for a significant proportion of the heterogeneity. However, the effect size in cyclosporine-A–treated animals (mean, 24.3; 95% CI, 13.2–35.3) is the same as in animals where no immune suppression was used (mean, 24.9; 95% CI, 18.3–31.6). This appears to confirm that immune suppression offers no advantage in experiments using allogeneic implants to treat SCI.
Intriguingly, in the subanalysis a dose–response relationship does emerge. As the mean number of cells implanted is 6.3×105 rather than 7.4×108 in the full motor dataset, this is consistent with the hypothesis that such an effect was previously masked by a ceiling effect.
The study protocol is available at www.camarades.info/index_files/Protocols.html. A completed PRISMA checklist and flow diagram for this systematic literature review can be found in Text S1.
We define a “publication” as a discrete piece of work (including abstracts); each publication may report data from a number of experiments. Each experiment may describe outcome in a number of different experimental cohorts, and the contrast between outcomes in a single treatment cohort with that in a control cohort we define as an “individual comparison.” We define “nesting” as combining the effect sizes from different functional outcomes measured in the same cohort of animals to give a single summary estimate of effect in that individual comparison (a nested individual comparison).
Using prespecified inclusion and exclusion criteria we identified all publications reporting relevant experiments (see below) by searching (December 2011) three electronic databases (PubMed, EMBASE, and ISI Web of Science) using the search strategy “(stem cell OR stem OR haematopoietic OR mesenchymal) AND (spinal cord injury OR hemisection OR contusion injury OR dorsal column injury OR complete transection OR corticospinal tract injury),” with search results limited to those indexed as describing animal experiments.
Two investigators (A.A. and E.S.) independently reviewed retrieved publications. We included experiments where functional outcome in a group of animals exposed to traumatic spinal cord injury and treated with allogeneic or autologous stem cells was compared with functional outcome in a control group of animals. We excluded individual comparisons that did not report (or where we could not calculate) the number of animals, the mean outcome, or its variance in each group. We excluded experiments where interventions such as growth factors were used to mobilise endogenous stem cells or where nontraumatic models of spinal cord injury were used.
From each individual comparison we extracted data for reported outcomes. This included extraction of mean and variance data from each cohort exposed to an intervention (controls and active therapy) and from sham cohorts of normal (unlesioned and untreated) animals, and by imputation where the performance of a normal animal could be imputed from the description of the scoring scale. Stem cells were characterised as “autologous” where cells were extracted from an animal, might be manipulated in some way, then returned to the same animal; or “allogeneic” where embryonic or adult cells derived from a different animal were administered to a recipient animal. Where a publication reported more than one experiment, or where an experiment reported more than one individual comparison (for instance, increasing numbers of stem cells transplanted), we considered these separately and extracted data for each, correcting the weighting of these studies in meta-analysis to reflect the number of experimental groups served by each control group. Where different functional outcomes were reported in a single cohort of animals, we combined these outcomes using fixed effects meta-analysis (nesting), to give a summary estimate of functional outcome in that cohort, described here as a comparison. Where a test involved exposing the animal to increasing intensities of the same stimulus (for instance, in allodynia testing), we used data for the median intensity. For sensory tests, only data for stimulation distal to the lesion were included. Where functional outcome was measured at different times, we extracted data for the last time point reported.
Study quality was assessed using a checklist adapted from good laboratory practice guidelines for in vivo stroke modelling [59] and the CAMARADES quality checklist [60]. The checklist comprised (i) publication in a peer-reviewed journal, (ii) statements describing control of temperature, (iii) randomisation to treatment group, (iv) allocation concealment, (v) blinded assessment of outcome, (vi) avoidance of anaesthetics with known marked intrinsic neuroprotective properties, (vii) sample size calculation, (viii) compliance with animal welfare regulations, and (ix) whether the authors declared any potential conflict of interest.
For each individual comparison, we calculate a normalised effect size [normalised mean difference) as the percentage improvement (“+” sign) or worsening (“−” sign) of outcome in the treatment group using the following formula:where and are the mean reported outcomes in the control and treatment group, respectively, and is the mean outcome for a normal (unlesioned and untreated) animal. In this calculation, the score achieved by the sham animals acts as the “fixed zero value” or baseline allowing the difference between the sham and treatment groups to be expressed as a ratio. This ratio takes into account differences in the “direction” of individual neurobehavioural scales.
Its corresponding standard error was calculated using:where refers to the number of animals in the control group and refers to the number of animals in the treatment group. and are the normalised standard deviations for the control and treatment group, respectively. These were calculated using the formulae:where SDc and SDrx are the reported standard deviation for the control and treatment group, respectively.
We then used DerSimonian and Laird random effects weighted mean difference meta-analysis to calculate a summary estimate of effect size; results are presented as the percentage improvement in outcome and its 95% confidence intervals. The variability of the outcomes assessed is presented as the heterogeneity statistic (χ2) with n−1 degrees of freedom.
The analysis was stratified according to (i) the approach to stem cell therapy (allogeneic, autologous, embryonic, source of cells, ex vivo manipulation), (ii) biological factors (number of cells, time and route of administration, time of assessment of outcome), (iii) aspects of study design (anaesthesia, species of animal, immunosuppression, model and severity of spinal cord injury), and (iv) elements of study quality.
The extent to which study design characteristics explained differences between studies was assessed using meta-regression with the metareg function of STATA/SE10, and the significance level was set at p<0.05. The meta-regression was univariate rather than multivariate; and we calculated adjusted R2 values (a measure of how much residual heterogeneity is explained by the model) to explain the proportion of the observed variability in the observed effect size for a group of experiments explained by variation in the independent variable in question [61].
We sought evidence of publication bias using a funnel plot, Egger regression, and Trim and Fill [62]. A detailed description of the statistical methods used for meta-analysis and meta-regression can be found in [63].
|
10.1371/journal.pntd.0005548 | The gut microbiota as a modulator of innate immunity during melioidosis | Melioidosis, caused by the Gram-negative bacterium Burkholderia pseudomallei, is an emerging cause of pneumonia-derived sepsis in the tropics. The gut microbiota supports local mucosal immunity and is increasingly recognized as a protective mediator in host defenses against systemic infection. Here, we aimed to characterize the composition and function of the intestinal microbiota during experimental melioidosis.
C57BL/6 mice were infected intranasally with B. pseudomallei and sacrificed at different time points to assess bacterial loads and inflammation. In selected experiments, the gut microbiota was disrupted with broad-spectrum antibiotics prior to inoculation. Fecal bacterial composition was analyzed by means of IS-pro, a 16S-23S interspacer region-based profiling method. A marked shift in fecal bacterial composition was seen in all mice during systemic B. pseudomallei infection with a strong increase in Proteobacteria and decrease in Actinobacteria, with an increase in bacterial diversity. We found enhanced early dissemination of B. pseudomallei and systemic inflammation during experimental melioidosis in microbiota-disrupted mice compared with controls. Whole-genome transcriptional profiling of the lung identified several genes that were differentially expressed between mice with a normal or disrupted intestinal microbiota. Genes involved in acute phase signaling, including macrophage-related signaling pathways were significantly elevated in microbiota disrupted mice. Compared with controls, alveolar macrophages derived from antibiotic pretreated mice showed a diminished capacity to phagocytose B. pseudomallei. This might in part explain the observed protective effect of the gut microbiota in the host defense against pneumonia-derived melioidosis.
Taken together, these data identify the gut microbiota as a potential modulator of innate immunity during B. pseudomallei infection.
| Melioidosis is a common cause of community-acquired pneumonia and sepsis in Asia. The causative agent, Burkholderia pseudomallei, is listed as a potential bioterror weapon. The intestinal microbiota has been suggested to be a modulator of innate immune defenses against bacterial infections. Here we investigated in mice whether the intestinal microbiota affects the clinical course of melioidosis and vice versa. The composition of the gut microbiota changed strongly during melioidosis. Mice with a disrupted gut microbiota showed increased bacterial dissemination when compared with controls following intranasal infection with B. pseudomallei, indicating that the intestinal microbiota acts as a protective factor in host defense against melioidosis. Macrophages in microbiota-disrupted mice showed a diminished capacity to phagocytose B. pseudomallei. Further research is needed to explore whether this knowledge could be used to the advantage of patients.
| Melioidosis is a frequent cause of community-acquired sepsis in Southeast Asia and northern Australia [1, 2]. Pneumonia is the presenting symptom in most adult patients [3] and results in a rapidly progressive illness with a high mortality up to 40% [1, 3]. The disease is caused by Burkholderia pseudomallei, a facultative intracellular Gram-negative bacterium that is commonly found in the soil from countries located between 20° north latitude and 20° south latitude [1, 2, 4]. Due to its high lethality, poor sensitivity to antibiotics, wide availability and easy dissemination, it has been classified as a Tier 1 biological threat agent. The global burden of melioidosis is probably much larger than previously anticipated: it was recently estimated that each year 165,000 (95% credible interval 68,000–412,000) people suffer from this debilitating disease resulting in 89,000 (36,000–227,000) fatalities [5]. Melioidosis is probably underreported in as many as 45 countries due to a lack of adequate diagnostic facilities [5]. In the near future, the management of patients may be compromised by emergence of resistance due to increased use of antibiotics in endemic regions [6]. Thus, there is an urgent need to better understand the pathogenesis of melioidosis.
The intestinal microbiota not only provides direct colonization resistance against invading pathogens, but is increasingly recognized as an important modulator of systemic immunity [7–9]. It was first suggested by Clarke and colleagues that bacterial cell wall components such as peptidoglycan are translocated into the bloodstream and at distant sites ‘prime’ immune effector cells [10]. This way, the effectiveness of bone-marrow derived neutrophils in killing pathogens such as Streptococcus pneumoniae and Staphylococcus aureus is increased [10]. Subsequent studies could demonstrate a protective effect of a healthy microbiota in a variety of in vivo murine models of infection: S. aureus, Pseudomonas aeruginosa or Klebsiella pneumoniae pneumonia [11–13] and Listeria monocytogenes or Escherichia coli induced sepsis [14, 15]. In line, we recently demonstrated that the gut microbiota plays a protective role in pathogenesis of pneumococcal pneumonia by enhancing primary alveolar macrophage function [16].
To the best of our knowledge, the role of the gut microbiota in the host defense against melioidosis has never been investigated. The importance of this subject is underscored by the notion that melioidosis has a notoriously protracted course for which cure can only be achieved through long-term antibiotic therapy. The minimum of two weeks of intravenous antibiotics followed by three months of oral antibiotics [1, 2] will have a profound effect on the microbiota. We hypothesized that a healthy microbiota supports the host defense against B. pseudomallei infection. In order to address this question, we made use of our well-established murine model of pneumonia-derived melioidosis [17, 18] and in selected experiments disrupted the gut microbiota with oral antibiotics before infection following a standard protocol [10, 16]. We show that the intestinal microbiota changes significantly during melioidosis, independent of antibiotic treatment. Secondly, we show that antibiotic disruption of the intestinal microbiota is associated with a less effective innate immune defense against experimental B. pseudomallei infection.
To first obtain insight into the composition of the intestinal microbiota during melioidosis, we inoculated mice intranasally with live B. pseudomallei to induce pneumonia-derived melioidosis and collected fecal pellets at baseline (t = 0) and after 72 hours, when all mice had symptoms of systemic infection. The gut microbiota was analysed by IS-pro technique, using the number of nucleotides between the genes for the 16S and 23S ribosomal subunits in bacterial DNA as a unique classification characteristic [19, 20]. Infection with B. pseudomallei was associated with profound changes in the composition of the intestinal microbiota (Fig 1). Clustering analysis, by unweighted pair group method with arithmetic mean (UPGMA) on cosine distances of all samples, resulted in separation of all pre- and post-infection samples (t = 0 vs t = 72), indicating that pre- and post-infection samples from each mouse were highly dissimilar. In all mice, a strong increase in Proteobacteria was seen as well as a decrease in Actinobacteria. The composition of Bacteroides and Firmicutes also changed in strikingly similar patterns. Of note, B. pseudomallei was not detected in any fecal sample. Total microbial diversity was significantly increased (p = 0.006), mostly due to increased diversity of Bacteroides (p = 0.055) and Proteobacteria (p = 0.007) (Fig 1).
To investigate whether gut microbiota composition impacts on host defense during melioidosis, we pre-treated mice with broad-spectrum antibiotics in drinking water in order disrupt the intestinal microbiota (Fig 2A) [10, 16]. We then inoculated mice intranasally with live B. pseudomallei (150 colony forming units (CFU), LD50) [17, 18] and sacrificed them after 24 or 72 hours. The antibiotic treatment caused dramatic changes in the intestinal microbiota compared to untreated control mice, with a marked reduction in the number of species (Fig 2B). Relative to control mice, antibiotic pre-treated mice displayed significantly increased bacterial loads in lung and liver 24 hours after infection (Fig 2C–2E). Bacterial loads in blood and broncho-alveolar lavage fluid (BALF) were not affected (Fig 2D and S1 Fig). To determine whether the effect of gut microbiota disruption on bacterial growth was dependent on the infectious dose, we next infected mice with 500 CFU B. pseudomallei (LD100). Using this higher infectious dose, the observed increase in bacterial dissemination in antibiotic treated mice was also present at the early time-point following infection (Fig 2F–2H). In addition, we tested whether the observed effects were specific for this combination of antibiotics by performing the same experiment, using only metronidazole and ampicillin in drinking water (S2 Fig). Similar to the previous experiments, we again observed increased bacterial loads after 24 hours in lungs of antibiotic pre-treated mice compared to controls—indicating that the intestinal bacteria targeted by these two antibiotics are involved in the observed effects.
Having found an inoculum-dependent effect of antibiotic pre-treatment on bacterial growth both at the primary site of infection and at distant sites, we next studied the impact of antibiotic pre-treatment on local and systemic cytokine release in mice infected with 500 CFU B. pseudomallei. In lung homogenate, cytokine levels were similar between groups at all time-points (Table 1). Plasma levels of tumor necrosis factor (TNF)-α and interferon (IFN)-γ however were significantly increased in mice with a disrupted gut microbiota, 72 hours after infection (Table 1).
To evaluate whether the above findings would lead to impaired survival in the experimental group, we followed groups of 20 mice for 14 days after intranasal inoculation with 150 CFU B. pseudomallei. The lower dose was chosen since this LD50 [17, 18] would allow to demonstrate a potential detrimental effect of gut microbiota depletion. Mice in the antibiotic pre-treated group showed a trend toward increased mortality but this did not reach statistical significance (Fig 3A). Likewise, a clinical observation score reflected a trend towards increased morbidity in the experimental group (Fig 3B). The marked organ injury in this model of melioidosis is reflected by elevated plasma markers of hepatocellular damage (aspartate aminotranspherase, AST and alanine aminotranspherase, ALT), renal failure (urea) and general cellular damage (lactate dehydrogenase, LDH), especially shortly before mortality occurs [17, 18] (Fig 3C–3F). However, in line with survival, no significant differences in these parameters were observed, indicating a limited influence of gut microbiota disruption on the extent of organ damage.
As the microbiota has been reported to be an important regulator of neutrophil homeostasis [10, 14, 15, 21] and neutrophils play an essential role in the host defense against melioidosis [22, 23], we hypothesized that these might play a role in the observed differences. 72 hours after infection, all mice showed extensive lung infiltrates characterized by neutrophil influx, necrosis, bronchitis, endothelialitis and oedema (Fig 4A and 4B). However, when we analysed HE-stained lung tissue sections using a semi-quantitative pathology scoring system, no differences were found between control and antibiotic pre-treated mice (Fig 4C). Quantification of a Ly-6GC staining demonstrated a similar pulmonary influx of neutrophils in both groups (Fig 4D–4F). In line, a similar pulmonary influx of cells was observed in BALF during melioidosis (Fig 4G). Equal neutrophil degranulation was confirmed by lung myeloperoxidase levels in control and antibiotic treated mice after infection with B. pseudomallei (Fig 4H). Similar results were obtained for mice inoculated with 150 or 500 CFU; only the latter are shown. Lastly, since a healthy microbiota is proposed to stimulate granulopoiesis [10, 14, 15, 21], we studied neutrophil numbers in bone marrow and blood of naïve control and antibiotic treated mice; however, we did not find any differences (Fig 4I).
To obtain insight into the mechanism by which the gut microbiota exerts its effects during pneumonia-induced melioidosis, we investigated the effect of antibiotic microbiota disruption on lung transcriptomes. Comparing lung transcriptomes of uninfected intestinal microbiota disrupted mice to control mice revealed 40 significantly altered genes (Fig 5A, 21 genes under-expressed and 19 genes over-expressed in antibiotic pre-treated mice). Ingenuity pathway analysis revealed that genes with elevated expression in antibiotic treated mice significantly enriched several cellular biological pathways, including acute phase response signaling, coagulation system and, notably, IL-12 signaling and production in macrophages as well as production of nitric oxide (NO) and reactive oxygen species (ROS) in macrophages (Fig 5B). Of note, these gene expression differences were not biased by altered neutrophil infiltration (S3 Fig). However, since the analysis was performed on whole lung tissue, it is possible that the genes in these pathways were upregulated in other cell types than macrophages. Altogether, these data suggest that disruption of the intestinal microbiota by antibiotic treatment may impact on lung homeostasis, with macrophages more likely influenced.
As alveolar macrophages are crucial in the first line of defense during pneumonia and are important in the innate immune response in melioidosis [24], we further studied the influence of gut microbiota disruption on the function of alveolar macrophages. Responsiveness of alveolar macrophages derived from gut microbiota-disrupted mice towards PAM3CSK4, LPS or heat-killed B. pseudomallei was not different from controls in terms of proinflammatory cytokine production (Fig 5C and 5D). In line, no differences in metabolic profiles of alveolar macrophages derived from naïve control and antibiotic pre-treated mice were observed; for this we used extracellular flux technology, which enables assessment of mitochondrial function in live cells, simultaneously measuring oxygen consumption and glycolysis (S4 Fig). In a last set of experiments, we investigated the influence of gut microbiota disruption on the capacity of alveolar macrophages to phagocytose B. pseudomallei, since an effect of the gut microbiota hereon has been described previously in a setting of S. pneumoniae, S. aureus and K. pneumoniae infection [10, 12, 16]. Cells from BALF were plated and adhering cells were incubated with heat-killed, FITC-labeled B. pseudomallei, after which the phagocytosis index was determined by flow cytometry (Fig 5E). To confirm this finding, we inoculated mice with heat-killed, FITC-labeled B. pseudomallei and performed broncho-alveolar lavage three hours later, followed by flowcytometry. Again, we found that alveolar macrophages derived from antibiotic pre-treated mice had a diminished capacity to phagocytose B. pseudomallei when compared with controls (Fig 5F and 5G). These data suggest that an unperturbed gut microbiota enhances the capacity of alveolar macrophages to phagocytose B. pseudomallei in vivo. The observed effect was compartment specific; in contrast to alveolar macrophages, ex vivo phagocytosis capacity of blood neutrophils, peritoneal macrophages and bone-marrow derived macrophages derived from gut microbiota-disrupted mice was equal compared with controls, as well as cytokine production (S5 Fig). Of note, we did not observe any differences in pulmonary microbiota composition between control- and antibiotic treated mice (S6 Fig).
To the best of our knowledge, this study is the first to investigate the role of the intestinal microbiota during melioidosis. Our data suggest a bidirectional interplay between intestinal microbiota and innate host defenses against B. pseudomallei. Firstly, we observed significant changes in fecal microbiota composition during melioidosis, independent of antibiotic treatment. A strikingly similar pattern of increased Proteobacteria, decreased Actinobacteria and increased diversity was observed in all mice. Secondly, a well-balanced gut microbiota appears to have a protective effect during melioidosis, especially when B. pseudomallei has its first encounter with alveolar macrophages in the lung. Antibiotic disruption of the intestinal microbiota affects the capacity of these cells to internalize the pathogen, which was associated with increased bacterial proliferation and dissemination after 24 hours. In this model, the subsequent effects of a disturbed gut microbiota on distant organ injury and survival were limited.
As far as we know, significant changes in the intestinal microbiota within 72 hours of systemic bacterial infection have never been demonstrated before in any model of sepsis. Human studies describing microbiota perturbation during sepsis are confounded by the universal use of antibiotics [25, 26]. We here demonstrate that the systemic inflammatory response itself can lead to marked alterations in the gut microbiota. Our findings are in line with a recent report on intestinal dysbiosis, caused by influenza infection [27]. Another study associated pulmonary Mycobacterium tuberculosis infection in mice with loss of intestinal microbiota diversity after six days, with a subsequent recovery during the following weeks [28]. As virtually no M. tuberculosis was detected in feces, it was suggested that these changes in intestinal microbiota were due to alterations in the adaptive immune system, which in the tuberculosis model becomes effective at controlling the infection around the same time [28]. We therefore expected to find lower microbial diversity three days after infection with B. pseudomallei, but observed the opposite. A possible explanation could be the elimination of several “big players” of the gut microbiota by the host immune system during severe systemic bacterial infection, giving way to other bacteria to proliferate.
The amount of data that demonstrates a beneficial effect of the intestinal microbiota on the systemic innate immune system in infection is rapidly expanding. Previous studies described a protective effect of the intestinal microbiota during E. coli and L. monocytogenes sepsis via stimulation of granulopoiesis in the bone marrow [14, 15]. Crosstalk between microbiota and bone marrow has been suggested to happen via interleukin-17, -22 and granulocyte colony-stimulating factor (G-CSF) [14, 15]. In addition, neutrophil function is affected in both germ free and antibiotic pre-treated mice, resulting in decreased killing of S. pneumoniae and S. aureus [10]. In contrast, we did not find any indications for a central role for neutrophils in the antibacterial effect of the microbiota that we observed during melioidosis. The so-called microbiota-bone marrow axis could be more important in younger mice, as were used in above mentioned studies. Also, many studies use germ free mice, which could display more pronounced phenotypes than antibiotic treated mice. Our findings are in line with earlier reports that suggest a positive effect of healthy intestinal microbiota on alveolar macrophages [11, 12, 16, 29]. As alveolar macrophages constantly adapt to their environment, one can imagine them being affected by the level of circulating compounds derived from the intestinal microbiota (e.g. cell wall components or metabolites). Microbial disturbances may induce an altered phenotype of these cells, leading to decreased phagocytosis of B. pseudomallei, which in turn may lead to decreased intracellular killing. This is in line with our previous findings in a mouse model of pneumococcal pneumonia, in which we found that phagocytic capacities of alveolar macrophages are affected by antibiotic gut microbiota disruption [16]. We found changes in a cholesterol synthesis pathway in the transcriptome of these alveolar macrophages, which could be important as cholesterol-rich membrane rafts are involved in phagocytosis [30].
This study has a number of limitations. The antibiotics were chosen based on similar experiments in the literature [10, 12, 16]; however, other antibiotic regimens may have different effects. Also, mice from different suppliers could have a different intestinal microbiota and as a result elicit different immune responses, as was recently demonstrated in a mouse model for malaria [31]. In addition, we cannot exclude a direct effect of antibiotics on the host response; however, our data are in line with previous reports on the effect of the gut microbiota on the innate immune response during infection [10, 12, 16]. Alterations in the respiratory tract microbiota could be another contributing factor to the observed phenotype [32]. We did however not find any differences in pulmonary microbiota between control- and antibiotic treated mice, making it less likely that this is of influence. Lastly, the situation in actual melioidosis patients is very different from this experimental murine setting; comorbidities, medications and interindividual differences all may influence a possible interplay between microbiota and innate immune system.
In summary, we observed increased bacterial dissemination in mice with a disrupted gut microbiota during pneumonia-derived B. pseudomallei sepsis, indicating that the intestinal microbiota improves host defense against melioidosis. Alveolar macrophages from microbiota-disrupted mice showed a diminished capacity to phagocytose B. pseudomallei. It will be very interesting to study if disruption of the microbiota by antibiotics affects susceptibility to melioidosis. There is evidence for the incidence of severe sepsis being higher after events known to be associated with disturbance of the intestinal microbiota, such as hospitalization for Clostridium difficile infection [33]. Hopefully, further research into the interplay between intestinal microbiota and melioidosis will tell us whether and how this knowledge could be used to the advantage of patients.
A detailed description of methods is available in the online supplement.
Specific pathogen-free C57BL/6 mice were purchased from Charles River (Maastricht, The Netherlands). In selected experiments, antibiotic treatment was started at six weeks of age (see below); infection was induced in all experiments at nine weeks of age. The animals were housed in IVC cages in rooms with a controlled temperature and light cyclus. They were acclimatized for one week prior to usage, and received standard rodent chow and water ad libitum.
The Institutional Animal Care and Use Committee of the Academic Medical Center approved all experiments (permit number DIX21, sub-protocols 21BB, 21CJ and 21DJ) and ethical approval was obtained to use B. pseudomallei strain 1026b for animal experiments (08–150; see Supplemental Methods). B. pseudomallei strain 1026b strain was received by our lab in 2004 as a kind gift from the Donald E. Woods lab, University of Calgary, Alberta, Canada. Samples were anonymized if applicable. Experiments were carried out in accordance with the Dutch Experiments on Animals Act.
Experimental melioidosis was induced by intranasal inoculation with 150 or 500 colony forming units (CFU) of B. pseudomallei strain 1026b as described [17, 18]. At 24 or 72 hours post-infection, mice were euthanized and sacrificed by bleeding from the heart, after which organs were harvested. For survival studies, mice were observed for 14 days.
Fresh stool pellets were obtained and stored at -80°C. DNA isolation followed by IS-pro bacterial profiling was performed as described before (IS-diagnostics, Amsterdam, The Netherlands) [19, 20]. In short, the length of the 16S-23S rDNA interspace (IS) region is used to classify bacteria by PCR, combined with phylum-specific fluorescent labelling of PCR primers. IS fragment analysis was performed on an ABI Prism 3500 Genetic Analyzer (Applied Biosystems). Data were analysed with IS-pro proprietary software (IS-diagnostics, Amsterdam, The Netherlands).
Mice received broad-spectrum antibiotics (ampicillin 1 g/L; neomycin 1 g/L, both from Sigma, Zwijndrecht, The Netherlands; metronidazole 1 g/L, Sanofi-Aventis, Gouda, The Netherlands and vancomycin 0.5 g/L, Xellia pharmaceuticals, Copenhagen, Denmark) in drinking water for 19 days [10, 16]. This cocktail disrupts the intestinal microbiota and significantly lowers microbial diversity [16]. In selected experiments, only ampicillin and metronidazole were used. After a washout period of two days with normal drinking water, mice were inoculated with B. pseudomallei or sacrificed naïve. Weights of control- and antibiotic treated mice were equal at the moment of inoculation.
Alveolar macrophages, peritoneal macrophages, blood and bone marrow derived macrophages were obtained and incubated as described previously [16, 17, 34, 35]. Briefly, cells were seeded, washed and stimulated overnight with LPS or heat-killed B. pseudomallei. For in vivo phagocytosis, mice were inoculated intranasally with 5x106 CFU heat-killed, FITC (fluoresceine isothiocyanate)-labeled B. pseudomallei. After three hours, mice were anesthetized and broncho-alveolar lavage (BAL) was performed. FITC-positivity of alveolar macrophages was determined by FACS analysis. Details are provided in the online supplement.
RNA was isolated from lung homogenates using the RNeasy mini kit (Qiagen, Venlo, The Netherlands). Biotinylated cRNA was hybridized onto the Illumina MouseRef-8v2 Expression BeadChip and an Illumina iScan array scanner (Eindhoven, The Netherlands) was used to scan samples [16, 36]. Detailed methods are available in the supplemental material.
Differences between groups were analyzed by Mann-Whitney U test. Differences in microbiota diversity over time were analysed by paired t-test. For survival, Kaplan-Meier analysis followed by log-rank test was performed and the clinical scores by matched two-way ANOVA. Analyses were performed using GraphPad Prism 5. Values of P<0.05 were considered statistically significant.
|
10.1371/journal.ppat.1001311 | The Coxsackievirus B 3Cpro Protease Cleaves MAVS and TRIF to Attenuate Host Type I Interferon and Apoptotic Signaling | The host innate immune response to viral infections often involves the activation of parallel pattern recognition receptor (PRR) pathways that converge on the induction of type I interferons (IFNs). Several viruses have evolved sophisticated mechanisms to attenuate antiviral host signaling by directly interfering with the activation and/or downstream signaling events associated with PRR signal propagation. Here we show that the 3Cpro cysteine protease of coxsackievirus B3 (CVB3) cleaves the innate immune adaptor molecules mitochondrial antiviral signaling protein (MAVS) and Toll/IL-1 receptor domain-containing adaptor inducing interferon-beta (TRIF) as a mechanism to escape host immunity. We found that MAVS and TRIF were cleaved in CVB3-infected cells in culture. CVB3-induced cleavage of MAVS and TRIF required the cysteine protease activity of 3Cpro, occurred at specific sites and within specialized domains of each molecule, and inhibited both the type I IFN and apoptotic signaling downstream of these adaptors. 3Cpro-mediated MAVS cleavage occurred within its proline-rich region, led to its relocalization from the mitochondrial membrane, and ablated its downstream signaling. We further show that 3Cpro cleaves both the N- and C-terminal domains of TRIF and localizes with TRIF to signalosome complexes within the cytoplasm. Taken together, these data show that CVB3 has evolved a mechanism to suppress host antiviral signal propagation by directly cleaving two key adaptor molecules associated with innate immune recognition.
| Mammalian cells utilize a variety of defenses to protect themselves from microbial pathogens. These defenses are initiated by families of receptors termed pattern recognition receptors (PRRs) and converge on the induction of molecules that function to suppress microbial infections. PRRs respond to essential components of microorganisms that are broadly expressed within classes of pathogens. The relative non-specificity of this detection thus allows for a rapid antimicrobial response to a variety of microorganisms. Coxsackievirus B3 (CVB3), a member of the enterovirus genus, is associated with a number of diverse syndromes including meningitis, febrile illness, diabetes, and is commonly associated with virus-induced heart disease in adults and children. Despite its significant impact on human health, there are no therapeutic interventions to treat CVB3 infections. Here we show that CVB3 has evolved an effective mechanism to suppress PRR signal propagation by utilizing a virally-encoded protein, termed 3Cpro, to directly degrade molecules that function downstream of PRR signaling. By targeting these molecules, CVB3 can evade host detection and escape antiviral defenses normally induced by mammalian cells. These findings will lead to a better understanding of the mechanisms employed by CVB3 to suppress host antiviral signaling and could lead to the development of therapeutic interventions aimed at modulating CVB3 pathogenesis.
| The innate immune system is the first line of defense against pathogen infiltration and is activated by the binding of conserved microbial ligands to pattern recognition receptors (PRRs). Activation of these receptors culminates in nuclear factor (NF)-κB and/or IFN regulatory factor (IRF)-mediated induction of type 1 interferons (IFN-α and -β), key components of antimicrobial host defenses.
PRRs, including Toll-like receptors (TLRs) and DExD/H box RNA helicases, signal through an assortment of downstream adaptor molecules to propagate innate immune signaling. TLRs signal through adaptor molecules such as myeloid differentiation factor 88 (MyD88), Toll/IL-1 receptor domain containing adaptor protein (TIRAP), Toll/IL-1 receptor domain containing adaptor inducing interferon-beta (TRIF), and TRIF-related adaptor molecule (TRAM) to activate cellular defenses [1]. These adaptors often display specificity with regard to the TLR family members with whom they interact with and from which they are activated. The specificity of TLR ectodomain-ligand recognition and concomitant specificity in the signaling networks that are engaged by this interaction provides an efficient strategy for microbial recognition. In contrast, activated DExD/H box RNA helicases, which include melanoma differentiation associated gene (MDA5) and retinoic acid induced gene-I (RIG-I), signal to a common downstream adaptor molecule, mitochondrial antiviral signaling [(MAVS), also known as VISA/IPS-1/Cardif] to activate NFκB and IRF3 [2], [3], [4]. MAVS is localized to the mitochondrial membrane and to peroxisomes via a C-terminal transmembrane domain, which is essential for innate immune signaling [5], [6]. PRR-associated adaptor molecules thus serve critical roles in the activation of cellular defenses associated with microbial recognition.
As host cells have developed highly specialized strategies for microbial detection and clearance, it is not surprising that many viruses have evolved strategies to counter these defenses in order to promote their replication and spread. In some cases, virally-encoded proteases directly target components of the innate immune system to abolish antiviral signaling via TLRs and/or DExD/H box helicases. Targeted proteolysis of adaptor molecules serves as a powerful means to eliminate antiviral signaling by suppressing common downstream targets of key innate immune signaling pathways. For example, MAVS is cleaved by the NS3/4A serine protease of hepatitis C virus (HCV) [7], the 3Cpro cysteine protease of hepatitis A virus (HAV) [8], the HCV-related GB virus B NS3/4A protease [9], and the 2Apro and 3Cpro proteases of rhinovirus [10]. HCV also utilizes the same NS3/4A serine protease to cleave TRIF in order to silence TLR3-mediated signaling [11]. Thus, the targeting of MAVS and/or TRIF by virally-encoded proteases in order to suppress antiviral signaling is emerging as a common theme in the evasion of host defenses.
Enteroviruses, which belong to the Picornaviridae family, are small single-stranded RNA viruses that account for several million symptomatic infections in the United States each year. Coxsackievirus B3 (CVB3), a member of the Enterovirus genus, is associated with a number of diverse syndromes, including meningitis, febrile illness, and diabetes [12] and is an important causative agent of virus-induced heart disease in adults and children [13], [14], [15], [16]. The induction of type I IFN signaling is essential for the control of CVB3 infection, as evidenced by enhanced virus-induced lethality in type I IFN receptor (IFN-α∼β R) null mice [17] and increased susceptibility to CVB3 infection in IFNβ-deficient mice [18]. Both TLR3- and MDA5-mediated type I IFN signaling have been implicated in the response to CVB3 infections and mice deficient in either TRIF or MAVS show an enhanced susceptibility to viral infection [19], [20], [21].
In this study, we determined the potential mechanisms employed by CVB3 to antagonize type I IFN signaling. We found that infection of cells with CVB3 led to the cleavage of the adaptor molecules MAVS and TRIF. Both MAVS and TRIF were cleaved by the CVB3-encoded cysteine protease 3Cpro, indicating that a single protease suppresses innate immune signaling through two powerful pathways. We found that 3Cpro cleaves specific residues within MAVS and TRIF that render these molecules deficient in type I IFN signaling and apoptotic signaling. Taken together, these data suggest that CVB3 has evolved a mechanism to cleave adaptor components of the innate immune system to escape host immunity.
The induction of type I IFNs is the earliest cellular immune response initiated to combat viral infections and is coordinated by the activation of transcription factors such as interferon regulatory factor (IRF)-3, IRF7, and NFκB downstream of PRR activation. We found that CVB3 infection of HEK293 cells led to only a modest induction of IRF3 activation as assessed by immunofluorescence microscopy for nuclear translocation (Figure 1A), western blot analysis of nuclear extracts (Figure 1B), and luciferase activity assays for IFNβ (Figure 1C). In contrast, transfection of cells with poly I:C induced pronounced IRF3 activation (Figure 1A–C). We also observed little activation of NFκB signaling in response to CVB3 infection as determined by luciferase activation assay (Figure 1C).
Because CVB3 did not elicit a pronounced translocation of IRF3 into the nucleus during infection of HEK293 cells, we investigated the role of several PRRs in mediating CVB3 recognition–TLR3, RIG-I, and MDA5. Both MDA5 [22] and TLR3 [19] have been proposed to act as sensors for CVB3 infection. Although infection of cells with CVB3 is sensitive to IFNβ (Supplemental Figure S1A), we observed less enhancement of IFNβ promoter activity as assessed by luciferase activation in CVB3-infected HEK293 cells overexpressing MAVS, MDA5, RIG-I, and TLR3/TRIF than in uninfected controls (Figure 1D). Instead, we observed the partial ablation of IFNβ promoter activity in response to ectopic expression of MAVS, RIG-I, MDA5, and TLR3/TRIF in CVB3 infected cells (Figure 1D). We also found that CVB did not induce potent IFNβ production in HEK293, HeLa, or Caco-2 cells in comparison to VSV controls (Figure 1E).
Because CVB3 infection was inefficient at inducing IRF3, we assessed the pattern of expression of MAVS in CVB3-infected HEK293 cells. By immunoblot analysis, we found that CVB3 infection induced the cleavage of MAVS (Figure 2A). Similar results were obtained in HeLa cells (Supplemental Figure S2A). This effect was specific for CVB3 as infection with VSV did not alter MAVS migration (Supplemental Figure S2B). In uninfected cells, full-length MAVS migrated as a single band of ∼75 kD. However, in cells infected with CVB3, there was a decrease in the expression level of full-length MAVS and the appearance of a distinct MAVS cleavage fragment migrating at ∼40–50 kD (Figure 2A). Because MAVS cleavage is induced in cells undergoing apoptosis [23], [24] and CVB3 is known to induce apoptosis in many cell types [25], [26], we investigated the role of apoptosis in CVB3-induced MAVS cleavage. We found that incubation of CVB3-infected HEK293 cells with the broad caspase inhibitor z-VAD-FMK and the proteosome inhibitor MG132 had little effect on CVB3-induced MAVS cleavage (Figure 2A). (The slight reduction in MAVS cleavage observed in the presence of MG132 is likely attributable to a reduction in replication in MG132-exposed cells, consistent with previously published results [27], [28]). The kinetics of MAVS cleavage was also not consistent with apoptosis: MAVS cleavage was evident by 3 hrs post-infection (p.i.) whereas apoptosis (as measured by caspase-3 cleavage) did not occur until 5–6 hrs p.i. (Figure 2B).
MAVS is localized to the mitochondrial membrane via a C-terminal transmembrane domain [5]. We found that CVB3 infection induced a pronounced decrease in MAVS mitochondrial localization as assessed by immunofluorescence microscopy with a mitochondrial marker (Figure 2C). We also found that the expression and mitochondrial localization of ectopically expressed MAVS was significantly reduced in CVB3-infected cells (Figure 2D, 2E). The appearance of cleavage fragments was evident in CVB3-infected cells overexpressing MAVS (Figure 2D). Moreover, we found that mutation of the caspase cleavage site of MAVS (D429E, [24]) had no effect on CVB3-induced MAVS cleavage (Supplemental Figure S2C), indicating a caspase-independent mechanism of action.
Another common pathway upstream of IRF3 activation is the engagement of TLR3 by viral dsRNA, which is produced as a replication intermediate during viral infection. As we observed cleavage of MAVS in CVB3-infected cells, we sought to determine if CVB3 might also target TRIF, the specific adaptor molecule downstream of TLR3, to repress IRF3 activation. Similar to our findings with MAVS, we found that TRIF expression was significantly reduced in HeLa cells infected with CVB3 (Figure 2F). Although TRIF can be cleaved during apoptosis [23], [24], we found that z-VAD-FMK and MG132 had little effect at antagonizing the CVB3-mediated reduction in TRIF expression (Figure 2F), consistent with our findings with MAVS (Figure 2A). The kinetics of TRIF cleavage also paralleled that of MAVS as we observed a marked reduction in TRIF levels by 3 hrs p.i. (Figure 2G), a time prior to the induction of caspase-3 cleavage (Figure 2B). Ectopically expressed CFP-fused TRIF was also significantly decreased in cells infected with CVB3 and coincided with the appearance of several cleavage fragments (Figure 2H).
We next investigated whether cleavage of MAVS and TRIF occurred in cells infected with other enteroviruses including echovirus 7 (E7) and enterovirus 71 (EV71). Infection of HeLa cells with both E7 and EV71 led to the significant reduction of MAVS and TRIF expression, which corresponded with the appearance of the newly replicated viral protein VP1 (Supplemental Figure S3A). However, in contrast to our findings with CVB3 (Figure 2A), we did not observe the appearance of any significant cleavage fragments in either E7 or EV71-infected cells. This may indicate that the cleavage fragments are short-lived in E7 or EV71-infected cells or that cleavage occurs at different residues within the molecule that alter antibody binding. These results may indicate that members of the enterovirus family target MAVS and TRIF to evade host immunity, but further studies are required to definitively show which members of the enterovirus family utilize this mechanism.
CVB3 infections are commonly associated with virus-induced heart disease in adults and children and have been detected in approximately 20-25% of patients with dilated cardiomyopathy and myocarditis [13], [14], [15], [16]. To determine whether MAVS and TRIF are degraded in vivo, mice were infected with CVB3 and the hearts of infected animals were probed for MAVS and TRIF. In contrast to uninfected controls, there was an almost complete absence of both MAVS and TRIF in murine hearts infected with CVB3 (Supplemental Figure S3B). These data indicate that the cleavage of MAVS and TRIF may also occur during CVB3 infection in vivo.
Enteroviruses encode specific proteases that are required for the processing of viral proteins and the establishment of replication, but which also cleave a variety of host cell molecules [29]. Because we observed the cleavage of MAVS in CVB3-infected cells, we investigated whether virally-encoded proteases might mediate this effect. We cotransfected HEK293 cells with N-terminal Flag-MAVS and various CVB3 viral proteins fused to EGFP. Of these proteins, we found that expression of the protease 3Cpro was sufficient to induce a significant reduction in MAVS expression (Figure 3A). In fact, in order to observe significant levels of full-length Flag-MAVS (or cleavage fragments) in EGFP-3Cpro co-transfected cells, cells had to be transfected with twice as much Flag-MAVS as vector control or other CVB3 viral proteins. The apparent lack of cleavage products in cells overexpressing proteases is a phenomenon that has also been observed for HCV-mediated cleavage of TRIF [11] and likely reflects the high efficiency of cleavage (which may result from protease overexpression) and that cleavage fragments are unstable and/or short-lived. For our subsequent studies, we transfected cells with equivalent amounts of MAVS cDNA to compare the level of full-length MAVS in control versus 3Cpro-transfected cells. The cleavage of MAVS required the cysteine protease activity of 3Cpro, as cotransfection of a catalytically inactive N-terminal EGFP-tagged 3Cpro mutant (C147A) [30] had no effect on MAVS expression (Figure 3B). In some cases, significant levels of GFP signal alone can be detected in EGFP-3Cpro WT transfected cells which is likely indicative of 3Cpro cleaving itself from the N-terminal EGFP tag.
To confirm that 3Cpro was directly cleaving MAVS, we incubated recombinant wild-type or C147A mutant 3Cpro with Flag-MAVS purified by Flag column affinity purification from overexpressing HEK293 cells. Whereas incubation with wild-type 3Cpro induced the appearance of a MAVS cleavage fragment as determined by Flag immunoblotting, the C147A mutant did not induce the appearance of a MAVS cleavage product (Figure 3C). Moreover, whereas expression of wild-type EGFP-3Cpro induced the relocalization of MAVS as assessed by immunofluorescence microscopy, expression of EGFP-3Cpro C147A had no effect (Figure 3D).
Because we also observed cleavage of TRIF in CVB3-infected cells, we determined whether 3Cpro was responsible for its cleavage as well. We cotransfected HEK293 cells with TRIF and either EGFP-2Apro, 3A, or -3Cpro. Expression of 3Cpro, but not 2Apro or 3A, led to the cleavage of TRIF, demonstrated by a reduction in the expression of full-length TRIF and the appearance of several TRIF cleavage fragments (Figure 3E). 3Cpro-mediated cleavage of TRIF required the cysteine protease activity of 3Cpro as expression of 3Cpro C147A did not lead to TRIF cleavage (Figure 3F). We also confirmed that 3Cpro was directly cleaving TRIF by incubation of Flag-TRIF purified by Flag column affinity purification from overexpressing HEK293 cells with recombinant wild-type or C147A mutant 3Cpro. Similar to our findings with MAVS (Figure 3C), we found that only recombinant wild-type 3Cpro induced the appearance of TRIF cleavage fragments (Figure 3G). Note that the pattern of TRIF cleavage by in vitro proteolysis assay (Figure 3G) differs from our experiments with overexpressed 3Cpro in HEK293 cells (Figure 3E, 3F) due to the use of C-terminal CFP- versus N-terminal Flag-tagged TRIF between experiments. Taken together, our data show that 3Cpro directly cleaves both MAVS and TRIF.
To assess whether expression of 3Cpro abrogated MAVS-dependent signaling, we transfected HEK293 cells with wild-type or C147A EGFP-3Cpro or vector control, with a luciferase reporter fused to the IFNβ promoter region (p-125-Luc), and with either Flag-MAVS or the caspase activation and recruitment domains (CARDs) of MDA5 or RIG-I. Expression of the CARDs of MDA5 and RIG-I alone results in the constitutive activation of type I IFN signaling even in the absence of stimulus [31]. We found that whereas there was pronounced induction of IFNβ activity in cells expressing vector alone or EGFP-3Cpro C147A, expression of wild-type EGFP-3Cpro led to a significant reduction in promoter activity (Figure 4A).
We next determined whether 3Cpro attenuated TRIF-mediated signaling. TRIF is involved in the activation of IRF3 and IFNβ induction downstream of dsRNA-TLR3 engagement. While the expression of TRIF and vector control enhanced IFNβ promoter activity, expression of TRIF in combination with 3Cpro significantly impaired IFNβ promoter activity (Figure 4B). We found that 3Cpro-mediated inhibition of IFNβ signaling was indeed occurring upstream of IRF3 activation as coexpression of wild-type 3Cpro and IRF3 had no effect on IRF3-mediated activation of IFNβ promoter activity (Supplemental Figure S4A). Furthermore, we found that expression of wild-type, but not C147A 3Cpro reduced IFNβ activation in response to infection with VSV (Supplemental Figure S4B).
In addition to their roles in type I IFN signaling, ectopic expression of MAVS [32] and TRIF [33] potently stimulate intrinsic apoptotic machinery to induce cell death. We found that expression of MAVS or TRIF induced pronounced apoptosis as demonstrated by enhanced Annexin V binding [which identifies the externalization of phosphatidylserine in cells undergoing apoptosis] (Figure 4C, 4D). In contrast, expression of MAVS or TRIF in the presence of 3Cpro potently reduced apoptosis (Figure 4C, 4D). Taken together, these data show that 3Cpro represses both the apoptotic and type I IFN signaling mediated by MAVS and TRIF.
3Cpro preferentially cleaves glutamine-glycine (Q-G) bonds in both the viral polyprotein and cellular targets, but may also exhibit proteolytic activity against glutamine-alanine (Q-A) bonds, amongst others [29]. In order to identify the residue(s) within MAVS cleaved by 3Cpro, we constructed a panel of site directed mutants within MAVS at residues that may serve as 3Cpro cleavage sites (Q148, Q211, and E480) (Figure 5A). Of these mutants, only one (Q148A) was resistant to 3Cpro-mediated cleavage in HEK293 cells and by in vitro protease assay (Figure 5B). Moreover, whereas wild-type Flag-MAVS was relocalized from the mitochondrial membrane upon expression of EGFP-3Cpro, the Q148A Flag-MAVS mutant retained its mitochondrial localization (Figure 5C). The 3Cpro cleavage site within MAVS (Q148) is located in the proline rich region, which mediates its interaction with a number of signaling molecules including TRAF2 [4], TRAF3 [34], TRAF6 [4], RIP1 [2], and FADD [2].
We next determined whether the Q148A mutant of MAVS was resistant to 3Cpro-mediated abatement of MAVS signaling. We found that whereas there was a pronounced reduction in IFNβ activity in cells expressing wild-type Flag-MAVS and EGFP-3Cpro, there was no effect of EGFP-3Cpro expression on IFNβ signaling in cells transfected with Q148A Flag-MAVS (Figure 5D, Supplemental Figure S2D). These data show that CVB3 3Cpro cleaves MAVS at Q148 to suppress MAVS signaling.
MAVS requires an intact CARD and localization to the mitochondrial membrane (via a C-terminal transmembrane domain) to remain functionally active [5]. Because we found that 3Cpro cleaved MAVS at a specific residue (Q148) within the proline rich region, we next determined whether either of the possible 3Cpro-induced cleavage fragments of MAVS would remain active. To that end, we constructed EGFP-fused constructs expressing wild-type MAVS, the N-terminal (residues 1-148), or C-terminal (residues 149-540) fragments of MAVS that would result from 3Cpro cleavage (Figure 5F). We found that the N-terminal fragment of MAVS (1-148) no longer localized to the mitochondrial membrane (Figure 5G) and induced NFκB or IFNβ signaling significantly less that full-length MAVS (Figure 5H). Whereas the C-terminal fragment of MAVS (149-540) retained its mitochondrial localization (Figure 5G), it also exhibited significantly less NFκB and IFNβ activation in comparison to full-length MAVS (Figure 5H). These data indicate that 3Cpro-mediated cleavage of MAVS likely inactivates MAVS-mediated downstream signaling by directly cleaving a residue that separates the CARD and transmembrane regions.
Overexpressed TRIF forms multimers and localizes to punctate cytoplasmic structures referred to as the TRIF ‘signalosome’ [35]. Downstream components of TRIF signaling localize to signalosomes as a mechanism to stimulate TRIF signaling [35], [36]. We found that EGFP-3Cpro and EGFP-3Cpro C147A were recruited to TRIF signalosomes when co-expressed with TRIF (Figure 6A). This recruitment was specific for 3Cpro as we did not observe the recruitment of either EGFP-2Apro (not shown) or EGFP-3A (Supplemental Figure S5) to TRIF signalosomes. Although TLR3 (and presumably TRIF) can localize to endosomal membranes [37], we did not observe any colocalization of overexpressed TRIF with markers of both early and late endosomes (Supplemental Figure S6).
Because we observed the relocalization of 3Cpro to the signalosome complex, we next determined whether 3Cpro and TRIF interact within this specialized complex. HEK293 cells were transfected with TRIF and either vector (EGFP alone), EGFP-3Cpro wild-type, or EGFP-3Cpro C147A and co-immunoprecipitation studies were performed. We found that whereas EGFP-3Cpro C147A and TRIF co-immunoprecipitated, wild-type EGFP-3Cpro and TRIF did not (Figure 6B). These findings indicate that 3Cpro forms an interaction with TRIF that is likely abolished upon 3Cpro-mediated cleavage.
TRIF contains a proline-rich N-terminal region, a Toll/Interleukin-1 receptor (TIR) domain, and a C-terminal region. To determine which TRIF domain is responsible for interacting with 3Cpro and recruiting it to the signalosome, we constructed N-terminal (NT, 1–359aa), C-terminal (CT, 360–712aa), and TIR (390–460aa) domain expression constructs of TRIF containing a HA-tag at the N-terminus and a Flag-tag at the C-terminus (Figure 6C). We then coexpressed these constructs with wild-type and C147A versions of 3Cpro and performed fluorescence microscopy and immunoprecipitation analysis. We found that 3Cpro C147A specifically interacted with the C-terminal domain of TRIF, but not the N-terminus (Figure 6D). However, the TIR domain did not mediate the interaction between TRIF and 3Cpro as we observed no co-immunoprecipitation between HA-TIR-Flag and 3Cpro (not shown).
Previous studies have shown that expression of the C-terminus of TRIF is required for the formation of the TRIF signalosome [36]. We found that expression of HA-CT-Flag was sufficient to induce the relocalization of 3Cpro C147A to signalosomes (Figure 6E). In contrast, 3Cpro C147A did not localize with either HA-NT-Flag or HA-TIR-Flag (Figure 6E). The formation of tubule-like structures induced by the expression of the TRIF TIR is consistent with previous work by others [36]. Taken together, these data indicate that 3Cpro interacts with the C-terminus of TRIF that is sufficient for its recruitment into the TRIF signalosome.
We did not observe any interaction between wild-type 3Cpro and either full-length or C-terminal TRIF (Figure 6D and 6E) suggesting that the interaction between TRIF and 3Cpro is diminished following cleavage. Interestingly, we observed the appearance of cleavage fragments of both HA-NT-Flag and HA-CT-Flag when coexpressed with wild-type 3Cpro (Figure 6D). To further define the extent of 3Cpro-mediated proteolysis of the N- and C-terminal regions of TRIF, we coexpressed dually HA- and Flag-tagged constructs of TRIF (described in Figure 6C) and wild-type or C147A EGFP-3Cpro and subjected lysates to dual-color (700 nm and 800 nm) immunoblot analysis using a LI-COR Odyssey infrared imaging system and antibodies specific for HA and Flag. This technique could therefore allow for the detection of a variety of TRIF cleavage fragments simultaneously. We found that expression of wild-type 3Cpro (but not the C147A mutant) induced the cleavage of both the N- and C-termini of TRIF (Figure 7A). In contrast, we observed no cleavage of the TIR domain (Figure 7A). Additionally, our data indicate that the C-terminus of TRIF is cleaved more abundantly than the N-terminus as we observed a marked decrease in the expression of full-length HA-CT-Flag and the appearance of several HA- or Flag-tag-positive cleavage products induced by 3Cpro overexpression (Figure 7A).
The N- and C-terminal regions of TRIF differ in their capacities to induce type I IFN and NFκB signaling—whereas overexpression of the N-terminal region of TRIF activates both IFNβ and NFκB signaling, the C-terminal domain fails to activate IFNβ but potently induces NFκB activation [38], [39]. Moreover, the C-terminus of TRIF is sufficient to induce apoptosis [33]. The RIP homotypic interaction motif (RHIM) at the C-terminus of TRIF is essential for both NFκB and apoptotic signaling [33], [40]. Because we observed pronounced 3Cpro-mediated cleavage of the C-terminus of TRIF (Figure 7A), we investigated whether NFκB and apoptotic signaling mediated by the C-terminus of TRIF was abolished. We found that expression of wild-type 3Cpro potently abrogated NFκB and apoptotic signaling induced by expression of the C-terminus of TRIF (Figure 7B, 7C). These findings are consistent with those indicating that 3Cpro also inhibits full-length TRIF-mediated apoptotic signaling (Figure 4C, 4D).
In order to identify the residue(s) within TRIF cleaved by 3Cpro, we constructed a panel of site directed mutants within the TRIF N- and C-terminal domains at residues that may serve as 3Cpro cleavage sites (Figure 7D). [We omitted any potential sites within the TRIF TIR domain as we did not observe 3Cpro-induced cleavage of this domain (Figure 7A)]. We found that a specific residue (Q190) within the N-terminal region of TRIF was targeted by 3Cpro as mutagenesis of this site abolished 3Cpro-induced cleavage (Figure 7E). Because several sites in the C-terminal domain of TRIF can serve as possible 3Cpro cleavage sites, and because these sites lie within close proximity to one another, we mutated these sites simultaneously. We found that simultaneous mutagenesis of four potential 3Cpro cleavage sites (Q653, Q659, Q671, and Q702) was sufficient to prevent 3Cpro cleavage (Figure 7F). These findings are consistent with our observation that the C-terminus of TRIF likely undergoes 3Cpro cleavage at several sites (Figure 7A).
Because the N- and C-terminal domains of TRIF function in unique capacities to induce IRF3, NFκB, and apoptotic signaling, we next explored whether possible 3Cpro cleavage fragments of TRIF could remain functional in these pathways. We constructed EGFP-fused full-length TRIF and various possible cleavage fragments of TRIF (encoding residues 190-653, 190-671, or 190-702). We found that all three possible 3Cpro TRIF cleavage fragments maintained their capacity to activate type I IFN signaling (as assessed by luciferase assays for IFN- stimulated response element (ISRE), an IRF3-dependent promoter) (Figure 7G, 7H). In contrast, two of these fragments, 190–653 and 190-671, lost their ability to activate NFκB or induce apoptotic signaling (Figure 7G, 7H). These data indicate that 3Cpro-mediated cleavage of TRIF may primarily function to suppress TRIF-mediated NFκB and apoptotic signal propagation.
The host innate immune response to viral infections often involves the activation of parallel PRR pathways that converge on the induction of type I IFNs and NFκB activation. Several viruses have evolved sophisticated mechanisms to evade the host innate immune response by directly interfering with the activation and/or downstream signaling events associated with PRR signal propagation. Here we show that the 3Cpro cysteine protease of CVB3 targets MAVS and TRIF, two key adaptor molecules in the innate immune response as a mechanism to suppress type I IFN and apoptotic signaling. By targeting these adaptors, CVB3 has evolved a strategy to suppress antiviral signal propagation through two powerful pathways—TLR3 and RIG-I/MDA5. 3Cpro cleaves MAVS at a specific site within its proline-rich region (at Q148) and suppresses MAVS-mediated induction of type I IFNs and apoptosis. Moreover, 3Cpro targets both the N- and C-terminal domains of TRIF to abrogate its type I IFN, NFκB, and apoptotic signaling capacities. Interestingly, we found that 3Cpro localized to TRIF signalosomes and interacted with the C-terminal domain of TRIF. Taken together, these data highlight the strategies used by CVB3 to evade the host innate immune response.
Many viruses target molecules upstream of IFN induction as a means to escape host immunity. Similar to our findings with CVB3 3Cpro, the 3Cpro of HAV directly cleaves MAVS to escape host immunity [8], but it is not known if HAV 3Cpro also cleaves TRIF. However, although HAV 3Cpro is responsible for mediating MAVS cleavage, the protease must be localized to the mitochondrial membrane via a transmembrane domain within the 3A viral protein in order to facilitate this event [8]. In contrast, CVB3 3A localizes to the ER membrane where it disrupts ER-Golgi vesicular trafficking [41], [42] and is thus not targeted to the mitochondrial membrane. Our studies indicate that in contrast to HAV, CVB3 3Cpro alone is sufficient to induce MAVS cleavage despite it not being localized to the mitochondrial membrane.
Although MAVS and TRIF are potent inducers of type I IFN signaling downstream of PRR activation, they have also been shown to induce apoptotic signaling–another powerful pathway used by host cells to suppress viral replication and progeny release. Enteroviruses are lytic viruses, and as such, possess no known mechanism for progeny release other than the destruction of the host cell membrane. Lytic viruses often develop efficient strategies to tightly regulate host cell death pathways in order to avoid killing the host cell prematurely (and terminating viral replication). CVB3 possesses anti-apoptotic strategies, which are mediated by the 2B and 2BC viral proteins [43], [44]. In addition, it has been shown that 3Cpro targets the inhibitor of κBα as a means to stimulate apoptosis and suppress viral replication [45]. Our results show that 3Cpro may also serve in an anti-apoptotic capacity to suppress MAVS- and TRIF-mediated apoptotic signaling as a means to tightly regulate host cell apoptotic pathways. The pro-apoptotic signaling mediated by MAVS requires its localization to the mitochondrial membrane and the presence of intact CARDs, but not the presence of an intact proline-rich region [32]. Although 3Cpro cleaves MAVS within the proline-rich region (Q148, Figure 5B, 5C), this cleavage both induces the relocalization of MAVS from the mitochondrial membrane (Figure 5C) and inhibits MAVS signals (Figure 5D). Furthermore, 3Cpro cleavage fragments of MAVS are non-functional (Figure 5H). Thus, the loss of MAVS-induced apoptosis in CVB3 3Cpro-expressing cells is likely the result of both the relocalization of MAVS from the mitochondrial membrane and the inhibition of signaling via the CARD regions. Moreover, CVB3 3Cpro targets the C-terminal region of TRIF, which has been shown to induce apoptosis via direct binding to receptor interacting protein 1 (RIP1) via its RIP homotypic interaction motif (RHIM) [33]. Specifically, we found that 3Cpro targeted several sites within the C-terminal domain of TRIF that could effectively remove the RHIM domain, a domain of TRIF known to be critically involved in NFκB and apoptotic signaling (Figure 7D, F). In support of this, we found that 3Cpro cleavage fragments were deficient in NFκB activation and apoptosis (Figure 7G). Taken together, these data indicate that 3Cpro suppresses MAVS and TRIF-induced apoptotic signals both by their direct cleavage and by their relocalization from either the mitochondria or signalosome, respectively.
The N- and C-terminal domains of TRIF serve disparate functions in the initiation of innate immune signaling. Whereas the N-terminus of TRIF activates type I IFN induction via the phosphorylation of IRF3, the C-terminal domain activates NFκB [38], [39]. Interestingly, we found that 3Cpro cleaves both of these domains—likely as a mechanism to suppress global TRIF-generated signaling capacities. Upon ligand stimulation of TLR3 (or upon overexpression), activated TRIF forms signalosomes enriched in TRIF-associated signaling components including RIP1 and NFκB -activating kinase-associated protein 1 (NAP1) [35], [36]. We found that 3Cpro localizes to the TRIF signalosome and that expression of the C-terminal domain of TRIF is sufficient to induce this localization (Figure 6A, 6E). Moreover, we found that 3Cpro interacts with the C-terminal domain of TRIF (Figure 6D). However, it remains unclear whether this interaction is direct or mediated via an adaptor molecule that also localizes to the signalosome. Additionally, we found that 3Cpro cleavage of the TRIF C-terminal domain leads to the disruption of TRIF signalosome formation (Supplemental Figure S7), which is required for the initiation of TRIF-mediated IRF3 and NFκB activation [36]. It is thus conceivable that 3Cpro attenuates TRIF-dependent signaling via direct cleavage, the degradation of the signalosome complex, and inhibition of the interactions between TRIF and downstream molecules that are required to propagate TRIF-dependent signals.
Although we found that 3Cpro cleavage fragments of TRIF were deficient in NFκB and apoptotic signaling, they retained their capacity to induce type I IFN signaling (Figure 7G). These data may indicate that the cleavage fragments of TRIF generated by 3Cpro cleavage are short-lived and do not accumulate within the cell. In support of this, we failed to identify TRIF cleavage products induced by CVB3 infection endogenously (Figure 2F, 2G). Alternatively, it remains possible that 3Cpro-mediated disruption of TRIF signaling is not involved in the suppression of type I IFN signaling, but may instead target type II IFN signaling. Previous studies in TLR3 and TRIF deficient mouse models indicate that TLR3- and TRIF-mediated IFNγ production plays an important role in CVB3 infections in vivo [19]. Thus, TLR3 signaling via TRIF to induce type II IFNs may function as a parallel pathway to MDA5 and/or RIG-I-mediated induction of type I IFNs. In this scenario, 3Cpro would suppress the downstream propagation of both type I and II IFN signaling in order to evade host immunity.
Viruses often utilize elegant strategies to attenuate innate immune signaling in order to promote their propagation. Here we show that the 3Cpro cysteine protease of CVB3 (and likely other enteroviruses) attenuates innate immune signaling mediated by two potent antiviral adapter molecules, MAVS and TRIF. By utilizing a variety of methods to abate MAVS and TRIF signaling, including both cleavage and retargeting from sites of signal propagation, 3Cpro can efficiently suppress both type I IFN and apoptotic signals aimed at clearing CVB3 infections. A better understanding of the mechanisms employed by enteroviruses to suppress host antiviral signaling could lead to the development of therapeutic interventions aimed at modulating viral pathogenesis.
Human embryonic kidney (HEK) 293, HeLa, and U2OS cells were cultured in DMEM-H supplemented with 10% FBS and penicillin/streptomycin. Human intestinal Caco-2 cells were cultured in MEM supplemented with 10% FBS and penicillin/streptomycin. Cells were screened for mycoplasma using a PCR-based mycoplasma test (Takara Bio USA) to prevent abnormalities in cellular morphology, transfection, and growth.
All experiments were performed with CVB3-RD, expanded as described [46]. Vesicular stomatitis virus (VSV) was kindly provided by Sara Cherry (University of Pennsylvania, Philadelphia, PA). Experiments measuring productive virus infection were performed with 0.1-1 plaque forming units (PFU)/cell for the indicated times. HeLa cells were infected with echovirus 7 and enterovirus 71 at a MOI = 0.1 for the indicated times.
Mouse infections were performed as described previously [47] and lysates kindly provided to us by Jeffrey M. Bergelson, Children's Hospital of Philadelphia.
Plasmid transfections were performed using FuGENE 6 according to the manufacturer's protocol (Roche Applied Science). Following transfection, cells were plated as described above and used 48–72 hrs later.
Cells cultured in collagen-coated chamber slides (LabTek, Nunc) were washed and fixed with either 4% paraformaldehyde or with ice-cold methanol. Cells were then permeabilized with 0.1% Triton X-100 in phosphate buffered saline (PBS) and incubated with the indicated primary antibodies for 1 hr at room temperature (RT). Following washing, cells were incubated with secondary antibodies for 30 min at room temperature, washed, and mounted with Vectashield (Vector Laboratories) containing 4′,6-diamidino-2-phenylindole (DAPI). For detection of apoptosis, cells were washed in cold PBS and incubated with Alexa-Fluor-488 conjugated-annexin V and propidium iodide for 15 min at room temperature. Cells were then washed, fixed in 4% paraformaldehyde, and images captured as described below. Images were captured using an Olympus IX81 inverted microscope equipped with a motorized Z-axis drive. Images were generated by multiple-section stacking (0.2 mm stacks) and deconvolved using a calculated point-spread function (Slidebook 5.0). Confocal microscopy was performed with a FV1000 confocal laser scanning microscope (Olympus).
Rabbit polyclonal and mouse monoclonal antibodies directed against GFP (FL, B-2), GAPDH, HA (Y-11, F-7) and IRF3 (FL-425) were purchased from Santa Cruz Biotechnology. Mouse monoclonal anti-Flag (M2) was purchased from Sigma. Rabbit polyclonal antibodies to TRIF and MAVS (human and rodent specific) were purchased from Cell Signaling Technologies or Bethyl Laboratories, respectively. Mouse anti-enterovirus VP1 (Ncl-Entero) was obtained from Novocastra Laboratories (Newcastle upon Tyne, United Kingdom). Mitochondria antibody [MTC02] was purchased from Abcam. Mouse anti-enterovirus 71 antibody was purchased from Millipore. Alexa Fluor-conjugated secondary antibodies were purchased from Invitrogen.
Flag-MDA5, Flag-MAVS, and Flag-TRIF plasmids were kindly provided by Tianyi Wang (University of Pittsburgh). pUNO2-hTRIF was purchased from Invivogen. EGFP-2Apro 2B, 2C, 3A and -3Cpro were constructed by amplification from CVB33 cDNA (kindly provided by Jeffrey Bergelson, Children's Hospital of Philadelphia) and cloned into the NT-GFP TOPO fusion vector (Invitrogen) following PCR amplification. EGFP-fusion constructs expressing cleavage fragments of MAVS and TRIF were generated by PCR amplification from Flag-MAVS or pUNO2-TRIF and cloned into the NT-GFP TOPO fusion vector (Invitrogen). CFP-TRIF was purchased from Addgene (plasmid 13644). Dual HA- and Flag-tagged TRIF constructs were generated by amplification of TRIF cDNA with primers encoding a N-terminal HA or C-terminal Flag tags and cloned into the XhoI and EcoRI sites of pcDNA3.1(+). Mutagenesis of 3Cpro, MAVS and TRIF constructs were performed using Quickchange mutagenesis kit following the manufacturer's protocol (Stratagene). Primer sequences are available upon request.
Cell lysates were prepared with RIPA buffer (50 mM Tris-HCl [pH 7.4]; 1% NP-40; 0.25% sodium deoxycholate; 150 mM NaCl; 1 mM EDTA; 1 mM phenylmethanesulfonyl fluoride; 1 mg/ml aprotinin, leupeptin, and pepstatin; 1 mM sodium orthovanadate), and insoluble material was cleared by centrifugation at 700×g for 5 min at 4°C. Lysates (30-50 µg) were loaded onto 4–20% Tris-HCl gels (Bio-Rad, Hercules, CA) and transferred to polyvinylidene difluoride membranes. Membranes were blocked in 5% nonfat dry milk or 3% bovine serum albumin, probed with the indicated antibodies, and developed with horseradish peroxidase-conjugated secondary antibodies (Santa Cruz Biotechology), and SuperSignal West Pico or West Dura chemiluminescent substrates (Pierce Biotechnology).
Immunoblots in isolated mouse hearts and dually HA- and Flag-tagged TRIF constructs were conducted using an Odyssey Infrared Imaging System (LI-COR Biosciences). Tissue homogenized in lysis buffer (100 µg) or whole-cell lysates from transfected HEK293 cells (30 µg) were loaded onto 4–20% Tris-HCl gels, separated electrophoretically, and transferred to nitrocellulose membranes. Membranes were blocked in Odyssey Blocking buffer and then incubated with the appropriate antibodies overnight at 4°C in Odyssey Blocking buffer. Following washing, membranes were incubated with anti-rabbit or anti-mouse antibodies conjugated to IRDye 680 or 800CW and visualized with the Odyssey Infrared Imaging System according to the manufacturer's instructions.
For immunoprecipitations, HEK293 cells transiently transfected with the indicated plasmids were lysed with EBC buffer (50 mM Tris [pH 8.0], 120 mM NaCl, 0.5% Nonidet P-40, 1 mm phenylmethylsulfonyl fluoride, 0.5 µg/ml leupeptin, and 0.5 µg/ml pepstatin). Insoluble material was cleared by centrifugation. Lysates were incubated with the indicated antibodies in EBC buffer for 1 hr at 4°C followed by the addition of Sepharose G beads for an additional 1 hr at 4°C. After centrifugation, the beads were washed in NETN buffer (150 mm NaCl, 1 mm EDTA, 50 mm Tris-HCl (pH 7.8), 1% Nonidet P-40, 1 mm phenylmethylsulfonyl fluoride, 0.5 µg/ml leupeptin, and 0.5 µg/ml pepstatin), then heated at 95°C for 10 min in Laemmli sample buffer. Following a brief centrifugation, the supernatant was immunblotted with the indicated antibodies as described above.
Bacterial expression vectors encoding wild-type or C147A 3Cpro were constructed in pET-SUMO (which encodes a linear fusion consisting of an N-terminal 6xHis tag for affinity purification followed by SUMO) following PCR amplification according to the manufacturers protocol (Invitrogen). pET-SUMO-3Cpro constructs were expressed in bacteria and purified by metal chelation resin columns (Qiagen). The resulting 6xHis-SUMO 3Cpro fusion proteins were treated with SUMO protease to create untagged proteins as per the manufacturer's instructions (Invitrogen) and then dialyzed.
For purification of MAVS and TRIF, 10 cm dishes of HEK2393 cells were transfected with Flag-MAVS or Flag-TRIF and lysed 48 hrs post-transfection. Lysates were purified over anti-Flag affinity gel columns, washed several times, and protein eluted by competition with five washes of 3x Flag peptide using the Flag M purification kit for mammalian expression systems (Sigma-Aldrich). Eluted protein was quantified by BCA protein assay and verified for purity by SDS-PAGE and immunoblot analysis.
Activation of the NFκB and IFNβ promoter was measured by reporter assay. Cells were transfected in 24-well plates with p-125 luc (IFNβ) or NFκB reporter plasmid together with the indicated plasmids. Luciferase activity was measured by the Dual-Luciferase assay kit (Promega). All experiments were performed in triplicate and conducted a minimum of three times.
To measure IFNβ production, the indicated cells were infected with either CVB or VSV and samples of culture supernatant removed at the indicated times. IFNβ levels in culture supernatant were determined by IFNβ ELISA according to the manufacturer's instructions (PBL Biomedical Laboratories).
Nuclear extracts were prepared from HEK293 cells after infection with CVB3 for 12 hr. Cells were washed in ice-cold PBS and isolated by incubation in 10 mM EDTA for 10 min. Cells were pelleted at 1000×g for 5 min, washed in ice-cold PBS, and incubated with buffer A (10 mM HEPES [pH 7.9], 1.5 mM MgCl2, 10 mM KCl, 0.5 mM DTT, 0.5 mM PMSF, and 0.1% NP-40). The pellets were then resuspended in buffer B (20 mM HEPES [pH 7.9], 25% glycerol, 0.42 M NaCl, 1.5 mM MgCl2, 0.2 mM EDTA, 0.5 mM DTT, 0.5 mM PMSF, 5-µg/ml leupeptin, 5-µg/ml pepstatin, 5-µg/ml aprotinin). Samples were incubated on ice for 15 min before being centrifuged at 10,000×g. Nuclear extract supernatants were diluted with buffer C (20 mM HEPES [pH 7.9], 20% glycerol, 0.2 mM EDTA, 50 mM KCl, 0.5 mM DTT, 0.5 mM PMSF).
Data are presented as mean ± standard deviation. One-way analysis of variance (ANOVA) and Bonferroni's correction for multiple comparisons were used to determine statistical significance (p<0.05).
(numbers were taken from GenBank at Pubmed): mitochondrial antiviral signaling protein (MAVS) 57506; Toll/IL-1 receptor domain-containing adaptor inducing interferon-beta (TRIF) 148022, toll-like receptor 3 (TLR3) 7098; Retinoic acid-inducible gene-I (RIG-I) 23586; melanoma-differentiation-associated gene 5 (MDA5) 64135; interferon regulatory factor 3 (IRF3) 3661.
|
10.1371/journal.pcbi.1004036 | Multi-timescale Modeling of Activity-Dependent Metabolic Coupling in the Neuron-Glia-Vasculature Ensemble | Glucose is the main energy substrate in the adult brain under normal conditions. Accumulating evidence, however, indicates that lactate produced in astrocytes (a type of glial cell) can also fuel neuronal activity. The quantitative aspects of this so-called astrocyte-neuron lactate shuttle (ANLS) are still debated. To address this question, we developed a detailed biophysical model of the brain’s metabolic interactions. Our model integrates three modeling approaches, the Buxton-Wang model of vascular dynamics, the Hodgkin-Huxley formulation of neuronal membrane excitability and a biophysical model of metabolic pathways. This approach provides a template for large-scale simulations of the neuron-glia-vasculature (NGV) ensemble, and for the first time integrates the respective timescales at which energy metabolism and neuronal excitability occur. The model is constrained by relative neuronal and astrocytic oxygen and glucose utilization, by the concentration of metabolites at rest and by the temporal dynamics of NADH upon activation. These constraints produced four observations. First, a transfer of lactate from astrocytes to neurons emerged in response to activity. Second, constrained by activity-dependent NADH transients, neuronal oxidative metabolism increased first upon activation with a subsequent delayed astrocytic glycolysis increase. Third, the model correctly predicted the dynamics of extracellular lactate and oxygen as observed in vivo in rats. Fourth, the model correctly predicted the temporal dynamics of tissue lactate, of tissue glucose and oxygen consumption, and of the BOLD signal as reported in human studies. These findings not only support the ANLS hypothesis but also provide a quantitative mathematical description of the metabolic activation in neurons and glial cells, as well as of the macroscopic measurements obtained during brain imaging.
| The brain has remarkable information processing capacity, yet is also very energy efficient. How this metabolic efficiency is achieved given the spatial and metabolic constraints inherent to the designs and energy requirements of brain cells is a fundamental question in neurobiology. The major cell classes in mammalian nervous systems include neurons, glia and the microvasculature that supplies the molecular substrates of energy and metabolism. Together, this neuron-glia-vasculature (NGV) ensemble constitutes the functional unit that underlies the cost infrastructure of computation. In spite of its importance, a comprehensive understanding of this dynamic system remains elusive. While it is well established that glucose feeds the brain, few of the details regarding the destiny of glucose intermediates in metabolic pathways are known. Controversy remains regarding the degree of cooperativity between glia and neurons in sharing lactate, the product of aerobic glycolysis (Warburg effect) and one of the substrates for further energy extraction by oxidative processes. Specifically, while experimental data support the occurrence of a flow of lactate from glia to neurons, the astrocyte-neuron lactate shuttle (ANLS), some theoretical considerations have been proposed to support the occurrence of lactate transport in the other direction (NALS). Our computational model is the first to integrate multiple timescales of the NGV unit. It provides a quantitative mathematical description of metabolic activation in neurons and astrocytes, and of the macroscopic measurements obtained during brain imaging that uses metabolism as a proxy for neuronal activity.
| The mammalian brain exhibits remarkable processing power. It is at the same time energy efficient. The design features that allow such efficient computation are mapped in cellular and molecular components and their roles in information processing. Concurrently, these features are anchored in, and constrained by, the universal metabolic chains that provide energy to cells. Deciphering the metabolic code and the neural code are thus tandem requirements for a comprehensive understanding of brain function. Understanding the metabolic underpinnings of information processing is also of added value to understanding the etiology and progression of neuropsychiatric and neurodegenerative disorders [1, 2]. The picture that emerges from this dynamical system will reflect the cooperative function of neurons, glia and the vascular system.
Glutamate, the brain’s major neurotransmitter, effects numerous cascades and processes in brain cells [3, 4]. Among them, astrocytes couple synaptic activity to energy metabolism via a sodium-dependent uptake of glutamate [5]. The ensuing cascade of molecular events leads to the glycolytic processing of glucose and the release of lactate by astrocytes. A comprehensive model of brain energy metabolism must consider oxidative and non-oxidative glucose consumption, intracellular and extracellular compartmentalization and transport of choke-point metabolic intermediates such as lactate and pyruvate, as well as feedback mechanisms that report local synaptic and intrinsic neuronal activity [6, 7]. These pathways are in turn complicit in the molecular and cellular mechanisms that contribute to the still poorly understood read-out of functional brain imaging [8].
The role of astrocytes and how they metabolically interact with neurons is well supported experimentally; some mostly theoretical considerations, however, have challenged this view. Magistretti and colleagues proposed that clearance of glutamate from the synaptic cleft by astrocytes could be coupled to glycolysis and subsequent lactate production [5]. Lactate produced in this way would then be transported to the extracellular space. Controversy remains surrounding directionality and timing of lactate flow in the brain; while a neuron-to-astrocyte lactate system (NALS) is proposed by some [9, 10], an astrocyte-to-neuron direction (ANLS) is supported by a large set of experimental evidence [11]. Biophysical models also weigh-in on the conditions and sequences of events required for lactate production and consumption [12–15].
The existence of an extracellular pool of lactate likely used as an energy reservoir at the onset of stimulation has been observed in rats and humans [16, 17]. The distribution of monocarboxylate transporters at the membrane of neurons and astrocytes supports the hypothesis of a net transfer of lactate from astrocytes to neurons through the extracellular space [18]. Glial cells have been observed to take up most glucose [19, 20], while neurons are responsible for the largest part of brain oxygen consumption [21, 22]. Additional evidence comes from the direct measure of NADH transients in brain slices, showing that neurons display early oxidative metabolism following presynaptic activity, while astrocytes display a delayed activation of glycolysis but no detectable oxidative response [23]. Nevertheless, the ANLS-hypothesis is still debated and challenged with arguments focusing now on the exact interpretation of the above observations [24, 25].
Nicotine adenine dinucleotide, either oxidized or reduced (NAD+, or NADH), is a workhorse cofactor that acts as a central electron broker for metabolic redox cycles including glycolysis, the citric acid cycle (Krebs, TCA) and oxidative phosphorylation. Owing to its high UV wavelength absorption, it is also responsible for cellular auto-florescence. This coincidence makes it a useful indicator of metabolic activity. NADH is an important metabolic signal because it is produced or used during both mitochondrial activity and activation of the glycolytic pathway, and because it cannot diffuse freely through the mitochondrial membrane but needs to be transported by appropriate shuttles. Fluctuations of the NADH concentration measured in the appropriate cellular compartments can then indicate increased or decreased oxidative and glycolytic metabolism. A critical previous finding in this regard was the observation of early and late activity-dependent phases of metabolic activity with the early phase taking the form of a NADH “dip” and the late phase appearing as a NADH “overshoot” with a longer time constant of decay [23]. Interestingly, these phases also correlate with the fluctuations of the extracellular lactate concentration as determined in animals [16] and humans [26, 27].
The emerging consensus is that the early phase represents NADH depletion in the dendrites of active neurons and that the overshoot represents glycolytic activity that results in the accumulation of NADH. This activity results in the high production of lactate in astrocytes as rapid glycolysis overtakes the subsequent consumption by oxidative pathways [23, 26, 27]. The accumulation and transportation of lactate between glial cells and neurons may in turn serve as an activity-dependent buffer that is informed by the neuronal release and glial uptake of glutamate [5]. It might also act as a signaling molecule to the vasculature [28] or to brain cells via binding to the G-protein coupled receptor GPR81 [29].
The preference for lactate over glucose as an energy substrate in neurons has been demonstrated in vivo as well as in vitro [30, 31], as has a neuroprotective role for lactate in the case of insulin-dependent hypoglycemia [32] and other conditions [33, 34]. The role of the ANLS in homeostatic maintenance involves the regulation of blood glucose [35] and sodium [36]. While the ANLS hypothesis, since its initial formulation [5], does not preclude the use of glucose by neurons as an energy substrate, it has been challenged by some studies defending the view that glucose, rather than lactate, is the sole energy substrate for oxidative metabolism in neurons [37, 38].
Previous modeling efforts have advanced our knowledge of this functional metabolic network by demonstrating that lactate consumption by neurons occurs early in the stimulus regimen and that the early and late lactate transients correspond to the activity of two distinct populations of cells, neurons and glia. The current study builds on and complements those and other models [9, 10, 12–15, 39–42] and addresses unresolved mechanisms of neuron-glial metabolic and vascular coupling. The model is based on several previous studies [13, 40] with five significant improvements: 1) the compartmentalization of NADH between cytosolic and mitochondrial compartments; 2) the linking of metabolic and Hodgkin-Huxley formalisms; 3) the input to the neuronal and astrocytic compartments formulated as a presynaptic glutamatergic stimulation; 4) the model explicitly and continuously updates reversal potentials; and 5) the model was constrained using in vitro data and correctly predicts in vivo results without the need of invoking glycogen (which is deliberately excluded from the model).
In this paper, we will show that a biophysical model of astrocyte-neuron metabolic interactions designed following these principles leads to the presence of an activity-dependent lactate shuttle from astrocytes to neurons and that this model can reproduce the evoked response of NADH in its various compartments as reported by Kasischke and colleagues [23]. We will subsequently show that our biophysical model correctly predicts qualitatively—and to some extent quantitatively—the evoked responses of tissue lactate and tissue oxygen as observed in the rat brain in vivo [16]. Finally, our biophysical model predicts the evoked responses of tissue lactate, of the BOLD signal and the glucose and oxygen consumption as observed in the human brain in vivo.
This in silico model represents a dynamic and integrative analysis of compartmentalized metabolism and its relation to neuronal signaling in the central nervous system. The model was designed based on knowledge of the underlying biophysics and required input parameters and equations from multiple species, time and spatial scales. The model is inspired by previous work from Aubert and colleagues [13, 40] and consists of four compartments: neuron, astrocyte, capillary and extracellular space (see Fig. 1). These compartments are referred to by the subscripts n, g (for glia/astrocyte), c and e respectively. In addition, the neuronal and astrocytic compartments are further divided between cytosolic and mitochondrial sub-compartments to account for the compartmentalization of nicotinamide adenine dinucleotide (NADH). These are referred to by the superscripts cyto and mito. Transport between compartments is noted with the subscripts of both compartments; for instance, transport from the neuronal compartment to the extracellular space is labeled with ne or, conversely, en.
The model is formulated as a series of 33 differential equations adapted from previous work [13, 40] with the following improvements: the compartmentalization of (mostly) NADH between the cytosolic and mitochondrial compartments; the model was joined to a Hodgkin-Huxley-type model [43, 44]; the input to the neuronal and astrocytic compartments is formulated as a glutamatergic input and not anymore as an abstract stimulation; the model explicitly models sodium entry and extrusion in both the neuronal and astrocytic compartments, continuously updating the corresponding reversal potentials. The model, which for the first time bridges mathematical descriptions of energy metabolism and Hodgkin-Huxley equations, was constrained on in vitro data and correctly predicts in vivo results.
Parameters that were difficult to determine experimentally such as transport constants were left free to vary [10, 42]. Free parameters were then optimized so that the model reproduces the experimental results presented in [23]. The number of free parameters was maintained as small as possible by enforcing constraints on the value of metabolites at steady state. Simulations with randomized fluctuations in parameter values (up to plus or minus 10% of reported values) did not reveal significant changes in behavior of the model. This approach successfully predicts qualitatively and quantitatively in vivo measurements in rodents and in humans (see Results, Fig. 5 and 6).
The pooling together of equations from various sources is unfortunately necessary for construction of such a broad and multi-dimensional model. There is no single source of equations that can be tapped for this model. Note however that most equations come from only two published sources. The variables, their steady-state values and the corresponding governing equations are given in Table 1. Equations (A.1)-(A.8), (A.13) and (A.16)-(A.22) are taken from Aubert and Costalat [40] and were originally introduced in ref. [45] for the equations describing metabolism and by Buxton and colleagues [46] for the equations describing the vascular dynamics. Equations (A.9)-(A.12), (A.14) and (A.15) are original. Equations (A.23)-(A.26) describe the neuronal membrane excitability following the Hodgkin-Huxley formalism and the neuronal calcium dynamics. They are adapted from [44].
All the fluxes and currents appearing in Table 1, as well the equations describing the dynamics of the gating variables, are given in Table 2. Like for Table 1, these equations are taken from [40, 44–46] except equations (A.35)-(A.37) which are original. All rates and state variables are given per unit cell volume (neuron or astrocyte) or per unit capillary volume to the exception of JGLCce, JLACec, LACe and GLCe that are given per unit extracellular volume. Mitochondrial and cytosolic NADH levels are given per unit mitochondrial or cytosolic volume respectively.
ADPx is given as a function of the ATP concentration (x stands for n or g). It reads:
A
D
P
x
=
A
T
P
x
2
[
−
q
A
K
+
q
A
K
2
+
4
q
A
K
(
A
/
A
T
P
x
−
1
)
]
(1)
with A = AMPx+ADPx+ATPx = 2.212 mM the total adenine nucleotide concentration and qAK = 0.92 the adenylate kinase equilibrium constant [40, 45]. As a consequence:
d
A
M
P
x
d
A
T
P
x
=
−
1
+
q
A
K
2
−
1
2
u
x
+
q
A
K
*
A
A
T
P
x
u
x
(2)
with ux = q2AK+4·qAK·(A/ATPx-1).
The model receives input from a presynaptic excitatory population. Glutamate released by excitatory presynaptic neurons drives the intracellular sodium concentration in neurons and astrocytes and activates AMPA receptors on neurons, thus inducing a synaptic current Isyn. The presynaptic population contains Nexc excitatory neurons discharging at frequency fexc(t). This presynaptic population thus generates an excitatory conductance gexc(t) given by:
g
e
x
c
(
t
)
=
N
e
x
c
g
¯
f
e
x
c
(
t
)
(3)
with g = 7.8·10-6 mS·cm-2·sec the total surface under the conductance evoked by one excitatory event [47, 48]. The corresponding synaptic current is then given by:
I
s
y
n
(
t
)
=
g
e
x
c
(
t
)
(
ψ
n
−
E
A
M
P
A
)
(4)
with ψn the neuronal membrane voltage and EAMPA = 0 mV the reversal potential of AMPA ionotropic receptors. It is estimated that about two thirds of the current generated at AMPA receptors is due to a flow of sodium ions [49]. Sodium also flows through voltage-dependent sodium channels when the neuron is active (INa). As a consequence, the sodium drive to the neuron is approximated by:
J
s
t
i
m
n
=
S
m
V
n
F
(
2
3
I
s
y
n
−
I
N
a
)
(5)
with Sm·Vn = 2.5·104 cm-1 the ratio between neuronal membrane surface and neuronal volume and F = 9.64853·104 C·mol-1 the Faraday constant. Finally, the glutamate is cleared from the synaptic cleft by excitatory amino acid transporters located on the astrocyte membrane. Those transporters use the electrochemical sodium gradient to transport glutamate with a stoichiometry of three sodium ions for one glutamate molecule. We thus write the sodium drive to the astrocyte as follows:
J
s
t
i
m
g
=
3
Δ
g
l
u
t
N
e
x
c
f
e
x
c
(
t
)
(6)
with ∆glut = 2.25·10-5 mM a constant, which corresponds to the total amount of glutamate released in the synaptic cleft by each presynaptic action potential multiplied by the ratio between synaptic and astrocytic fractional volumes. For the sake of simplicity, we assume that fexc(t) always follows the same temporal dynamics exponentially decaying from f0 = 3.2 Hz to f∞ = 0.5 Hz with a time constant tf = 2.5 sec and Nexc = 1500.
Following in vivo measurements in rodents [50, 51], the cerebral blood flow is modeled as a piecewise double exponential function delayed in time by t1 relatively to the onset of stimulation t0. It reads:
F(t)={F0{F01.1+1.5[exp(−t−t15)−exp(t−t12)]F0+[F(tend)−F0]exp(−t−tend5)}ift<t1t1≤t≤tendt<tend}
(7)
with F0 = 0.012 sec-1 [46]. Typical values are t0 = 0 sec and t0 = 1 sec, tend being the time at which stimulation ends. Two distinct simulation scenarios are considered to mimic in vitro and in vivo conditions. In the in vivo scenario, Equation (7) is used while in the in vitro scenario, the capillary state variables remain constant at their steady-state while the rest of the variables are left free to vary. Our simulations have shown that this is almost equivalent to taking a constant blood flow F(t) = F0.
The blood-oxygen-level-dependent (BOLD) signal is computed following [52]. It is written as a function of the deoxyhemoglobin concentration (dHb) and of the venous volume (Vv):
B
O
L
D
(
t
)
=
V
V
,
0
[
(
k
1
+
k
2
)
(
1
−
d
H
b
d
H
b
0
)
−
(
k
2
+
k
3
)
(
1
−
V
V
V
V
,
0
)
]
(8)
with dimensionless parameters k1 = 2.22, k2 = 0.46 and k3 = 0.43 [40]. The steady-state values of deoxyhemoglobin (dHb0) and venous volume (Vv,0) are given in Table A1.
As noted already by Aubert and colleagues [53], most models of energy metabolism concentrate on erythrocytes, muscles or other organs such as the liver. It is also not clear whether or not parameters drawn from experiments could be directly injected as such into a model without spatial dimensions and without diffusion processes like ours. To circumvent this problem, we proceeded as follows:
First, we chose target steady-state values for the concentration of metabolites following measures reported in the literature. Specifically, we chose the concentration of intracellular sodium following [54], the concentration of intracellular glucose, phosphoenolpyruvate, pyruvate, adenosine triphosphate and phosphocreatine following [55], and glyceraldehyde-3-phosphate following [56]. Finally, the NADH concentration in all four compartments where it appears in the model was chosen following [56] and calculations based on results by Kasischke and colleagues [23].
We then optimized a subset of model parameters (see Table 3) by fitting its predictions to the temporal dynamics of NADH fluorescence as measured by Kasischke et al. [23]. Namely, the dynamics of the NADH concentration in various compartments was extracted empirically from Fig. 4D in [23]. Data points were then fitted with sums of exponentials in order to obtain continuous curves. We then optimized the model by minimizing the distance between the temporal dynamics of NADH in the model and the one in the smoothed curve obtained from [23] using least-square distance as the error measure and using the downhill simplex algorithm. After the optimization converged, we rounded the value of the optimal parameter set and recomputed the steady-state value. All along optimization, we checked that the steady-state was stable by computing its Jacobian matrix (first order approximation) [45]. The parameter set in Table 3 is the set resulting from this procedure.
The simulations were run in MATLAB (The Mathworks, Natick MA, USA). The model was integrated with the ordinary differential equation solver with fixed and optimized parameters (ode15s) that is adapted to stiff systems. We used a time step ∂t = 10-4 sec when the neuron is spiking and ∂t = 1 sec starting one second after the end of presynaptic stimulation. ∂t = 10-4 sec is smaller than the fastest time constant appearing in the Hodgkin-Huxley equations [tauh(-80 mV) = 6.4·10-4 sec]. The second time step (∂t = 1 sec) is small enough for the slow metabolic processes and maintains simulation time and memory usage to reasonable values for an average desktop PC. Simulations take a couple of minutes to execute on a recent laptop.
We developed a model of the coupling between neuronal activity and metabolic response in neurons and astrocytes. The model employed to simulate the neural-glial-vascular (NGV) functional system is composed of four distinct computational units representing a neuron, an astrocyte, a capillary and the extracellular space (Fig. 1). The core of our model is composed of the compartmentalized model of brain energy metabolism recently proposed by Aubert and Costalat [13, 40]. This model connects a model of erythrocyte glycolytic metabolism [45, 56] together with the so-called “Balloon model” of blood flow dynamics [46]. From this starting point, we added a precise description of neuronal membrane excitability formulated within the Hodgkin-Huxley framework [57]. Channels dynamics is drawn from a model proposed by Wang [44]. It includes all the standard Hodgkin-Huxley currents plus a high-threshold calcium current and a calcium-gated potassium current inducing spike-frequency adaptation. The Hodgkin-Huxley model is connected to the metabolic pathways through the electrogenic Na, K-ATPase pump which is responsible for a net outward current and concomitant ATP consumption. We modified the metabolic pathways to include compartmentalization of NADH between the cytosol and mitochondria. To do so, we developed a very simple model of mitochondrial respiration and added NADH malate-aspartate shuttles between the cytosol and mitochondria, drawing inspiration from a model by [58]. Finally, the model is driven by external input modeled as a global excitatory presynaptic activity and coordinated increase of the cerebral blood flow. The presynaptic population is coarsely described through a time-dependent excitatory conductance. This conductance drives sodium flow in neurons through AMPA receptors and action potential-generating voltage-gated sodium channels, and in astrocytes through excitatory amino acid transporters which co-transport glutamate using the sodium gradient. The model is illustrated in Fig. 1 and extensively described in the Methods section.
We first tested the model for its responsiveness to an excitatory stimulus (Fig. 2A) and recorded its voltage response (Fig. 2B). In response to this stimulus, the model generated action potentials within the initial 7 sec of the stimulation. Because of spike-frequency adaptation and of the time course of the excitatory stimulus, the frequency of elicited action potentials quickly decreased until the neuron eventually ceased to fire (Fig. 2B inset). Response trajectories of intracellular sodium, in both the astrocyte (red) and neuron (blue), showed significant differences in both amplitude and duration, with astrocytes exhibiting a smaller but more sustained response and a delayed recovery (Fig. 2C).
We then examined the time course of critical intermediates in energy metabolism in response to the same excitatory stimulus. We first focused on the concentration changes of adenosine triphosphate (ATP) and phosphocreatine (PCr) in the glial and neuronal compartments (Fig. 3). In both cases, the response to the excitatory stimulation, evidenced as a consumption of these energy rich metabolites, was slower in the astrocytic (red) than in the neuronal compartment (blue) (Fig. 3A). And while the decrease in glial ATP surpassed that seen in the neuron (Fig. 3C), the consumption of PCr predominated in the neuron (Fig. 3B). In both compartments, the resulting decrease in ATP concentration was very limited.
Next, we examined the trajectories of nicotinamide adenine dinucleotide (NADH) in response to the stimulation in the astrocytic, neuronal and mitochondrial compartments (Fig. 4A). The dashed lines represent experimental data from Kasischke et al. [23]. Fig. 4A shows the temporal evolution of the concentration of NADH in the astrocytic cytosol, in the neuronal mitochondria and averaged over the whole tissue as evoked by a 20 sec stimulation episode (see Methods). All three curves are in excellent quantitative agreement with the results reported by Kasischke et al. [23]. In particular, the NADH concentration in the neuronal mitochondria displays an initial dip of about-10% indicating a strong increase of the oxidative metabolism in neurons (Fig. 4B). It then returns towards its baseline before the presynaptic bombardment has finished and finally slightly overshoots in the poststimulus period. On the contrary, the NADH concentration in the astrocytic cytosol increases significantly only about 10 sec after the onset of the stimulation and displays a long-lasting monophasic behavior. This corresponds to a strong and sustained increase of the glycolysis in this compartment (Fig. 4B).
The initial dip in the neuronal mitochondria is the result of consumption of NADH to produce ATP. A recovery and rebound results when NADH is produced from the consumption of lactate imported into the neuronal cytosol from the extracellular space (Fig. 4C). In both Fig. 4B and C, it can be seen that the astrocytic response is slower than the neuronal response. In particular, both oxygen and glucose consumption increase immediately at the beginning of the stimulation in the neuronal compartment. Partially supporting this metabolic activity, neurons immediately start to import lactate from the extracellular space (Fig. 4C). On the contrary, the increase in glucose consumption by astrocytes (Fig. 4B) is more gradual and the increase in lactate export by astrocytes to the extracellular space is slightly delayed (Fig. 4C). The initial release of presynaptic glutamate with subsequent neuronal activity and reuptake into astrocytes lead to the increase in intracellular sodium concentration and activation of the Na, K-ATPase imposing, along with the conversion of glutamate to glutamine in astrocytes, an increased metabolic demand. However, as can be seen in Fig. 2C, the increase in intracellular sodium is slower and more gradual in the astrocytic compartment leading to the 10 second delay in the glial cell metabolic response to the stimulation. Finally, the dynamics of tissue NADH (Fig. 4A) is mirrored in the predicted tissue and extracellular lactate concentrations (Fig. 4D).
The utilization of glucose and oxygen by neurons and astrocytes during this 20 sec stimulation episode is shown in Fig. 4B. For model optimization, we imposed that the largest fraction of glucose goes to astrocytes while the largest fraction of oxygen goes to neurons [42]. We observed that this bias is further increased during stimulation. The neuronal oxygen utilization immediately increased at the onset of stimulation in register with the initial dip of the NADH in the neuronal mitochondria. This is consistent with reports that the astrocytic fraction of glucose utilization increases during stimulation [19].
One of the hallmarks of a successful model is its ability to reproduce and explain empirical observations. We thus now turn to an in vivo situation and compare the predictions of the mathematical model we designed in the precedent sections to two experiments carried out in rats. Tissue oxygen and lactate during stimulus (Fig. 5B and C), and lactate transfer between compartments were compared (Fig. 5D).
Upon stimulation, CBF increases after a delay of ~1 sec (see Equation 7), quickly peaks before relaxing to an elevated plateau. This pattern matches neurovascular responses observed in rodents in response to sustained sensory stimulation (for instance by mechanical activation of the whiskers [50, 51] (Fig. 5A)). The 1 sec delay seen in panel A is hardcoded in the model (see Equation 7). This matches our own observations that CBF only starts to increase above its baseline ~0.5–1 sec after the onset of stimulation [51, 59].
Because neuronal activity slightly precedes functional hyperemia, the oxygen concentration initially dips below its resting value before rising as cerebral blood flow finally increases after the 1 sec delay (Fig. 5B inset). The dip in oxygen below baseline levels after stimulation has ceased reflects the rapid decrease in the replenishment rate by blood at a time when oxygen is still consumed to replenish the ATP that is used to fuel the Na, K-ATPase pump.
Extracellular lactate is consumed throughout the stimulation. Its concentration initially dips until the cerebral blood flow increases and leads to a sustained overproduction of lactate (Fig. 5C). At rest, the astrocytic compartment exports lactate to the extracellular space, part of which is taken up by the neuronal compartment for energy production (Fig. 5D), the rest being exported to the circulation. Upon stimulation, export of extracellular lactate to the circulation is reduced while import into neurons is increased. Export of lactate to the extracellular space by the astrocytic compartment is also increased but with a delay and explains the initial dip in concentration. In the recovery period, all transports slowly return back to their baseline values explaining the long lasting overshoot of extracellular and tissue lactate. Export of extracellular lactate to the circulation is durably increased in the recovery period (Fig. 5D; pink line).
Fig. 5 shows results of simulations independent from the simulations that yielded Fig. 2 to 4 where the model was constrained to reproduce the experimental results from ref. [23]. In these new simulations, not only is the temporal course of tissue oxygen qualitatively predicted, but the amplitude of fluctuations is, to some extent, quantitatively predicted as well. Our model predicts that the oxygen pressure first drops by -1.7% (inset), then overshoots at +17.7% before stabilizing 2.4% above its baseline in the last 20 sec of the stimulation. Finally, it undershoots to -2.6% in the post-stimulus period. These figures are to be compared with the values reported by [50], namely, an initial drop at -1.8%, an overshoot at +19.9%, a stabilization 1.7% above the baseline and a final undershoot at -3.1%.
The consumption of lactate closely tracks the stimulation dependent oxygen consumption but lacks the inflections corresponding to CBF changes as it is less affected by the blood flow (Fig. 5D). Lactate is exported by the astrocyte to the extracellular space (thick red line) and imported by the neuron from the extracellular space (thick blue line; net import is negative by convention in this model). A small amount of lactate is exported from the extracellular space to the capillary at baseline and this export increases by 69% after the end of the stimulation (pink line). The thin red line denotes the activity of the lactate dehydrogenase converting pyruvate into lactate in the astrocytic cytosol while the thin blue line denotes the activity of the lactate dehydrogenase converting lactate into pyruvate in the neuronal cytosol (again, the negative sign is a convention). These net transfers all contribute to the evolution of the tissue and extracellular lactate concentrations (Fig. 5C). This prediction closely matches the experimental results of Hu and Wilson [16] (their Fig. 1).
We then compared new simulations to measurements from human subjects [26, 52, 60]. As in Fig. 5, the neurovascular response was adapted from experimental measurements (Fig. 6A). Simulated levels of lactate (Fig. 6B), the cerebral metabolic rate for glucose (Fig. 6C), the cerebral metabolic rate for oxygen (Fig. 6D), the ratio of cerebral uptake of O2 to cerebral uptake of glucose (Oxygen-Glucose Index or OGI) (Fig. 6E), as well as the blood-oxygen-level dependent signal (Fig. 6F) all supported a validation of the model.
The model qualitatively and to some extent quantitatively predicted well-known features of these various macroscopic observables, including the BOLD signal [52, 60], despite extrapolation from multiple sources. For instance, after an initial dip [17], the tissue lactate concentration reached a plateau that last until the end of the stimulation [61]. The initial dip in extracellular lactate (Fig. 6B inset) is representative of a surge in lactate consumption at the beginning of neural activity, and a symptom of the intrinsic latency in the start-up of the ANLS and of functional hyperaemia. The time lag between the onset and offset of neuronal activation and the onset and offset of CBF also explain other transients such as the slow recovery of the lactate concentration after the cessation of neural activity [61]. Relative increase in glucose and oxygen consumption also matched experimental results and led to a decreased OGI during stimulation [26].
There is mounting evidence in support of a metabolic link between neurons and glial cells. The most prominent experimentally-based conceptual model of neuron-glia metabolic coupling, the astrocyte-neuron lactate shuttle (ANLS), has raised controversies, not so much for the proposed energy-dependent link between the two cell types, but because certain details and mechanisms are debated [11, 62]. Although experimental evidence gathered over the last two decades largely supports it [11], experimentally untangling this system is challenging.
Our detailed biophysical model of the NGV ensemble expands on previous models in four distinct ways. First, a shuttling of lactate from astrocytes to neurons emerged in response to activation. Second, the model is consistent with increased neuronal oxidative metabolism and delayed increased astrocytic glycolysis for generating the activity-dependent NADH transients. Third, the model correctly predicts the dynamics of tissue lactate and oxygen as observed in vivo in rats. Fourth, the model correctly predicts with good quantitative precision the temporal dynamics of tissue lactate, CMRglc, CMRO2 and of the BOLD signal as reported in human studies. These findings not only support the ANLS hypothesis but also provide a quantitative mathematical description of the metabolic activation in neurons and astrocytes, as well as the macroscopic measurements obtained with brain imaging techniques.
The blood oxygen-level dependent (BOLD) signal which forms the basis of the functional magnetic resonance imaging (fMRI) technology reports fluctuations in brain activity, the molecular and cellular mechanisms of which are still incompletely understood [8, 63]. Although somewhat limited in representing detailed processes, our use of the Buxton model was sufficient to correctly predict known features of the BOLD signal (Fig. 6F). Ultimately, modeling efforts that build on our work will have to include more detailed descriptions of blood flow regulation. For instance, blood flow regulation in the brain was recently suggested to happen in the microvasculature at the capillary level by active dilation of pericytes [64]. Subsequent efforts will need to focus on the daunting task of modeling the numerous pathways that relate neuronal activity to functional hyperaemia [65].
Recent modeling of the neuron-astrocyte cross-talk during oscillations linked to blood oxygenation levels verified the possibility that the slow fMRI BOLD signals might reflect the spontaneous ongoing activity of neuroglial networks [66]. Our results support this view by accurately modeling results from human imaging experiments [52, 60]. A course for future modeling will be to examine and model data from multi-modal imaging experiments [67].
As noted in the introduction, controversy surrounds the directionality of lactate flow in the brain with a neuron-to-astrocyte direction proposed by some. Here, following the arguments delineated in Jolivet et al. [42], we imposed that the largest fraction of glucose should go to astrocytes (there is no controversy that neurons are responsible for the vast majority of oxygen consumption). We derive confidence in our model from the fact that it correctly predicts a vast array of in vivo experimental findings while being only loosely constrained by in vitro experimental findings and by the imposed compartmentalization of glucose uptake between neurons and astrocytes. As argued in Jolivet et al. [42], metabolic shuttling between the astrocytic and neuronal compartments originates from the imbalance between the high oxygen consumption of neurons and their limited glucose utilization for ATP production purposes. Unlike astrocytes, neurons are unable to up-regulate their rate of glycolysis in response to increased activity due to constitutive inhibition of the rate limiting glycolytic enzyme Pfkfb3 (6-phosphofructo-2-kinase/fructose-2, 6-bisphosphatase 3) [68]. This mechanism is crucial to neuronal defenses against reactive oxygen species. Additional in vitro mechanisms support that compartmentalization includes the existence of glutamate-induced glycolysis in astrocytes [5, 69], glutamate-induced inhibition of glucose transport in neurons [70] and the involvement of extracellular increases in potassium inducing an Na, K-ATPase-dependent activation of glycolysis in astrocytes [71]. However, it is to be noted that if the proportion of glucose directly taken up by neurons was to be increased, the astrocyte-to-neuron lactate shuttle would be reduced in amplitude, and its direction eventually reversed if neurons were consuming glucose in excess of what they oxidize (see [42] for further discussion of this question).
Metabolic phenotypes have been suggested that convey a metabolic identity depending on how they utilize the various oxidative and non-oxidative pathways [14]. The re-equilibration of the constituents of extracellular space, a process involving physical flushing mediated by astrocytes, is also now thought to be one of the key housekeeping functions of sleep [72] and the disruption of normal metabolic processes is suggested to underlie the progression of neurodegenerative diseases such as Alzheimer’s [73]. Lactate, whether sourced from glia or plasma, is associated with neuroprotection [32–34]. Assuming the ANLS is indeed taking place, we are left to question: What is it good for?
Neurons do not appear to suffer functional consequences as a result of their metabolic peculiarities. Lactate can sustain prolonged firing in neurons more efficiently than glucose in culture and can preferentially support activity in both resting and active states in vitro and in vivo [30, 31, 74]. Recently, it was shown in the subfornical organ that this pathway can also affect the dynamics of the local neural network by modulating the excitability of GABAergic neurons through the regulation of ATP-dependent potassium channels [36]. It is not necessary perhaps to preclude the use of glucose by both neurons and glia under certain circumstances as has been suggested by a computational model studying the ATP supply to neurons under hypoxic conditions [75], and as is indeed suggested by our own results (see Fig. 4B, 5D and 6C). Finally, extracellular lactate might also act as a messenger to the vasculature [28] and it is thus possible that the ANLS plays a role as one of the pathways regulating functional hyperaemia.
Astrocytes also support memory formation by supplying neurons with lactate [76]. So central is the ANLS to the normal function of the brain that learning, as measured by LTP and long term memory formation in the hippocampus of rats, is abolished by interfering with the transport of lactate from astrocytes to neurons [76]. Consistent with those findings, lactate, but not glucose, has been show to induce the expression of plasticity genes such as Arc, Zif 268 and BDNF in vitro in neurons and in vivo [77]. Further, the mechanism by which glucose enhances memory storage has been shown to involve the neuronal consumption of lactate [78].
We present here the first temporal multi-scale model of the NGV that accurately reflects experimental observations in multiple settings and organisms. These findings not only support the ANLS hypothesis but also provide a quantitative mathematical description of the metabolic activation in neurons and astrocytes, as well as of the macroscopic measurements obtained with functional brain imaging techniques.
|
10.1371/journal.pgen.1004431 | Muscle Structure Influences Utrophin Expression in mdx Mice | Duchenne muscular dystrophy (DMD) is a severe muscle wasting disorder caused by mutations in the dystrophin gene. To examine the influence of muscle structure on the pathogenesis of DMD we generated mdx4cv:desmin double knockout (dko) mice. The dko male mice died of apparent cardiorespiratory failure at a median age of 76 days compared to 609 days for the desmin−/− mice. An ∼2.5 fold increase in utrophin expression in the dko skeletal muscles prevented necrosis in ∼91% of 1a, 2a and 2d/x fiber-types. In contrast, utrophin expression was reduced in the extrasynaptic sarcolemma of the dko fast 2b fibers leading to increased membrane fragility and dystrophic pathology. Despite lacking extrasynaptic utrophin, the dko fast 2b fibers were less dystrophic than the mdx4cv fast 2b fibers suggesting utrophin-independent mechanisms were also contributing to the reduced dystrophic pathology. We found no overt change in the regenerative capacity of muscle stem cells when comparing the wild-type, desmin−/−, mdx4cv and dko gastrocnemius muscles injured with notexin. Utrophin could form costameric striations with α-sarcomeric actin in the dko to maintain the integrity of the membrane, but the lack of restoration of the NODS (nNOS, α-dystrobrevin 1 and 2, α1-syntrophin) complex and desmin coincided with profound changes to the sarcomere alignment in the diaphragm, deposition of collagen between the myofibers, and impaired diaphragm function. We conclude that the dko mice may provide new insights into the structural mechanisms that influence endogenous utrophin expression that are pertinent for developing a therapy for DMD.
| Duchenne muscular dystrophy (DMD) is a severe muscle wasting disorder caused by mutations in the dystrophin gene. Utrophin is structurally similar to dystrophin and improving its expression can prevent skeletal muscle necrosis in the mdx mouse model of DMD. Consequently, improving utrophin expression is a primary therapeutic target for treating DMD. While the downstream mechanisms that influence utrophin expression and stability are well described, the upstream mechanisms are less clear. Here, we found that perturbing the highly ordered structure of striated muscle by genetically deleting desmin from mdx mice increased utrophin expression to levels that prevented skeletal muscle necrosis. Thus, the mdx:desmin double knockout mice may prove valuable in determining the upstream mechanisms that influence utrophin expression to develop a therapy for DMD.
| Duchenne muscular dystrophy (DMD) is an X-linked muscle disorder that affects approximately 1∶4000 boys [1]. DMD is caused by mutations in the large 2.2 Mb dystrophin gene [2], [3]. The dystrophin protein functions as a large molecular spring that connects the skeletal muscle cytoskeleton to the transmembrane dystrophin glycoprotein complex (DGC) [4]–[9]. The lack of dystrophin in DMD is accompanied by a significant reduction in the expression of the DGC leaving the membrane highly susceptible to contraction-induced injury and hypoxic stress [10]–[18]. DMD patients develop severe cardiorespiratory distress and generally live into their third decade with the help of palliative care.
The absence of dystrophin leads to various molecular and cellular homeostatic responses that slow the loss of skeletal muscle [19]. For instance, the dystrophin paralog, utrophin is expressed on the sarcolemma of dystrophic fibers acting to mitigate necrosis [20]–[25]. Skeletal muscle necrosis in the mdx mouse model of DMD is prevented by the expression of a full-length utrophin transgene when expressed at twice the levels of the endogenous utrophin [26]. Utrophin expression in DMD patients correlates with the severity of disease and time to wheelchair demonstrating the therapeutic potential of utrophin in humans [25], [27]–[31]. An utrophin therapy would benefit all DMD patients and circumvent a potential T-cell mediated immune response that could impair the long-term benefit of prospective dystrophin replacement strategies [32]–[34]. Accordingly, increasing the expression of utrophin is a primary target for therapy of DMD [33]. While promising utrophin-mediated therapies are being tested in clinical trials [33], [35], the mechanisms that influence utrophin expression are not fully understood.
Utrophin is normally expressed on the sarcolemma of developing and regenerating muscle fibers [21], [22], [36]. Utrophin is ultimately replaced by dystrophin in the sarcolemma of normal maturing fibers and remains concentrated at the neuromuscular and myotendinous junctions. However, low levels of utrophin can remain on the sarcolemma of dystrophin-deficient mdx mouse skeletal muscle fibers independent from muscle regeneration [37]. While various factors that influence utrophin expression and stability within the sarcolemma are well described [33], [38], [39], the upstream mechanisms are less clear. We recently discovered an increase in utrophin expression in mdx4cv mice expressing the microdystrophinΔR4–R23 transgene [40]. The polyproline site within hinge 2 of microdystrophinΔR4–R23 led to myotendinous strain injury and the formation of ringed fibers where the peripheral sarcomeres surround the central sarcomeres [40], [41]. Notably, we found a significant increase in utrophin expression within the limb muscles that contained ringed fibers, but not in the diaphragm muscles that did not contain ringed fibers [40]. Accordingly, we hypothesize that structural changes within skeletal muscle can influence utrophin expression, independent from muscle regeneration.
To examine the role of muscle structure on the pathogenesis of DMD we generated mdx:desmin double knockout (dko) mice. Desmin is an intermediate filament protein that maintains the highly ordered structure of striated muscles by connecting the sarcomeres to the sarcolemma and organelles [42]–[45]. Desmin influences the organization of dystrophin and ankyrin in a costameric lattice that connects the Z-disks of peripheral sarcomeres to the sarcolemma [46], [47]. Desmin−/− mice develop a severe dilated cardiomyopathy with a mild skeletal myopathy [45], [48]. The skeletal myopathy is associated with misaligned sarcomeres and changes to the distribution and function of mitochondria [45], [48]. Here, we report that a ∼2.5-fold increase in utrophin expression in dko skeletal muscle fibers prevented necrosis in a fiber-type specific manner.
We initially found that desmin expression was increased in mdx4cv mouse skeletal muscles by western analysis of whole muscle lysates (Fig. 1A), confirming previous reports in mdx mice [49], [50]. To examine the role of desmin in the pathogenesis of DMD we bred mdx4cv:desmin+/− mice to generate the dko pups (N = 5, F>4). The dko pups were born in the expected Mendelian ratios [71 (25%) +/+; 144 (51%) +/−; 67 (24%) −/−]. We examined only the male mice for this study, as DMD patients are males. The dko mice developed a mild kyphosis (Fig. 1B). The genotype was confirmed by immunohistological analyses of dystrophin and desmin expression in skeletal muscle (Fig. 1C). The dko mice gained less body mass than the wild-type (24%), desmin−/− (18%), and mdx4cv controls (36%; P<0.001 one way ANOVA; Fig. 1D). The desmin−/− and dko mice were euthanized when they lost body mass and/or exhibited labored breathing and reduced mobility consistent with cardiorespiratory failure. Kaplan-Meyer survival analyses demonstrated a significantly reduced lifespan in the dko mice with a median survival of 76 days for males compared to a median survival of 609 days for the desmin−/− males (Fig. 1E, P<0.001). The average lifespan for mdx4cv males is 21.5 months [51]. We chose a time point of 11 weeks for the experiments in this study, unless otherwise stated. Approximately a quarter of the dko mice (22%) developed malocclusion, which contributed to the reduced body mass and increased mortality rate particularly in mice younger than 8 weeks of age. The malocclusion was treated with trimming the teeth every week and feeding the mice crushed food pellets mixed with hydrated gel. Malocclusion consistently presented in dko mice through various backcrosses suggesting that this was likely a phenotype of the dko mice and not a separate genetic defect. Furthermore, none of the wild-type, desmin−/− or mdx4cv mice developed malocclusion during the course of this study. The dko mice that developed malocclusion were included for body mass and survival analysis, but not for further analyses.
We next examined the gross dystrophic histopathology in various limb and respiratory muscles. Wild-type mice had few central nuclei (<1%), and no detectable calcified or necrotic fibers (Fig. 2A-D). Desmin−/− mice had a mild skeletal myopathy with a low level of central nuclei (∼5%) and rare necrotic fibers (Fig. 2AD), but no calcification was evident (Fig. 2A,C), as previously described [45], [48], [52]. The mdx4cv skeletal muscles were highly dystrophic with predominantly centrally nucleated fibers (Fig. 2A,B). Of the different limb and respiratory muscles we examined, only the mdx4cv diaphragms consistently contained calcified fibers (Fig. 2A,C), whereas all mdx4cv muscles contained patches of necrotic fibers (Fig. 2A,D). The proportion of dko limb and respiratory skeletal muscles with central nuclei was significantly reduced when compared to the mdx4cv muscles (Fig. 2A,B). None of the dko skeletal muscle fibers were calcified and there were 96% fewer necrotic fibers than the mdx4cv gastrocnemius muscles (P<0.001; Fig. 2A,C,D). Inflammation was also reduced in the dko gastrocnemius muscle with a 93% reduction in macrophages (P<0.01; Fig 2A,E) and an 82% reduction in CD3 positive T-lymphocytes (P<0.001; Fig. 2A,F) when compared to the mdx4cv controls. Thus, multiple indices of dystrophic histopathology in the mdx4cv mice were improved by the absence of desmin.
We next examined whether the dystrophic pathology in the dko muscles was improved by an increase in utrophin expression. We examined the gastrocnemius muscle because of its distinct fiber-type distribution. Utrophin was restricted to the neuromuscular junctions in mature (11 week) wild-type and desmin−/− skeletal muscle fibers (Fig. 3A). Utrophin was expressed at low levels on the extrasynaptic sarcolemma in mdx4cv muscles (Fig. 3A), as previously described in mdx mice [21], [22], [36]. Utrophin was highly expressed in the dko extrasynaptic sarcolemma in some, but not all of the gastrocnemius muscle fibers (Fig. 3A). We next performed a titration of dko utrophin by western analyses to generate a non-linear regression to quantitate the changes in utrophin expression (Fig. S1). The significant increase in utrophin expression in mdx4cv mice compared to wild-type mice was confirmed by western analysis of total gastrocnemius muscle lysates (Fig. 3B; P<0.001). Importantly, we found a 2.54-fold increase in utrophin expression in the dko when compared with the mdx4cv controls (Fig. 3B; P<0.001). Because not all myofibers express utrophin in the dko we next quantitated the level of utrophin fluorescence intensity on the sarcolemma. We quantitated utrophin fluorescence in the wild-type sarcolemma as the negative control and the wild-type neuromuscular synapse as the peak of detection to ensure our quantitation is not beyond the limits of detection. The fluorescence intensity of utrophin was significantly increased in mdx4cv muscles compared to wild-type muscles (P<0.001; Fig. 3C). The utrophin fluorescence intensity increased by 2.86-fold in the dko sarcolemma when compared to the mdx4cv (P<0.001). To test whether this increase in fluorescence intensity in the dko reached therapeutic levels, we compared mdx:utrophin double knockout muscles treated with microutrophinΔR4–R21 using the same gastrocnemius muscles from our previous study [53], which demonstrated that microutrophinΔR4–R21 prevented skeletal muscle necrosis. We found that the sarcolemmal fluorescence intensity of utrophin was increased by 22% in the dko muscles when compared to the mdx:utrophin double knockout muscles expressing microutrophinΔR4–R21 (P<0.01). We found no change in utrophin mRNA in the gastrocnemius muscles of wild-type, desmin−/−, mdx4cv and dko mice, when measured by qPCR (Fig. 3D). Upregulation of utrophin was associated with a reduction in necrosis and regeneration in the dko, as only 9% of the fibers with extrasynaptic utrophin had central nuclei compared with 46% central nuclei in fibers without extrasynaptic utrophin (P<0.001; Fig. 3E). Thus, an increase in utrophin expression in a fraction of the dko muscle fibers prevented cycles of necrosis and regeneration.
Utrophin expression is found on the sarcolemma of all developing wild-type muscle fibers and subsequently becomes restricted to the neuromuscular junctions [21], [54]. The prevention of skeletal muscle necrosis in the dko mice implied that the developmental loss of utrophin expression from the extrasynaptic sarcolemma did not occur. Furthermore, the expression of utrophin on the extrasynaptic sarcolemma of a portion of dko fibers suggests that utrophin may be expressed in certain muscle fiber types. To test this, we compared the expression of the utrophin A isoform relative to muscle fiber types at 3 weeks of age (Fig. 4). We found that utrophin was near absent from the extrasynaptic sarcolemma of wild-type gastrocnemius muscles by 3 weeks of age (Fig. 4). We found utrophin in the cytoplasm of a portion of the wild-type fast 2b fibers (Fig. 4). Furthermore, antibodies to the utrophin A isoform labeled blood vessels in wild-type muscles at 3 weeks of age (Fig. 4), but not at 11 weeks of age (Fig. 3), which was similar to the immunohistochemical staining pattern of the utrophin A isoform in humans [55]. Utrophin expression was absent from the extrasynaptic sarcolemma in most fast 2b fibers in desmin−/−, mdx4cv and dko muscles (Fig. 4). However, utrophin remained at low levels on the sarcolemma of 1a, 2a and 2d/x fiber types in desmin−/− and mdx4cv gastrocnemius muscles. The reduced utrophin expression in the extrasynaptic sarcolemma of mdx4cv muscles coincided with the appearance of patches of necrotic fibers (Fig. 4). In contrast, utrophin prevented skeletal muscle necrosis in the dko muscles by remaining on the extrasynaptic sarcolemma of maturing 1a, 2a and 2d/x fiber-types (Fig. 4). We next performed a titration of utrophin by western analyses of the 3-week-old dko muscles to generate a non-linear regression to quantitate the changes in utrophin expression (Fig. S2). We found a 29.6% increase in utrophin in the mdx4cv muscles compared to wild-type controls (Fig. 4B; P<0.05). Utrophin in the dko was increased by a further 60.9% compared to the mdx4cv muscles (P<0.001). Similar to 11 weeks of age (Fig. 3D), we found no change in the relative amounts of mRNA at 3 weeks of age when comparing all genotypes (Fig. 4C). Thus, utrophin expression was increased in the dko in a fiber-type specific manner to prevent necrosis.
To examine whether utrophin prevented necrosis by maintaining the integrity of the muscle membrane, we systemically delivered 200 µl of 1% (w/v) Evan's blue dye (EBD) into the mdx4cv and dko mice and looked for permeable skeletal muscle fibers (Fig. 5A). We found large patches of skeletal muscle fibers in the mdx4cv mice that were permeable to EBD (Fig. 5A), as previously described [40]. Utrophin was selectively expressed in the dko 1a, 2a and 2d/x fiber types and prevented the infiltration of EBD into these fibers (Fig. 5A). This correlated with an ∼80% reduction in centrally nucleated 1a, 2a and 2d/x fiber types in the dko compared to the corresponding mdx4cv muscles (P<0.001; Fig. 5B). Only the fast 2b fibers in the dko were permeable to EBD, which correlated with an ∼5 fold increase in centrally nucleated 2b fibers when compared with the other fiber-types in the dko (P<0.001; Fig. 5B). The total number of permeable fibers in the dko gastrocnemius muscles was ∼91% less than the mdx4cv muscles (Fig. 5C; P<0.001). Thus, utrophin prevented necrosis in the dko 1a, 2a and 2d/x fiber types by maintaining the integrity of the membrane.
We found a distinct separation of the fast 2b fibers from the 1a, 2a and 2d/x fiber types in the dko gastrocnemius muscles suggestive of a fiber-type switch in the dko muscles (Fig. 5A). We examined the fiber-type proportions in the smaller soleus muscle that contains all fiber-types in wild-type C57Bl/6mice. Analysis of fiber-type proportions in the soleus muscles at 11 weeks of age revealed a significant shift from the 2a fibers in the wild-type toward the slow 1a fibers in the desmin−/−, mdx4cv and dko muscles (P<0.001; Fig. S3). However, we found no significant change in fiber-type proportions when comparing between the desmin−/−, mdx4cv and dko muscles (Fig. S3). Thus, the skeletal muscle fiber-types were redistributed in the dko muscles, but we found no evidence of a fiber-type switch.
The increase in utrophin on the dko sarcolemma (Fig. 3C) may have resulted from reduced surface area of the 1a, 2a and 2d/x fibers compared with the corresponding mdx4cv muscles. However, the fiber area of 1a, 2a and 2d/x fiber types within the gastrocnemius muscles was unchanged when comparing wild-type, desmin−/−, mdx4cv and dko muscles (Fig. 5D). The fast 2b fibers in the desmin−/− and mdx4cv gastrocnemius were hypertrophic when compared to wild-type muscles (Fig. 5D). In contrast, the fast 2b fibers in the dko muscles were selectively atrophic. The desmin−/− muscles contained some smaller caliber fibers that increased the overall variability in muscle fiber area. The muscle fiber areas were highly variable in the mdx4cv muscles. Thus, the increase in utrophin expression on the dko sarcolemma did not result from changes in the average area of 1a, 2a and 2d/x fiber types.
We also found a 36% reduction in the proportion of centrally nucleated fast 2b fibers in the dko when compared to the mdx4cv fast 2b fibers (P<0.01; Fig. 5B), which was consistent with the low level of central nuclei in utrophin negative fibers in the dko (46%) compared to all mdx4cv control fibers (76%) (Fig. 3E). To directly test whether utrophin-independent mechanisms were influencing the dystrophic pathology we performed a more detailed examination of the most superficial region of the gastrocnemius muscles that contained a near pure population of fast 2b fibers (Fig. 6). We found a significant reduction in the extrasynaptic utrophin expression on the fast 2b fibers in the dko compared with mdx4cv muscles (Fig. 6A,B; P<0.001). Moreover, there was a significant reduction in the number of fast 2b fibers expressing extrasynaptic utrophin in the dko when compared to the mdx4cv fast 2b fibers (Fig. 6A,C; P<0.05). Utrophin was expressed on the extrasynaptic sarcolemma in groups of regenerating mdx4cv 2b fibers as the myofibers expanded toward the basal lamina shell (Fig. 6D). Utrophin expression was maintained on the mdx4cv sarcolemma as the muscles matured and developmental myosin heavy chain dissipated (Fig. 6D). In contrast, examination of four dko gastrocnemius muscles revealed that the regenerating 2b fibers were directly enveloped by the basal lamina rather than utrophin (Fig. 6D). Together, these results demonstrate that utrophin expression was reduced in the extrasynaptic sarcolemma of dko fast 2b fibers. Thus, utrophin-independent mechanisms were also mitigating the dystrophic pathology of dko muscles.
The regenerative capacity of skeletal muscles depleted of desmin is profoundly impaired in cell culture [56], [57]. However, muscle generation in desmin−/− skeletal muscles in vivo is apparently normal [58]. Desmin−/− muscles injured with cardiotoxin can lead to persistent expression of developmental myosin heavy chain [59]. We found that regenerating fibers in uninjured gastrocnemius muscles were rare (up to 2 fibers) in the wild-type and desmin−/− mice (Fig. 7A,B). The mdx4cv muscles contained patches of regenerating fibers (Fig. 7). However, the dko muscles contained 47% fewer regenerating fibers than the mdx4cv muscles (P<0.01; Fig. 7A,B). To examine whether the regenerative capacity of muscles was impaired in the dko we delivered notexin to injure the gastrocnemius muscles and examined the muscles 4 and 6 days post injury. We found that regenerating fibers were expressing developmental myosin in wild-type, desmin−/−, mdx4cv and dko treated muscles at 4 days post injury (Fig. 7). At 6 days post injury we found that half (2 out of 4) of the injured wild-type muscles expressed developmental myosin (Fig. 7). Neither the desmin−/−, mdx4cv or dko muscles expressed developmental myosin 6 days post notexin injury (Fig. 7). We found no other overt changes in the regenerative capacity of the muscles when comparing the different strains of mice (Fig. 7). Thus, the improved dystrophic pathology in the dko muscles did not result from overt changes to the regenerative capacity of the skeletal muscles.
We next examined whether the significant increase in utrophin expression in the dko muscles restored the expression of β-dystroglycan and the NODS complex to the sarcolemma. Adjacent sections of gastrocnemius muscles revealed that β-dystroglycan and members of the NODS complex were concentrated within the sarcolemma of wild-type and desmin-/- skeletal muscles (Fig. 8A). The expression of β-dystroglycan and the NODS complex were increased in the desmin-/- mice (Fig. 8B,C), as previously described [60]. The expression of β-dystroglycan and the NODS complex at the sarcolemma of mdx4cv skeletal muscles were significantly diminished (Fig. 8), as previously described [40], [61] (Fig. 8). The increase in utrophin expression in the dko sarcolemma was accompanied by the increased concentration of β-dystroglycan (Fig. 8A). Immunoblots of whole muscle lysates revealed no significant difference in β-dystroglycan expression when comparing the wild-type or the mdx4cv controls with the dko (Fig. 8B,C). However, the expression of the NODS complex on the sarcolemma of dko muscles was not restored (Fig. 8).
Desmin can interact with α-dystrobrevin in the NODS complex indirectly through synemin, syncoilin and dysbindin [46]. Therefore, we examined whether desmin expression influenced the restoration of the NODS complex (Fig. S4). We found that utrophin was expressed on the sarcolemma of 4-week-old mdx4cv soleus muscles with minimal expression of the NODS complex (Fig. S4). Thus, the lack of the NODS complex on the sarcolemma of dko skeletal muscle fibers did not result from the absence of desmin.
We next examined whether diaphragm function in the dko was influenced by structural defects within and around the muscles. We measured the specific contractile force of diaphragm strips in vitro. We found that the specific force production of the desmin−/− diaphragm was similar to wild-type at 11 weeks of age (Fig. 9A). In contrast, the specific force production of both mdx4cv and dko diaphragms were significantly diminished (Fig. 9A; P<0.001). Detailed histological analyses of the mdx4cv and dko diaphragms revealed that utrophin colocalized with α-sarcomeric actin in a costameric lattice (Fig. 9B). However, the alignment of α-sarcomeric actin in the dko was severely perturbed similar to the rectilinear pattern of utrophin (Fig. 9B). Electron microscopy analyses revealed that the sarcomeres aligned in wild-type muscles, but this alignment was impaired in desmin−/− muscles (Fig. 9C), as previously described [44], [45]. The alignment of sarcomeres in mdx4cv myofibers was similar to wild-type (Fig. 9C). However, the alignment of sarcomeres in the dko was severely impaired within and between individual muscle fibers (Fig. 9C). Gross histological analyses of the diaphragm revealed a 1.83-fold increase in the deposition of collagen in desmin−/− compared to wild-type (P<0.05; Fig. 9D,E). The mdx4cv diaphragms were significantly larger and contained proportionally more collagen than wild-type (4.05-fold increase; P<0.001) and desmin−/− controls (2.21-fold increase; P<0.001; Fig. 9D,E). The dko diaphragm was similar in size to the wild-type and desmin−/− controls (Fig. 9D), but contained proportionately similar amounts of collagen as the mdx4cv diaphragm (28% in the dko compared to 29% in mdx4cv; Fig. 9D,E). Together, these results demonstrate that the impaired respiratory function in the dko mice resulted, at least in part, from the impaired alignment of sarcomeres and deposition of collagen between the myofibers in the diaphragm.
Increasing utrophin expression is a promising target for treatment of DMD [33]. While the downstream signaling pathways that influence utrophin expression are well described [33], [38], [39], the upstream mechanisms are less clear. Here, we found that perturbing the highly ordered structure of striated muscle by genetically deleting desmin from mdx4cv mice increased utrophin expression to levels that prevented skeletal muscle necrosis. We report a ∼2.5-fold increase in utrophin expression in the dko sarcolemma of 1a, 2a and 2d/x fiber types, which prevented necrosis by maintaining the integrity of the sarcolemma. Understanding the structural mechanisms that influence utrophin expression in the dko skeletal muscles may contribute to development of a therapy for DMD.
We found that the onset of necrosis in the mdx4cv gastrocnemius muscles was coincident with the loss of utrophin expression from the maturing fibers (Fig. 4), as previously described [22], [36]. MyoD initiates skeletal muscle differentiation and maturation by activating many skeletal muscle genes and suppressing others [62]. MyoD activates the transcription of miR-206, which targets the utrophin mRNA for degradation leading to the loss of utrophin expression from the sarcolemma and its replacement by dystrophin [63]. Analysis of C2C12 cells suggests that several other miRNAs may also repress the expression of utrophin [64]. The loss of utrophin expression from the sarcolemma of maturing fibers was delayed in desmin−/− muscles and prevented in the dko muscles. It will be interesting to test whether desmin can influence the expression, trafficking, or function of miRNA's that knock-down utrophin expression.
An alternate possibility is that an early pulse in utrophin transcription [65] increased utrophin expression to levels that could overcome the knockdown effects of the miRNA's. Muscle contraction can change the shape of nuclei [66], which can change gene expression [67]–[69]. Desmin interacts with myonuclei via plectin and lamin A/C [70]–[72]. The myonuclei in the desmin−/− muscles remain oval shaped in response to muscle contraction [66]. This could potentially lead to the persistence of a developmental gene expression program that underlies the increased utrophin expression in the dko.
Utrophin is normally expressed at low levels on the sarcolemma of the slower oxidative fibers in wild-type mice [73]. Inducing the oxidative myogenic program can alleviate the dystrophic pathology in mdx mice by stimulating utrophin expression. For instance, activation of PGC1α [74]–[76], calcineurin A/NFAT [77]–[80], GA binding protein [74], Ca2+/calmodulin [81], AMP activated protein kinase [82], and the transcriptional activator PPARβ/δ [83] can each induce the slow oxidative program in mdx muscle and increase utrophin expression. Metabolic changes to the muscle can also influence utrophin expression [84]. While we found no significant change in fiber-types when comparing mdx, desmin−/− and dko soleus muscles (Fig. S3), we did find utrophin expression on the extrasynaptic sarcolemma of 1a, 2a and 2d/x fiber-types, but not in the fast 2b fibers. Thus, our results are consistent with the activation of the slower oxidative myogenic pathways that can induce utrophin expression.
The absence of desmin in stressed muscle is associated with a shift in the expression of muscle proteins to those found in slow-twitch fibers [85], [86]. These changes may be mediated in part by changes in the activity of calcineurin linked to alter myoplasmic Ca2+ levels, which could result from a loss of local protein kinase A (PKA) signaling linked to the loss of desmin. The copolymerization of desmin with synemin in the intermediate filament reticulum contributes to synemin's localization around Z-disks [87], [88]. As synemin is an A kinase anchor protein (AKAP) [89] the absence of desmin in the dko is likely to alter local PKA activity associated with the sarcomere. Calcium homeostasis is likely to be affected locally as PKA can regulate many channels and transporters essential for normal excitation-contraction coupling [90]–[93]. However, our finding that the mRNA levels for utrophin do not change in extracts of dko gastrocnemius muscle, compared to age-matched wild-type, desmin−/− and mdx4cv muscles, argue against this mechanism.
While there are various signaling pathways that can activate utrophin transcription in mdx mice, we found no changes in utrophin mRNA in the dko total gastrocnemius muscle lysates when compared to the mdx4cv, desmin−/− or wild-type muscles. The persistence of utrophin on the dko sarcolemma of maturing skeletal muscle fibers is consistent with increased utrophin stability and post-transcriptional mechanisms. Proteins experimentally over-expressed within the mdx extrasynaptic sarcolemma such as sarcospan [94], [95], cytotoxic T cell GalNac transferase [96] and biglycan [35] can stabilize utrophin to prevent skeletal muscle necrosis. RhoA, a small GTPase also increases utrophin expression without apparently influencing transcription [97]. Stabilizing RNA, a known function for the type III intermediate filament protein vimentin [98], is another potential mechanism that can increase utrophin expression without changing transcription [99], [100]. Desmin may also influence protein degradation pathways by trafficking lysosomes through the muscle via its interaction with myospryn [101], [102].
Increasing utrophin expression by increasing utrophin transcription or stabilization can restore the expression of the DGC to the sarcolemma [53], [96], [103]–[105], except for nNOS [106]. We found that utrophin was able to concentrate β-dystroglycan to the sarcolemma in the dko 1a, 2a and 2d/x fiber types. However, the expression of the NODS sub-complex was not restored in the dko muscles. nNOS influences blood flow to the skeletal muscles and can lead to hypoxic stress injury post-exercise [12]. However, the long-term effects of the lack of nNOS are difficult to predict considering Becker muscular dystrophy patients expressing truncated dystrophins can have a mild phenotype without restoring nNOS to the sarcolemma [12], [107]. The low level of α-dystrobrevin on the sarcolemma may have contributed to the low level of central nuclei in the dko mice, as α-dystrobrevin−/− mice have a mild dystrophy and residual expression of α2-dystrobrevin mitigates the dystrophic pathology in mdx muscles [13]. While α1-syntrophin is an important adapter protein that is required for the localization of nNOS and aquaporin to the sarcolemma of striated muscle [108], [109], its role in the pathogenesis of DMD is unclear. The low level of the NODS complex in the dko muscles did not result from the lack of desmin (Fig. S4). Thus, the low level of central nuclei (∼9%) in the dko muscle fibers with extrasynaptic utrophin likely resulted from the lack of desmin in combination with the reduced expression of the NODS complex from the extrasynaptic sarcolemma.
We found that the dystrophic pathology in the fast 2b fibers was also improved in the dko despite a significant reduction in extrasynaptic utrophin expression when compared with mdx4cv fast 2b fibers. Most striking was the fact that utrophin expression was reduced in the extrasynaptic sarcolemma of regenerating fast 2b fibers in the dko. However, we found no overt change in the regenerative capacity of the muscle stem cells in the dko gastrocnemius muscles injured with notexin. In contrast, Agbulut and colleagues found that desmin−/− muscles injured with cardiotoxin displayed persistent expression of developmental myosin, small caliber fibers and the infiltration of adipocytes [59]. Here, we found no evidence of increased adipocytes in the desmin−/− or dko muscles. Therefore, the discrepancy between our studies may have resulted from the different myotoxins. In any case, we found a significant reduction in the number of necrotic fibers in the dko supporting a mechanism that prevents dystrophy rather than influencing regeneration. Desmin is also likely to play a structural role in linking the contractile apparatus to the sarcolemma [47], [52], [101] and in regulating the passive mechanical properties of skeletal muscle [66], [110]. We found that utrophin could form costameric striations with α-sarcomeric actin in dko mice, but the rectilinear pattern was severely impaired. The exacerbated loss of sarcomere alignment in dko diaphragms suggests the absence of desmin and potentially the NODS complex could weaken the sarcomeric connections to the membrane. However, it is important to note that the specific force production of mdx4cv and dko diaphragms was comparable. The mdx4cv mice have a compensatory hypertrophy that can potentially maintain peak force production [111]. However, the dko diaphragms lack this cellular hypertrophy suggesting that the impaired diaphragm function could contribute to the respiratory distress and shortened lifespan. Considering the dko mice die prematurely from apparent cardiorespiratory failure, it is possible that reduced mobility in the cage could mitigate contraction-induced injury to the muscles. We are currently investigating whether desmin influences contraction-induced injury to the sarcolemma in mdx4cv muscles.
In conclusion, we report a significant increase in utrophin expression in dko skeletal muscles that prevented necrosis in a fiber-type specific manner. The fact that utrophin expression was elevated ∼2.5-fold on the dko sarcolemma when compared with mdx4cv muscles is of considerable interest for developing treatments for DMD [26]. Clearly, deleting desmin is not a therapeutic option for DMD as the dko mice die from apparent cardiorespiratory distress, but understanding the upstream mechanisms that influence utrophin expression may lead to novel treatment strategies for DMD. Furthermore, an utrophin-mediated therapy developed from the dko mice would treat all muscle fiber-types in the human as humans lack the fast 2b fiber types. Considering desmin functions to maintain the highly ordered structure of striated muscles [44], [45], it is likely that utrophin expression in the dko is initiated by changes to muscle structure/signaling relationships. We also found that utrophin-independent mechanisms were improving the dystrophic pathology in dko fast 2b fibers, which will be of interest for understanding the pathophysiology of DMD. Thus, the dko mice may provide new insights into the regulation of utrophin expression that are relevant for the treatment of DMD.
We utilized C57Bl/6 wild-type mice, desmin−/− mice, mdx4cv mice and mdx:desmin dko mice. All experiments were in accordance with the Institute of Animal Care and Use Committee of the University of Washington. The desmin−/− mice were a kind gift from Professor Yassemi Capetanaki. We generated the dko mice by first backcrossing the desmin−/− mice from the FVB strain to the wild-type C57Bl/6 strain for five generations (N5). The resulting desmin−/− mice on the C57Bl/6 strain were then inbred for at least four generations to obtain desmin−/− controls (>F4) or they were crossed with the mdx4cv strain on the C57Bl/6 background and inbred for at least four generations to obtain the dko mice (>F4). Therefore, the mice generated for this study were B6.FVB-Desmin and B6.FVB-Desmin-mdx4cv incipient congenic with ∼96.9% homozygosity with the C57Bl/6 background. We genotyped the mice using standard PCR for desmin and performed sequence analysis of the mdx4cv genomic DNA to avoid potential false positives as previously described [112]. The desmin−/− and dko mice were sacrificed if they lost body mass or exhibited signs of cardiorespiratory distress. Kaplan-Meyer survival analysis was performed with 16 desmin−/− male mice and 13 dko male mice.
The diaphragm physiology was performed as previously described [113]. Briefly, the diaphragm from wild-type (n = 6), desmin−/− (n = 5), mdx4cv (n = 5) and dko (n = 5) was placed in oxygenated KREBS (2 mM Ca2+, 24 mM NaHCO3, 137 mM NaCl, 5 mM KCl, 1 mM MgSO4, 1 mM NaH2PO4, D-Glucose). Strips of the diaphragm were dissected and the optimum length and peak tetanic contractile force was measured over 350 ms. Because the diaphragm strips vary in size, a direct comparison of peak contractile force is not plausible. After contraction, the diaphragm strip is weighed and specific force was calculated as peak tetanic force production × length × density (1.04) × pennation (1 for the diaphragm)/muscle mass.
Costamere analysis was performed as previously described [52]. Briefly, the mice were anaesthetized with 2,2,2-tribromoethanol (Sigma) and perfused with 2% paraformaldehyde (Electron microscopy sciences). The muscles were incubated in 2% paraformaldehyde for 2 hours at 4°C, then washed 3 times with 1× PBS, and incubated in 10% sucrose for 1 hour at 4°C, and then 20% sucrose overnight at 4°C. The muscles were then placed in cryovials and flash frozen in liquid N2. The frozen samples were placed on a frozen chuck with OCT and 40 µm thick sections were cut using a cryostat. The sections were immunostained with 1∶800 utrophin A polyclonal antibody (kind gift from Stanley Froehner) and 1∶500 α-sarcomeric actin monoclonal antibody (SIGMA). The thick sections were imaged using a Leica SP5 confocal microscope.
The electron microscopy was performed on longitudinal sections of diaphragm muscle as previously described [114].
Muscles were frozen directly in OCT cooled in 2-methylbutane in liquid N2. Ten micrometer transverse sections of skeletal muscles were stained with hematoxylin and eosin, alizarin red and Sirius red using manufacturer protocols (Electron Microscopy Sciences; Hatfeild, PA). The Sirius red staining of collagen was measured using the manufacturers protocols in Image J analyses software. Transverse frozen sections were also immunostained as previously described [40]. Briefly, the sections were incubated in blocking buffer (1% BSA, 0.05% Triton X-100 in 1× phosphate buffered saline (PBS)) for 30 minutes and immunostained with antibodies to desmin (1∶50; DAKO Corp), N-terminal dystrophin antibody (1∶800), utrophin (1∶800), α-dystrobrevin 1 (1: 500), α-dystrobrevin 2 (1∶1000), α1-syntrophin (1∶500; the latter four antibodies were kind gifts from Stanley C. Froehner), β-dystroglycan (1∶100; Transduction Laboratories), MHCd (1∶40; Novocastra), α2-laminin (1∶800; Sigma) or nNOS (Zymed; 1∶100) for 1 hour. The sections were washed 3 times in 1× PBS for 10 minutes each and incubated in Alexa-488, Alexa-555, Alexa-594 or Alexa-647 secondary antibodies for 30 minutes (1∶800; Invitrogen). To label necrotic fibers we immunostained the muscles with mouse IgG1 antibodies conjugated to Alexa 488 (1∶800; Invitrogen). For labeling of acetylcholine receptors we incubated the sections in α-bungarotoxin conjugated to TRITC for 1 hour (1∶800; Invitrogen). The sections were washed 3 times for 10 minutes each and coverslipped with ProLong Gold mounting medium containing DAPI (Invitrogen). Muscle fiber typing was performed using conjugated monoclonal antibodies as previously described [115]. Sections were imaged with either a Leica SP5 confocal (Fig. 1, 3, 6), Nikon eclipse E1000 (Fig. 2, 7, 8) or an Olympus SZX16 dissection fluorescent microscope (Fig. 4, 5).
Quantitation of maximal sarcolemmal utrophin fluorescence intensity was performed as previously described for dystrophin [116]. Briefly, gastrocnemius muscle sections and images were processed identically for quantitation. We utilized the FIJI analyses software to quantitate maximal fluorescence intensity. The utrophin fluorescence intensity on the wild-type sarcolemmal was used as a negative control and the utrophin fluorescence intensity at the wild-type synapse was used as the peak of detection. We drew a line across the images to ensure unbiased quantitation and measured the peak fluorescent intensity that coincided with extrasynaptic sarcolemma staining. The sarcolemmal utrophin fluorescence intensity from mdx4cv, dko and microutrophinΔR4–R21 treated mdx∶utrophin double knockout muscles all fell within these limits. The mean +/− S.D. fluorescence intensity from n = 4 mice from 92 wild-type, 99 desmin−/−, 100 mdx4cv, 112 dko, and 77 microutrophinΔR4–R21 treated mdx:utrophin double knockout myofibers were compared.
The mdx4cv and dko mice (n = 4) were administered 200 µl of 0.22 µm filter sterilized 1% (w/v) EBD solution in HBSS intravenously by retro-orbital injection. Mice were sacrificed 3 hours after EBD administration. The gastrocnemius muscles were frozen in OCT in 2-methylbutane in liquid N2. Ten micrometer sections were cut and stained for utrophin (1∶800; kind gift from Stanley Froehner). Utrophin was labeled with Alexa-488 goat anti-rabbit secondary antibody (Invitrogen). The sections were viewed and imaged using the Olympus SZX16 dissection fluorescent microscope.
The gastrocnemius muscles of wild-type, desmin−/−, mdx4cv and dko (n = 8) were administered 30 µl of 1 µg/ml notexin in PBS at 11 weeks of age. The mice were sacrificed 4 days (n = 4) and 6 days (n = 4) post-injury. The gastrocnemius muscles were frozen in OCT. Ten micrometer sections were immunostained with α2-laminin (1∶800; Sigma) and developmental myosin heavy chain (1∶40; Novocastra) and directly compared to adjacent sections stained with hematoxylin and eosin. Considering monoclonal antibodies can label necrotic fibers, we defined regenerating fibers as those fibers that expressed developmental myosin heavy chain and contained centrally located nuclei.
Western blots were performed on whole muscle lysates as previously described [40]. Briefly, the gastrocnemius muscles of 3 and 11-week-old wild-type, desmin−/−, mdx4cv and dko (n = 6) were ground in liquid N2 and homogenized in extract buffer (50 mM Tris-HCl, 150 mM NaCl, 0.2% SDS, 24 mM Na Deoxycholate, 1% NP40, 47.6 mM Na Fluoride, 200 mM Na Orthovanadate, Roche). Protein concentration of whole muscle was determined by Coomassie Plus Bradford Assay (Pierce). Equal amounts of protein (10 µg) were resolved on a 4–12% SDS polyacrylamide gel. The blots were incubated in utrophin (1∶1000; kind gift from Stanley C. Froehner) overnight at 4°C. The α-sarcomeric actin primary antibody (1∶500; Sigma) was used as a loading control as its expression was unchanged when comparing the different strains of mice, as previously described for wild-type versus mdx4cv [40], [117]. We also loaded 20 µg of total protein to compare the expression of desmin, β-dystroglycan (1∶100; BD Transduction laboratories), α1-syntrophin (1∶500; kind gift from Stanley C. Froehner), pan α-dystrobrevin (1∶1000; BD Transduction laboratories) primary antibodies. The primary antibodies were detected with IgG HRP secondary antibodies (1∶6000; Jackson ImmunoResearch Labs). The blots were developed with ECL plus (Pierce) and scanned with the Storm 860 imaging system (Amersham Biosciences). The band intensity was measured using Image J software (NIH). The relative amount of utrophin in each blot was determined using a non-linear regression generated by a titration of utrophin from the dko from 1.25 µg up to 20 µg of total loaded protein and examined using the PRISM statistics software (Figures S1, S2; n = 4 for wild-type and desmin−/− and n = 8 for mdx4cv and dko samples).
To isolate the RNA, approximately 20 µg of gastrocnemius muscle previously ground by mortar and pestle in liquid N2 was used to extract total RNA following manufacturers instructions (TRI Reagent, Molecular Research Center). We used gastrocnemius muscles from 11 week old (Fig. 3D) or 3 week old mice (Fig. 4C). The pelleted RNA was suspended in 50 µl nuclease free elution solution (Ambion, Austin, TX). Five µg of total RNA was treated with Turbo DNA-free (Ambion, Austin, TX) in order to remove trace amounts of contaminating DNA. The DNAase Treated RNA (0.5 µg) was diluted to 8 µl with nuclease free water followed by use of the SuperScript™ III First-Strand Synthesis kit (Invitrogen, Carlsbad, CA) to generate cDNA. Subsequently 2 µl of the cDNA was used for qPCR with utrophin primer-probe sets. The mouse utrophin primers sequences were: Forward 5′- ACCAGCTGGACCGATGGA-3′, Reverse 5′- CTCGTCCCAGTCGAAGAGATCT-3′, Probe 5′-6FAM- CGTTCAACGCCGTGCTCCACC-3′-BHQa1-Q. As a reference gene the oligonucleotide set was used to target the mouse Ywhaz gene sequence (Tyrosine 3-monooxygenase; [118]): Forward 5′- GCTGGTGATGACAAGAAAGGAAT-3′, Reverse 5′- GGTGTGTCGGCTGCATCTC-3′, Probe 5′-6FAM- TGGACCAGTCACAGCAAGCATACCAAGA-3′-BHQa1-Q.
The data were compared using a one-way ANOVA with a Tukey post-test that compares all data sets with a Student's t-test. The relative amounts of utrophin in western analyses were determined using a non-linear regression generated from a titration of utrophin in the dko gastrocnemius muscles (from 1.25 µg–20 µg of total added protein). All data analyses were performed using the PRISM software.
|
10.1371/journal.ppat.1004155 | Bacterial Superantigens Promote Acute Nasopharyngeal Infection by Streptococcus pyogenes in a Human MHC Class II-Dependent Manner | Establishing the genetic determinants of niche adaptation by microbial pathogens to specific hosts is important for the management and control of infectious disease. Streptococcus pyogenes is a globally prominent human-specific bacterial pathogen that secretes superantigens (SAgs) as ‘trademark’ virulence factors. SAgs function to force the activation of T lymphocytes through direct binding to lateral surfaces of T cell receptors and class II major histocompatibility complex (MHC-II) molecules. S. pyogenes invariably encodes multiple SAgs, often within putative mobile genetic elements, and although SAgs are documented virulence factors for diseases such as scarlet fever and the streptococcal toxic shock syndrome (STSS), how these exotoxins contribute to the fitness and evolution of S. pyogenes is unknown. Here we show that acute infection in the nasopharynx is dependent upon both bacterial SAgs and host MHC-II molecules. S. pyogenes was rapidly cleared from the nasal cavity of wild-type C57BL/6 (B6) mice, whereas infection was enhanced up to ∼10,000-fold in B6 mice that express human MHC-II. This phenotype required the SpeA superantigen, and vaccination with an MHC –II binding mutant toxoid of SpeA dramatically inhibited infection. Our findings indicate that streptococcal SAgs are critical for the establishment of nasopharyngeal infection, thus providing an explanation as to why S. pyogenes produces these potent toxins. This work also highlights that SAg redundancy exists to avoid host anti-SAg humoral immune responses and to potentially overcome host MHC-II polymorphisms.
| Streptococcus pyogenes is the most common cause of bacterial pharyngitis, also known as ‘strep throat’. However, this organism is also responsible for a range of other important human illnesses including necrotizing fasciitis and rheumatic heart disease (RHD). Indeed, complications from RHD and invasive infections by S. pyogenes are responsible for over one half million deaths per year in the world. S. pyogenes produces potent toxins called superantigens (SAgs), also known as the scarlet fever or erythrogenic toxins. SAgs have been studied for many years, yet we don't understand what purpose SAgs play in the life cycle of S. pyogenes. Rather than studying SAgs in the context of serious streptococcal disease, we studied the role of SAgs in a nasopharyngeal infection model. Our work demonstrates that for S. pyogenes to efficiently infect mice, the mice must express a human protein that is a receptor for the SAgs, and that S. pyogenes must produce SAgs. We further show that immunizing against SAgs prevents nasopharyngeal infection. This work demonstrates that SAgs are important factors for establishing infection by S. pyogenes and that SAgs may be potential candidates for inclusion within a S. pyogenes vaccine.
| S. pyogenes (commonly known as the β-hemolytic Group A Streptococcus) is a prominent bacterial pathogen that causes a diverse range of clinical manifestations. Globally, S. pyogenes is responsible for over 600 million cases of pharyngitis, and more than one half million deaths primarily from complications of autoimmune rheumatic heart disease and invasive infections [1]. In addition, approximately 12% of school-aged children are asymptomatic carriers of this organism [2], and this ‘carriage’ state can last for years without the development of disease [3]. Although humans remain the only known natural reservoir for S. pyogenes, closely related streptococci such as Streptococcus canis and Streptococcus dysgalactiae lack host-specific adaptation. This indicates that the ancestor to S. pyogenes was unlikely to be human specific, and this further suggests that a primary feature in the evolution of S. pyogenes was the stringent adaptation to the human host [4]. Although morbidity associated with S. pyogenes is dependent on the ability to colonize and transmit within human populations, the molecular basis of the human specific tropism by S. pyogenes remains poorly understood.
One group of ‘trademark’ virulence factors produced by S. pyogenes are the bacterial superantigens (SAgs), also commonly referred to as the erythrogenic toxins or the streptococcal pyrogenic exotoxins [5]. There are at least 14 genetically distinct streptococcal SAgs [6] often encoded within mobile, or putatively mobile, genetic elements [7]–[9]. Thus, different S. pyogenes strains typically encode for distinct repertoires of multiple SAgs. These toxins function by engaging lateral surfaces of both MHC class II (MHC-II) molecules and T cell receptor (TCR) β-chains [10]; these unconventional interactions can force the activation of enormous numbers of T cells. Indeed, human MHC-II molecules are known host factors fundamental to the development of severe streptococcal disease [11]–[14], and the ability of MHC-II to modulate severity of invasive streptococcal disease has been linked directly to SAgs [12], [15]. Although SAgs are well recognized in the pathogenesis of scarlet fever [16], [17] and the streptococcal toxic shock syndrome (TSS) [18], [19], in what context these toxins contribute to the fitness and life cycle of S. pyogenes is unknown. Thus, since a major biological niche for S. pyogenes is the upper respiratory tract, we hypothesized that SAgs have likely evolved to function in the context of asymptomatic nasopharyngeal colonization and/or pharyngitis, rather than in the context of severe invasive disease.
Here we show that mice expressing human MHC-II molecules are highly susceptible to acute nasopharyngeal infection by S. pyogenes. Furthermore, we demonstrate that S. pyogenes MGAS8232 requires the streptococcal pyrogenic exotoxin A (SpeA) SAg to cause acute nasopharyngeal infection. In addition, immunization with toxoid SpeA is protective for nasopharyngeal infection by wild-type S. pyogenes MGAS8232. This work indicates that the streptococcal SAgs play an important role in the life cycle of S. pyogenes by promoting the initial stages of infection, and that these toxins should be further considered as potential vaccine targets to prevent S. pyogenes nasopharyngeal carriage.
S. pyogenes infection of the nasopharynx in mice is a model for pharyngeal infection in humans [20], [21]. We first evaluated the influence of human MHC-II on mouse nasopharyngeal infection by S. pyogenes. Nasal inoculation of C57BL/6 (B6) mice with ∼1×108 CFU of S. pyogenes MGAS8232 resulted in very low bacterial recovery from the nasal mucosa [herein referred to as the complete nasal turbinates (cNT)] at 48 h (Figure 1A). However, ‘humanized’ B6 mice that express HLA-DR4 (DR4-B6 mice) contained ∼100-fold more CFUs of S. pyogenes compared to wild-type B6 mice, and mice that expressed HLA-DQ8 (DQ8-B6 mice), or both HLA-DR4 and HLA-DQ8 (herein referred to as HLA-B6 mice), contained ∼10,000-fold more CFUs than wild-type B6 mice (Figure 1A). In a time course analysis of this acute infection model, ∼5×103 bacterial CFUs of S. pyogenes MGAS8232 were recovered at 24 h, which increased by ∼100-fold at 48 h, and were subsequently cleared after one week (Figure 1B). Importantly, MGAS8232 did not become invasive as viable S. pyogenes cells were not recovered from the spleen (Figure 1B) or blood (data not shown) in HLA-B6 mice at any time point. These data reveal that human MHC-II molecules are important host factors for acute nasopharyngeal infection by S. pyogenes.
To investigate the potential role of SAgs for nasopharyngeal infection in HLA-mice, we mined the genome of S. pyogenes MGAS8232 and confirmed that this strain encodes for six genetically distinct SAgs (SpeA, SpeC, SpeG, SpeL, SpeM and SmeZ) [22]. The mature coding region for each SAg gene was cloned, recombinant SAgs (rSAg) were expressed and purified, and each was shown to activate human T cells (data not shown). To assess activity of the rSAgs in HLA-mice, ex vivo splenocyte activation was evaluated. B6 splenocytes demonstrated little capacity for activation by all six rSAgs as measured by mouse IL-2 production and proliferative responses (Figure 2A). However, splenocytes from HLA-B6 mice showed enhanced responses to both the SpeA and SmeZ rSAgs in a dose-dependent manner (Figure 2B). These data indicate that both SpeA and SmeZ function in HLA-B6 mice and suggest that SAgs encoded within S. pyogenes MGAS8232 may contribute to the ability of S. pyogenes to infect the nasopharynx of HLA-mice.
In order to determine empirically if SAgs were direct contributors to the enhanced nasopharyngeal infection phenotype in HLA-mice, a series of precise, in-frame deletions were generated within the coding regions for each SAg gene in S. pyogenes MGAS8232 (Figure 3A). As speL and speM are encoded in tandem, these SAgs were deleted together. In addition, we generated a complete SAg deletion strain lacking the coding regions for all six SAg genes (MGAS8232 ΔSAg) (Figure 3A). Each of the S. pyogenes deletion strains grew comparably to wild-type MGAS8232 in vitro (Figure 3B). Furthermore, the various mutants did not show any alterations in protease (SpeB) activity (Figure 3C). To evaluate SAg production in these strains in vitro, we generated rabbit polyclonal antibodies to each rSAg. The specific rabbit antisera recognized each of the expected SAgs without cross-reacting with others (Figure 3D). Next, we assessed in vitro production of the SAg proteins for the wild-type and isogenic SAg deletion strains. The production of both SpeC and SpeL from wild-type S. pyogenes MGAS8232 was clearly detectable by Western blot analysis, while SpeA was weakly detected (Figure 3E). Importantly, for each of these three SAgs, the toxin was not made by the appropriate deletion mutant. To evaluate SAg activity from the different S. pyogenes mutants, we tested the supernatants from the isogenic S. pyogenes strains for the ability to activate HLA-B6 mice splenocytes. Consistent with the Western blot analysis (Figure 3E), and the ability of select SAgs to activate these cells (Figure 2B), the largest reduction in activity for the individual deletion strains was for MGAS8232 ΔSpeA, whereas the MGAS8232 ΔSAg mutant did not activate splenocytes above background levels (Figure 3F).
We next evaluated all of the isogenic S. pyogenes MGAS8232 strains for the ability to infect HLA-B6 mice. MGAS8232 ΔSpeA demonstrated a striking reduction in bacterial CFUs recovered at 48 h compared to wild-type MGAS8232 (Figure 4A). Although some variability was seen for the other mutants in vivo (ΔSpeC, ΔSpeG, ΔSpeLM and ΔSmeZ), multiple animal experiments demonstrated that each was capable of infecting the HLA-mice to wild-type levels. MGAS8232 ΔSAg however, resembled wild-type MGAS8232 recovery in B6 mice (Figures 1A and 4A). In order to genetically complement the MGAS8232 ΔSAg strain, and to determine if a single SAg could promote infection, we introduced the wild-type speA gene (and the corresponding speA promoter) into the chromosome of MGAS8232 ΔSAg between the pepO and tsf genes. The wild-type SpeA complemented strain enhanced infection significantly above MGAS8232 ΔSAg that was not statistically different from wild-type MGAS8232 (Figure 4A). To assess the requirement for SAg activity for the infection phenotype, we constructed a SpeA MHC-II binding mutant based on a structural model of SpeA in complex with HLA-DQ8 (Figure 4B). This model predicted SpeA Tyr100 would interact with the conserved MHC-II α-chain Lys39 and both of these equivalent residues in the staphylococcal enterotoxins, and HLA-DR1, respectively, have been shown to be important for SAg function [23], [24]. Recombinant SpeAY100A (Figure 4C) demonstrated ∼100-fold reduction in potency relative to wild-type SpeA (Figure 4D) and thus we introduced speAY100A into MGAS8232 ΔSAg in a similar fashion to the wild-type speA complementation experiment. The speAY100A-complemented strain was significantly reduced compared to the wild-type SpeA complemented stain, and although CFUs were not statistically increased over MGAS8232 ΔSAg, there was a trend for enhanced infection by ∼1 log. We believe this potential increase may reflect the residual activity present in the SpeAY100A mutant (Figure 4A). Furthermore, using qRT-PCR we confirmed that the in vitro expression of speAY100A was similar to speAWT between the two complemented strains (data not shown). These collective data indicate that the enhanced nasopharyngeal infection by S. pyogenes MGAS8232 in HLA-B6 mice is due to the production of the SpeA exotoxin, and that SAg function is required.
To gain insight into the S. pyogenes infection process in HLA-B6 mice, we generated coronal sections of the nasal passage for sham, wild-type MGAS8232, or MGAS8232 ΔSAg treated mice, at both 24 and 48 h. Sections were stained with H&E, as well as DAPI (blue) and an anti-S. pyogenes fluorescent antibody (red). As predicted, there was no evidence of S. pyogenes infection seen in the sham treated mice at 24 or 48 h (Figure 5A). Both wild-type MGAS8232 and MGAS8232 ΔSAg demonstrated sparse, but detectable S. pyogenes cells, within the upper nasal turbinates at 24 h (Figure 5A). This was consistent with bacterial counts from MGAS8232 ΔSAg infected mice at 24 h (975±475 CFU/cNT; n = 3 mice) that did not differ significantly from wild-type MGAS8232 infected mice (Figure 1B). Consistent with the CFU data (Figure 4A), very few MGAS8232 ΔSAg cells were detected microscopically at 48 h, whereas wild-type MGAS8232 produced a robust infection that was localized to the upper nasal turbinates (Figure 5A; 48 h MGAS8232 boxed insets). H&E-stained sections from 48 h post-infection (n = 5 mice per group, 2 sections per mouse) were scored in a blinded fashion for the presence and severity of mucus, red blood cells, and nucleated cellular debris on the surface of the respiratory epithelium and neuroepithelium. This analysis revealed no significant findings for the sham treated mice, and mice infected with MGAS8232 ΔSAg displayed only mild neutrophilic infiltration in the sub-epithelial spaces, and mild epithelial disruption (Figure 5B). However, mice infected with wild-type MGAS8232 demonstrated significant signs of inflammation including sloughed cellular debris into the nasal cavity lumen, and marked neutrophilic infiltration and epithelial cell disruption with evidence of cocci on the epithelial surface, and evidence of hypersecretory activity from the neuroepithelium (Figure 5B). Flow cytometric analysis of total cNT preparations from mice treated with the 3 groups (n = 4 mice per group) however, revealed few overall changes in immune cell percentages, although a significant decrease in the dendritic cell (DC) (CD11c+) population was revealed in the wild-type MGAS8232 infected animals compared with MGAS8232 ΔSAg infect mice, and the analysis also showed a trend for increased neutrophils (GR1+) in wild-type MGAS8232 infected animals (Figure S1) that was consistent with the histological analyses. We did not observe any changes in immune cell populations in the spleen or lymph nodes (Figure S1). However, cytokine analysis of homogenized cNTs demonstrated an early inflammatory type or TH1-skewed response of the wild-type MGAS8232 infected mice compared with the MGAS8232 ΔSAg strain, including enhanced production of IL1α, IL-2, IL-6 and IL-17, as well as the chemokines KC, IP-10 and MCP-1 (Figures 5C and S2). By 48 h, the wild-type MGAS8232 infected mice displayed a robust chemokine response consistent with the high numbers of bacterial cells by this time point. Taken together, these data are consistent with an early SAg-dependent inflammatory environment within the nasal turbinates where S. pyogenes could both survive and rapidly expand to high numbers over the initial 48 h of infection.
Given the prominent nasopharyngeal infection phenotype for S. pyogenes MGAS8232 that was SpeA-dependent (Figures 4A and 5A), we tested the ability of humoral immunity to SpeA to influence nasopharyngeal infection by S. pyogenes MGAS8232 (Figure 6A). We vaccinated HLA-mice with the SpeAY100A toxoid and these mice developed high anti-SpeA IgG titres compared with sham-vaccinated mice (Figure 6B). Following nasal inoculation with ∼1×108 CFU of wild-type MGAS8232, SpeAY100A-vaccinated mice showed a dramatic reduction in infection (Figure 6C) that was similar to the MGAS8232 ΔSpeA mutant (Figure 4A), while sham-vaccinated mice were infected efficiently (Figure 6C). These data indicate that humoral immunity to appropriate SAgs can inhibit nasopharyngeal infection in SAg-sensitive mice.
Bacterial SAgs from S. pyogenes are well recognized as important virulence factors for the development of serious toxin-mediated diseases such as the streptococcal TSS [5]. However, given that streptococcal TSS is rare, and since the disease has a very high mortality rate (∼45%) [25], it is unlikely that the induction of streptococcal TSS provides an evolutionary advantage for S. pyogenes. Thus, the actual biological function of these toxins has remained an enigma. We reasoned that since a major niche for S. pyogenes is the epithelial surfaces of the nasopharynx, we sought to evaluate the influence of SAgs in this context. Herein, we provide important evidence that acute infection of the upper respiratory tract by S. pyogenes MGAS8232 is enhanced dramatically in mice expressing human MHC-II, and that this phenotype is directly attributed to SpeA. A logical interpretation of these findings is that the true biological function of SAgs, in terms of the life cycle of S. pyogenes, is to promote the initial establishment of infection in humans.
This work also provides further evidence that host MHC-II molecules are a central component of the human specific tropism of S. pyogenes. Human MHC-II [11]–[14], human plasminogen [26], [27], and human CD46 [28] are each known to contribute to the development of invasive streptococcal disease, although the influence of these latter host factors have not yet been tested in a nasopharyngeal infection model. S. pyogenes produces a hyaluronic acid capsular polysaccharide [29] that is important for nasopharyngeal colonization in mice [30], and the hyaluronic acid binding receptor CD44 is an important receptor for S. pyogenes [31]. However, the conserved amino acid sequences, structures, and binding affinities of human and mouse CD44 for hyaluronan [32] would argue against a specific role of CD44 in the human-specific tropism of S. pyogenes.
Although there were no significant differences in CFUs at 24 h comparing wild-type S. pyogenes and the SAg-deficient strain, wild-type MGAS8232 was able to expand rapidly between 24 h to 48 h by ∼2 logs (Figure 1B), which was entirely consistent with the immunofluorescence staining of the nasal passages (Figure 5A). Although there were few differences in immune cell percentages from the cNT by 48 h, other than a clear decrease in CD11c+ leukocytes for wild-type MGAS8232 infected mice (Figure S1), S. pyogenes did induce a SAg-dependent Th1/inflammatory cytokine response, which was accompanied by the influx of innate immune cells by 48 h (Figure 5A). Selective depletion of CD11c+ DCs in mice resulted in enhanced dissemination of S. pyogenes from a subcutaneous infection into lymph nodes and the liver, demonstrating that DCs are contributors to host defence against S. pyogenes [33]. However, some strains of S. pyogenes can also induce DC apoptosis in a streptolysin O-dependent manner [34], which may have contributed to the decrease in CD11c+ populations. Nevertheless, we favour the hypothesis that S. pyogenes has provoked a localized and SAg-mediated cytokine response that resulted in a state of transient immunosuppression allowing S. pyogenes to escape myeloid cell mediated killing. In support of this hypothesis, staphylococcal enterotoxin B has been demonstrated in vivo to induce a transient (∼48 h) Vβ-unrestricted immunosuppressive response in T cells with the inability to produce IL-2, that also caused decreased numbers of splenic CD11c+ dendritic cells [35]. These responses were very consistent with mouse cytokine responses (Figure 5C) and immune cell analyses (Figure S1) of MGAS8232 infected HLA-B6 mice. By 48 h, however, a robust IL-6 and IL-17 response was produced (Figures 5C and S2), concurrent with the high numbers of S. pyogenes, and each of these cytokines are known to be important for S. pyogenes control [21], [36]. Alternatively, inflammation induced epithelial cell damage within the nasal turbinates (Figure 5B) may have promoted access to host cell adhesive factors to allow for the initial establishment of infection. The very weak cytokine response to the MGAS8232 ΔSAg strain was somewhat surprising, yet this finding may support this second hypothesis where in the absence of functional SAg, S. pyogenes are rapidly cleared through primarily mucociliary clearance mechanisms. Although S. pyogenes is capable of internalization into epithelial cells, the evidence indicates that S. pyogenes does not replicate efficiently within epithelial cells [37], [38], and thus we do not predict a role for enhanced intracellular survival based on SAg expression.
SAgs have also been studied directly in the context of live invasive streptococcal disease using defined genetic knockout strains. Earlier work utilizing a S. pyogenes myositis model in HLA-DQ8 transgenic mice demonstrated clear Vβ-specific alterations during infection, as well as SpeA-dependent conjunctivitis, hyperplasia of the lymph nodes and spleen, and T cell infiltration into the liver, yet the SpeA knockout strain demonstrated no difference in overall mortality compared with the wild-type counterpart [14]. Additionally, although SmeZ is dominant in the speA- and speC-negative M89 isolate S. pyogenes H293, genetic disruption of smeZ did not alter bacterial clearance or mortality in a peritoneal infection model [39]. Thus, a picture has emerged where individual SAgs may not contribute to S. pyogenes survival during invasive disease. Nevertheless, streptococcal SAgs can cause TSS directly in experimental animal models [40]–[42], human MHC-II molecules contribute to mouse mortality during invasive infections [12], [14], and in severe invasive human infections, streptococcal TSS is a major contributor to overall mortality [25].
In the nasal infection model presented here, despite the high number of bacterial cells recovered from wild-type S. pyogenes infected HLA-B6 mice at 48 h, S. pyogenes did not become invasive (Figure 1B), and the infection was reduced to very low levels after about 7 days. In humans, symptomatic pharyngitis infections typically last for about one week without antibiotic treatment [43], and thus the presented model appears to be a reasonable approximation of acute upper respiratory tract infection in humans. However, whether the model could replicate a longer term asymptomatic colonization state is unlikely, as many patients can harbour S. pyogenes for extended time periods, potentially lasting over 2 years [3].
Although our data demonstrate that SpeA is critical for the infection phenotype in HLA-B6 mice, the genome of S. pyogenes MGAS8232 also encodes for 5 other SAgs, each which is known to activate human T cells [44]–[47]. Although SpeC was the predominant SAg secreted from MGAS8232 in vitro (Figure 3E), the lack of a phenotype for the MGAS8232 ΔSpeC mutant was somewhat predicted, as this SAg does not activate murine T cells [48] (Figure 2). Although recombinant SmeZ did potently activate splenocytes from HLA-B6 mice, and SmeZ is the primary immunoactive SAg for some S. pyogenes strains [39], we did not detect expression of SmeZ in vitro from wild-type MGAS8232 (Figure 3E), which likely contributed to the inability of SmeZ to compensate functionally in the MGAS8232 ΔSpeA strain. However, the MGAS8232 ΔSAg strain complemented with wild-type SpeA did not appear to fully restore the infection phenotype, and the lack of SmeZ production could potentially be responsible for this result. The remaining SAgs showed very weak activity for the activation of splenocytes from HLA-mice and thus they do not play an important role for infection in this model.
The vaccination experiments provide further evidence, independent of the genetic deletion strains, of the critical role played by SpeA during the infection. Neutralization of SAg activity by antibodies is consistent with clinical evidence that has established a link between the lack of streptococcal SAg-neutralizing antibodies and the development of streptococcal TSS [49]-[51]. Also, neutralizing antibodies are known to protect against experimental STSS in rabbits [40], [41]. Thus, pre-existing anti-SAg antibodies may potentially inhibit infection by specific strains of S. pyogenes, yet S. pyogenes could theoretically circumvent this through up-regulation of additional SAgs for which neutralizing antibodies are absent. Since it is clear that different strains of S. pyogenes encode different repertoires of SAgs [7], S. pyogenes may also potentially alter SAg expression patterns to engage different host MHC-II molecules.
The SAgs remain well-recognized virulence factors for S. pyogenes. However, this work demonstrates that their genuine contribution to the life cycle of this pathogen is likely to promote the establishment of pharyngitis, or potentially asymptomatic colonization, in genetically susceptible individuals expressing SAg-responsive MHC-II molecules. The redundancy of SAgs within S. pyogenes has also remained unexplained, and this work further illustrates that this may exist, in part, to overcome the highly polymorphic nature of human MHC-II molecules, and also to avoid natural host immunity to SAgs. Thus, SAgs are important contributors to the complex genotype-phenotype relationship that exists between S. pyogenes and humans, and these toxins should be considered further as valid targets for vaccination studies to impede the enormous burden of disease by this versatile pathogen.
All animal experiments were in accordance with the Canadian Council on Animal Care Guide to the Care and Use of Experimental Animals. The animal protocol (#2009-038) was approved by the Animal Use Subcommittee at Western University.
S. pyogenes MGAS8232 is an M18 serotype that was isolated from a patient with acute rheumatic fever in Utah in 1987 and the genome has been sequenced and fully annotated [22]. S. pyogenes strains were grown in Todd Hewitt media supplemented with 1% (w/v) yeast extract. Deletions were made for all of the SAg genes using the pG+host5 system [52], [53]. All recombinant plasmids were built with standard molecular procedures using Escherichia coli XL1-blue as the cloning host [54]. Briefly, deletion constructs were generated by amplification of ∼500 bp of DNA on either side of the relevant SAg gene (Primers are listed in Table S1) and cloned into pG+host5. Flanking DNA included the first and last 8 codons for each SAg gene to generate precise, markerless and in frame deletions of each SAg gene. Plasmids were electroporated (Bio-Rad Gene Pulser XCell) into S. pyogenes MGAS8232 and single crossover integrations were selected at 40°C under erythromycin (1 µg ml−1) selection. PCR confirmed single crossover integrations were subcultured without antibiotics at 30°C and single clones were screened for a loss of erythromycin resistance, and double crossover gene disruptions were confirmed by PCR. In each case, appropriate PCR products were sequenced to confirm the expected deletion. All mutants were confirmed to lack growth alterations using Bioscreen C (Piscataway, NJ, USA) assays. To assess for protease activity, S. pyogenes strains were grown on dialyzed brain heart infusion agar containing 1.5% skim milk. Five microliters of OD600 0.1 S. pyogenes mutants were inoculated into 2 mm holes punched in the plates and incubated at 37°C for 24 hours. The ability of each strain to hydrolyze casein was assessed by the diameter of the zones of clearing. For speA and speAY100A complementation experiments, wild-type and mutant speA from MGAS8232 (including the native promoter) were individually ‘knocked in’ using the pG+host5 system to the non-coding region between genes encoding endopeptidase O (pepO) and elongation factor-Ts (tsf) and confirmed by PCR and DNA sequencing.
Genes encoding for SpeA, SpeG, SpeL, SpeM, and SmeZ, lacking nucleotides encoding the predicted signal peptides, were PCR amplified and cloned into a modified pET-28a vector to introduce an engineered tobacco etch virus (TEV) protease cleavage site downstream from a His6 tag. Cloning of SpeC into the pET-41a vector has been described [55], and SmeZ was cloned in a similar manner to SpeC. All recombinant SAgs were produced by 200 µM isopropyl β-D-1-thiogalactopyranoside (IPTG)-induced expression in E. coli BL21(DE3), purified via nickel affinity chelation chromatography, and His6 tags were removed using autoinactivation resistant His6::TEV protease [56], as described [57], [58]. Proteins were run on 12% separating SDS-PAGE gels and Western blots were visualized on a LI-COR Odyssey (LI-COR Biosciences) using IRDye800 conjugated donkey anti-rabbit IgG as the secondary antibody (Rockland Inc.). All recombinant SAgs ran as discrete homogenous bands by SDS-PAGE (Figure 3D). Purified and lyophilized SpeA, SpeC, SpeG, SpeL, SpeM, and SmeZ were used to generate polyclonal rabbit antibodies from a commercial source (ProSci Incorporated, USA).
HLA-expressing humanized mice (B6-DR4, B6-DQ8, B6-DR4/DQ8) have been previously described [12], [59], [60]. These mice were bred in a barrier facility at Western University and were routinely genotyped for the appropriate transgene(s). The mouse nasopharyngeal infection model has been described [61], with modifications. Briefly, mice were used at 9–13 weeks. Freshly grown exponential phase S. pyogenes cells (OD600 0.2–0.3) were washed and suspended in Hanks balanced saline solution [HBSS; a total of ∼1×108 (range 0.6–1.4×108) CFU per 15 µl] and 7.5 µl was used to inoculate each nostril under Forane (isoflurane, USP) inhalation anesthetic (Baxter Corporation). Sham treated mice only received HBSS. Mice were sacrificed at the noted time points, and the cNTs, including the nasal associated lymphoid tissue, nasal turbinates, and maxillary sinuses were removed. Tissue was homogenized in HBSS, serially diluted, and plated on Trypticase Soy Agar with 5% Sheep Blood plates (Becton, Dickinson and Company, MD, USA), or used for cytokine analysis, or flow cytometry.
For splenocyte activation experiments, cells were harvested from B6 or HLA-B6 mouse spleens, treated with ammonium-chloride-potassium (ACK) lysis buffer (15 mM NH4CL, 10 mM KHCO3, 0.1 mM EDTA) and suspended (2×105 cells per well) in RPMI 1640 medium supplemented with 10% heat-inactivated fetal bovine serum (Sigma-Aldrich), 0.1 mm minimal essential medium (MEM) non-essential amino acids, 2 mm L-glutamine, 1 mm sodium pyruvate, 100 U ml−1 penicillin, 100 µg ml−1 streptomycin (all from Gibco Life Technologies) and 50 µM β-mercaptoethanol (Sigma). SAgs were added at the indicated concentrations and mouse IL-2 was determined after 18 h by enzyme-linked immunosorbent assay (eBioscience). Proliferation was measure by the addition of 1 µCi/well [3H]thymidine after 72 h and after another 18 h cells were harvested on fiberglass filters and [3H]thymidine incorporation was assessed on a 1450 Microbeta liquid scintillation counter (Wallac).
The SpeA Tyr100→Ala mutation was predicted to disrupt the low-affinity MHC-II binding domain of SpeA based on a model of the crystal structure of SpeA in complex with HLA-DQ8. To generate this model, the structures of SpeA (PDB: 1FNU) [62] and HLA-DQ8 (PDB: 1JK8) [63] were superpositioned onto the known SEC3:HLA-DR1 complex (PDB: 1JWM) [64], and visualized using Pymol (pymol.sourceforge.net). Mutagenesis of speA was conducted using megaprimer-PCR (Table S1) to introduce the SpeA Tyr100→Ala mutation.
Mice were fully anesthetized with Forane (isoflurane, USP) inhalation anaesthetic (Baxter Corporation) and perfused through the heart with sterile PBS containing heparin using a Gilson Minipuls 3 peristaltic pump (Middletown, WI, U.S.A) at a constant flow rate. Mice were then perfused with 10% neutral buffered formalin (BDH, VWR, West Chester, PA, USA) through the peristaltic pump. The head was soaked in 10 volumes of formalin for 24 hours and re-suspended in Shandon TBD-2 Decalcifier (TBD; Thermo Scientific, Kalamazoo, MI, USA) for 96 hours. TBD-decalcified heads were placed in formalin for 48 hours and washed with 1× PBS, and resuspended in 10-volumes of 1× PBS twice daily for 4 days, washed in 70% ethanol twice, and stored in 10-volumes of 70% ethanol. Cassettes were processed in Leica ASP300 fully enclosed paraffin wax tissue processor overnight using the ‘bone’ program and embedded in paraffin wax. Heads were serially sectioned between the first and second molar on a HM335E Microtome (Leica) into 5 micron sections, mounted on Fisherbrand Superfrost Plus microscope slides (Fisher Scientific, Fair Lawn, NJ, USA) and dried at 45°C for 48 h prior to storage/staining. Tissues were stained with H&E in a Leica Autostainer XL.
H&E stained slides were evaluated by an experienced mouse pathologist in a blinded fashion. The relative amount of mucus present covering the epithelia, the presence of red blood cells, and the presence of nucleated cellular debris on the surface of the epithelia was assessed and the presence and severity of these findings were used to assign a score of zero to two points to each of two sections per mouse (n = 5 mice per group). The scores were averaged to determine differences in histological pathology in the mice. Fluorescence staining was done with adjacent serial sections using a Goat α-S. pyogenes polyclonal (NB200-643; Novus Biologicals) at 1∶100 dilution, and donkey α-Goat Alexflour 595 (A-11058; Invitrogen) at 1∶1000 dilution. Images were captured using an upright BX61 fluorescent microscope (Olympus).
For flow cytometry analysis, isolated cells (from cNTs, lymph nodes, or spleens) were aliquoted at 500,000 cells per 5 ml tube, and pre-treated with Fc block (hybridoma clone 2.4G2) prior to cell staining. Staining was done in panels using the following antibodies: α-CD3-APC (clone 145-2C11), α-CD3-PE Cy7 (clone 53–7.3), α-CD8 PE Cy7 (clone 53–6.7), α-CD45 PE (clone 30-F11), α-CD19 (clone MB19-1), α-NK1.1 (clone PK136), α-CD11c (clone N418), α-F4/80 (clone BM8), and α-GR1 (clone RB6-8C5) (all from eBioScience); α-CD4 APC Cy7 (clone GK1.5; Biolegend); and α-CD45 Alexafluor 700 (clone 30-F11; BD Biosciences). Dead cells were excluded using 7-AAD (BD Biosciences). Antibodies to stain cells for each panel were added, mixed, and incubated on ice in the dark for 30 minutes. Cells were washed twice with 1× PBS +5% FBS and resuspended in 500 µl of 1× PBS +5% FBS. Stained cells were run on a BD FACS Canto II flow cytometer (BD Biosciences). Standard compensations were used for each tissue using FACSdiva software.
Cytokine concentrations were determined from cNT homogenates isolated from mice treated with HBSS (sham), wild-type S. pyogenes MGAS8232, or isogenic S. pyogenes MGAS8232 ΔSAg strains at either 24 or 48 hours post-infection in HLA-B6 mice. Multiplex bead arrays were performed using the Mouse Cytokine 32-plex Discovery Array (Eve Technologies). Heat maps were generated using the Matrix2png algorithm [65] and data is shown as the average cytokine responses from 3–4 mice per group. Quantitative data from the cytokine analyses are shown in Figure S2.
For vaccination experiments, 6–8 week old HLA-B6 mice were injected subcutaneously with 1 µg of recombinant SpeAY100A (or sham) emulsified in Imject Alum Adjuvant (Thermo Fisher Scientific Inc.) every 2 weeks for a total of three injections. Two weeks following the last injection, mice were bled for antibody titers as determined by direct ELISA against 1 µg wild-type SpeA per well, and calculated as the reciprocal of the lowest serum dilution with readings 4-fold above background. Mice were challenged with ∼1×108 CFU wild-type MGAS8232 as described above 24 hours after the final bleed and CFUs were determined 48 h post infection.
When appropriate, individual data points, or the mean ± SEM, are shown, and p values were calculated using the Student's t-test with Prism software (GraphPad). A p value of less than 0.05 was determined to be statistically significant.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.